<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Julia Leonardi on Medium]]></title>
        <description><![CDATA[Stories by Julia Leonardi on Medium]]></description>
        <link>https://medium.com/@julialeonardi?source=rss-7297bf51f197------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*nT1xpPZqPeUOJuoOmPdypQ.jpeg</url>
            <title>Stories by Julia Leonardi on Medium</title>
            <link>https://medium.com/@julialeonardi?source=rss-7297bf51f197------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 16:11:44 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@julialeonardi/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[What are the tools of a GeoAI practitioner? (The Roadmap to GeoAI — Part 4)]]></title>
            <link>https://medium.com/@julialeonardi/what-are-the-tools-of-a-geoai-practitioner-the-roadmap-to-geoai-part-4-aa4f5e678d4b?source=rss-7297bf51f197------2</link>
            <guid isPermaLink="false">https://medium.com/p/aa4f5e678d4b</guid>
            <category><![CDATA[geoai]]></category>
            <category><![CDATA[frameworks-and-libraries]]></category>
            <category><![CDATA[ai-tools-for]]></category>
            <dc:creator><![CDATA[Julia Leonardi]]></dc:creator>
            <pubDate>Sun, 30 Mar 2025 21:45:21 GMT</pubDate>
            <atom:updated>2025-03-30T21:45:21.115Z</atom:updated>
            <content:encoded><![CDATA[<h3>What are the tools of a GeoAI practitioner? (The Roadmap to GeoAI — Part 4)</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sI9QbzaveyZ7nKXbi5qXaA.png" /></figure><blockquote>This is the fourth part of a five-part series that encapsulates everything you need to know about the exciting field of GeoAI. If you missed the previous parts, check them out! <a href="https://medium.com/@julialeonardi/the-roadmap-of-geoai-part-1-e3fefa2575a9">Part 1</a>, <a href="https://medium.com/@julialeonardi/the-roadmap-of-geoai-part-2-23d5167d93a6">Part 2</a>, <a href="https://medium.com/@julialeonardi/traditional-ml-in-geoai-applications-the-roadmap-to-geoai-part-3a-ed670a1bb7fe">Part 3a</a> &amp; <a href="https://medium.com/@julialeonardi/deep-learning-in-geoai-applications-the-roadmap-to-geoai-part-3b-bc60ca96b266">Part 3b</a>.</blockquote><p>Now that you have read about what GeoAI is, what AI methods to learn, and all its cool applications, you may be excited to get to work and do some GeoAI analysis yourself! But you might be wondering — where do I start? What software should I use? Should I learn programming? The tools for GeoAI analysis are slowly emerging, but to get the best out of them, you should first learn the fundamental tools specific to the GeoAI fields, i.e., AI and geospatial analysis.</p><p>In this article, I will introduce the software and tools you should look into if you want to start your GeoAI journey. The post is divided into three chapters: geospatial analysis-specific software, AI analysis-specific software, and emerging GeoAI-specific tools. For each group of tools, I will list the most important ones. However, remember that there are many more you can stumble across when working on your analyses—it’s important to always be open to learning new software in this dynamically evolving field.</p><h4>Python Preface</h4><p>It’s inevitable — to perform any data analysis, including geospatial or AI, <strong><em>you need to learn a programming language, and it should be Python</em></strong>. Writing this article, I understood that Python cannot be categorized as a tool either for AI development or geospatial analysis — it is used in all of these fields individually and combined. Python is rich in frameworks for geospatial and machine learning analysis. Additionally, it’s known for its readability and simplicity, making it accessible for beginners and powerful for experts.</p><p>In the following chapters, I will still assume Python as a vital tool and add the frameworks and libraries that are domain-specific and useful to learn.</p><h4>Geospatial analysis-specific software</h4><ul><li><strong>QGIS — </strong>as you will deal with spatial data, it is useful to know how to work with GIS software. QGIS is free and open-source, making it an accessible tool for everybody. It offers a graphic user interface for visualizing, editing, and analyzing spatial data. You can find many good learning resources for QGIS, such as <a href="https://docs.qgis.org/3.34/en/docs/training_manual/index.html">this manual</a> or <a href="https://courses.spatialthoughts.com/introduction-to-qgis.html">this course.</a></li><li><strong>Google Earth Engine (GEE ) — </strong><a href="https://earthengine.google.com/">GEE</a> is a cloud-based platform for geospatial analysis, which provides the user with access to petabytes of satellite data, a coding environment, and a practical visualization interface. While it primarily uses JavaScript for its code editor, a Python API is also available for those more comfortable with Python workflows. GEE requires all users to sign up for an account. A walkthrough of the process is provided <a href="https://courses.spatialthoughts.com/gee-sign-up.html">in this tutorial</a>. <a href="https://courses.spatialthoughts.com/end-to-end-gee.html#introduction">A complete course on GEE can be found on the <em>Spatial Thoughts</em> platform</a>.</li><li><strong>PostgreSQL + PostGIS — </strong>When working with large amounts of spatial data, it’s necessary to understand the role of spatial databases. PostgreSQL is an open-source relational database, and its PostGIS extension adds support for geographic objects. Together, they allow you to perform spatial queries efficiently. Understanding how to set up and use PostGIS is essential for scaling your GeoAI projects, especially with large geospatial datasets. At the beginning of your journey, you may use more cloud-based tools, which don’t require this type of knowledge, but I believe that a good foundation in database usage is a powerful asset for your future endeavors. <a href="https://medium.com/@tjukanov/why-should-you-care-about-postgis-a-gentle-introduction-to-spatial-databases-9eccd26bc42b">This post gives</a> a good introduction to the topic and might convince you that it’s worth learning after all. Another resource for learning about these technologies <a href="https://www.bostongis.com/printerfriendly.aspx?content_name=postgis_tut01">is this guide</a>.</li><li><strong>SNAP — </strong>When working with satellite data, it’s good to have some experience with software such as SNAP, which is designed especially for reading, exploring, and analyzing satellite imagery from different sensors. Although it is intended for Sentinel missions, it also works for other data. It’s an excellent tool for understanding how satellite data is structured, especially for more complex data types such as SAR imagery. To learn more about SNAP, look at <a href="https://medium.com/@eomasters/eomasters-blog-list-of-great-tutorials-for-esa-snap-1dcf859c16fb">this tutorial list</a>.</li><li><strong>Python libraries for geospatial analysis<br></strong>-<strong> rasterio — </strong>a library for reading and writing raster datasets in Python.<br>- <strong>GeoPandas — </strong>an extension of the popular Pandas library to handle spatial vector data in a “Pythonic” way.<br>- <strong>Shapely / Fiona / Pyproj — </strong>are complementary libraries useful for geometry operations, reading/writing geospatial files, and handling coordinate reference systems (CRS).</li></ul><h4>AI analysis-specific tools</h4><p>Since GeoAI heavily relies on machine learning and deep learning methods, you’ll need a solid foundation in AI development environments and libraries:</p><ul><li><strong>PyTorch / TensorFlow</strong> — These two are the most widely used deep learning Python frameworks. You’ll use them to build neural networks, train models, and experiment with architectures like CNNs, RNNs, and transformers. In the case of geospatial applications, <strong><em>I’d recommend focusing on PyTorch</em></strong>, as most emerging GeoAI libraries are based on it. A really good resource to learn about PyTorch is <a href="https://www.learnpytorch.io/">this course</a>, which comes with a <a href="https://github.com/mrdbourke/pytorch-deep-learning/tree/main">GitHub repo</a> and walks you through explanations and examples of usage.</li><li><a href="https://huggingface.co/"><strong>Hugging Face</strong></a> 🤗— Hugging Face offers an open-source hub of pre-trained models and datasets. It’s becoming a must-know platform for working with state-of-the-art AI models, and most of the 🔥<strong>hottest</strong>🔥GeoAI models are also hosted there! To learn about Hugging Face, check out <a href="https://www.freecodecamp.org/news/get-started-with-hugging-face/">this post</a>!</li></ul><h4>GeoAI-specific tools</h4><p>This is where things get <strong><em>really </em></strong>interesting. GeoAI is still an emerging field, so many tools are in active development. Most are Python libraries that extend existing AI frameworks with geospatial capabilities.</p><ul><li><strong>TorchGeo</strong> — A PyTorch domain library designed for geospatial deep learning. It simplifies working with geospatial datasets in PyTorch and provides dataset loaders, models, and utilities tailored for tasks like land cover classification or change detection. The best way to learn about the framework is to browse the <a href="https://github.com/microsoft/torchgeo">GitHub repo</a>, which also provides links to the documentation, tutorial videos, and podcast episodes about TorchGeo.</li><li><strong>TerraTorch</strong> — Another PyTorch-based library that focuses on remote sensing data, based on TorchGeo. It’s a project focused more specifically on Geospatial Foundation Models (GFMs) — <em>on which I’ll definitely be expanding in my future posts, as they are the focus of my PhD research 😉.</em> It provides a toolkit to utilize these models for fine-tuning for specific geospatial applications. Check out their <a href="https://github.com/IBM/terratorch?tab=readme-ov-file">GitHub repository</a> for more information.</li><li><strong>Raster Vision</strong> — An open-source Python library for deep learning on aerial and satellite imagery. Raster Vision is a framework created by <a href="https://element84.com/">Element 84</a> that helps build pipelines for object detection, classification, and segmentation tasks on geospatial raster data. To learn more, check out their <a href="https://rastervision.io/">official webpage</a>, <a href="https://docs.rastervision.io/en/stable/#">documentation</a>, or <a href="https://github.com/azavea/raster-vision">GitHub repo</a>!</li></ul><h4>Wrapping Up</h4><p>This wraps up this post! Remember, the tools I’ve introduced here are just a starting point — there are many more out there, each offering unique capabilities for geospatial and AI analysis. These are the ones I believe provide a solid foundation for anyone beginning their GeoAI journey, but the field is evolving rapidly, with new frameworks and methodologies emerging all the time. Staying curious, continuously exploring new tools, and keeping up with advancements is key in this new and exciting field!</p><p>I plan one last post in the <em>Roadmap to GeoAI series </em>to wrap it up, and after that, I’ll be posting about topics more specific to my work, and those that spark your interest more.</p><p>So, as always, don’t hesitate to leave a comment if you have any questions, suggestions, curiosity, or want to read about a specific subject. Let’s learn together! ✨🐸</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=aa4f5e678d4b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Deep Learning in GeoAI applications (The Roadmap to GeoAI — Part 3b)]]></title>
            <link>https://medium.com/@julialeonardi/deep-learning-in-geoai-applications-the-roadmap-to-geoai-part-3b-bc60ca96b266?source=rss-7297bf51f197------2</link>
            <guid isPermaLink="false">https://medium.com/p/bc60ca96b266</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[geoinformatics]]></category>
            <category><![CDATA[transformers]]></category>
            <category><![CDATA[geoai]]></category>
            <category><![CDATA[deep-learning]]></category>
            <dc:creator><![CDATA[Julia Leonardi]]></dc:creator>
            <pubDate>Tue, 10 Dec 2024 10:26:05 GMT</pubDate>
            <atom:updated>2024-12-10T10:26:05.812Z</atom:updated>
            <content:encoded><![CDATA[<h3>Deep Learning in GeoAI applications (The Roadmap to GeoAI — Part 3b)</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sI9QbzaveyZ7nKXbi5qXaA.png" /></figure><p>In the previous parts of this series, I introduced the field of GeoAI, its applications, and its most fundamental concepts — check them out if you missed them: <a href="https://medium.com/@julialeonardi/the-roadmap-of-geoai-part-1-e3fefa2575a9">Part 1</a>, <a href="https://medium.com/p/23d5167d93a6">Part 2</a>.</p><blockquote>This post is one half of a two-part tandem on the most popular GeoAI methods. While each stands alone, they complement each other well. For a fuller understanding, I recommend reading <a href="https://medium.com/@julialeonardi/traditional-ml-in-geoai-applications-the-roadmap-to-geoai-part-3a-ed670a1bb7fe">the other post first</a> and then getting to this one.</blockquote><h3>What to expect from this post</h3><p>This post and <a href="https://medium.com/@julialeonardi/traditional-ml-in-geoai-applications-the-roadmap-to-geoai-part-3a-ed670a1bb7fe">its traditional machine learning counterpart</a> aim to create a toolkit with a selection of the most fundamental machine learning models and deep learning architectures used in the geospatial context. In this part, we’ll dive into the deep-learning architectures most often used in geospatial applications.</p><p>Below is a table of the content structure. For each topic, I provide a brief explanation of the architecture and examples of GeoAI use cases that utilize the specific algorithm along with resources to learn more.</p><ul><li><a href="#7197"><strong>Deep Learning Methods — A Quick Introduction</strong></a></li><li><a href="#ab15"><strong>Into the Deep — the Multilayer Perceptron (MLP)</strong></a></li><li><a href="#3af7"><strong>Convolutional Neural Networks (CNN)</strong></a></li><li><a href="#4347"><strong>Transformers</strong></a></li><li><a href="#62c9"><strong>Recurrent Neural Networks</strong></a></li><li><a href="#9baf"><strong>Generative Models</strong></a></li><li><a href="#072a"><strong>Summary</strong></a></li></ul><h3>Deep Learning Methods — A Quick Introduction</h3><p>Deep learning is a branch of machine learning that has seen a significant rise in popularity in recent years. Many people tend to think only of deep learning when discussing machine learning. However, deep learning is a subset of methods built on the concept of artificial neural networks (ANNs) with multiple hidden layers. To understand more complex DL architectures, learning about these methods&#39; fundamental building blocks is crucial.</p><p>Deep learning architectures come in many forms, each built on distinct core concepts. In this post, I’ll introduce the most fundamental architectures, which are the foundation for more advanced, hybrid models. I’ve also organized models commonly found in geospatial research within the sections on these fundamental architectures so that you’ll understand the core concept they’re built on when you come across their names.</p><h4>Into the Deep — the Multilayer Perceptron (MLP)</h4><p>Let’s begin with the first and most basic deep learning model—the multi-layer perceptron (MLP). The MLP is based on the concept of an ANN with multiple hidden layers, as introduced in the previous post. It is built with the same components (inputs, weights, biases, activation functions, etc.) and trained based on forward and backward propagation mechanisms.</p><p>Because the information of such a network flows in one direction during the forward pass—from the input notes through the hidden nodes and to the output nodes—without any loops or cycles (in contrast to other architectures, such as recurrent neural networks), MLPs are a type of <strong>feedforward neural network.</strong></p><p>In an MLP, all layers are <strong>fully connected</strong>, meaning each node in a layer is connected to every node in the subsequent layers. While this is the simplest architecture of a deep artificial neural network, it is also one of the most widely used. This is because the principle of fully connected layers is a fundamental building block in more complex architectures, such as CNNs and transformers, which are often incorporated in specific parts of the model.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/647/1*QRKpexTehqTys6mPwjKVmQ.png" /></figure><p><strong>An important detail to note in the context of GeoAI is the input layer of an MLP.</strong> In an MLP, the input consists of individual nodes, each representing a feature of the data. For example, in determining whether an area is susceptible to floods, input nodes might represent features such as elevation, slope, predicted annual precipitation, and similar factors. This structure works well for data that can be represented in tabular format.</p><p>However, when working with images, such as satellite datasets, the process becomes more complex because each input neuron corresponds to a single pixel. This introduces two main challenges:</p><ol><li><strong>Loss of spatial context in the image:</strong></li></ol><p>Since MLPs require the input layer to be a 1D vector of features, images must first be flattened from their 2D matrix form into a 1D vector. This operation removes the spatial relationships between pixels — such as edges and textures — that are often critical in geospatial analysis.</p><p>2. <strong>High parameter count:</strong></p><p>Flattening an image significantly increases the number of input neurons, leading to a large number of parameters. For instance, say you have an RGB (3-channel) image of size 224 × 224 pixels. After flattening this image we will have <strong>3 * 224 * 224 = 150 528 input neurons.</strong> If we use a simple MLP architecture with one hidden layer of 7 neurons and an output layer of 1 neuron, the total number of parameters would be:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/861/1*r6vLzueJe-zpT8n8d5eAQg.png" /></figure><p><strong>These are more than a million parameters for a single small RGB image!</strong> In remote sensing applications, where large patches of high-resolution multispectral images are processed, this quickly becomes very computationally inefficient. These limitations highlight the importance of convolutional neural networks (CNNs) in geospatial applications, which we’ll explore in the next section.</p><p><strong>Use cases of the MLP in geospatial analysis:</strong></p><ul><li><a href="https://medium.com/@ce23resch12001/mapping-flood-susceptibility-using-random-forest-and-mlp-classifier-to-predict-vulnerable-areas-f376eebf5e18"><em>Mapping Flood Susceptibility</em></a><em> — </em>a study using MLP to predict vulnerable areas based on 12 flood conditioning factors. The article also clearly outlines the workflow of developing a GeoAI project.</li><li><a href="https://www.mdpi.com/2072-4292/13/20/4034"><em>Urban Sprawl Modeling</em></a><em> — </em>an example of how the MLP architecture is implemented in an urban sprawl prediction workflow in Baghdad, Iraq. The authors simulate several short and long-term development plans for the city.</li><li><a href="https://www.mdpi.com/2071-1050/11/4/975"><em>Coastal Erosion Prediction Modeling</em></a><em> — </em>the study utilizes MLP to build a model that predicts the shoreline movements in the dune-beach system of Costa da Caparica, Lisbon, Portugal.</li></ul><p><strong>Resources to learn more about MLP:</strong></p><ul><li><a href="https://www.youtube.com/playlist?list=PLZHQObOWTQDNU6R1_67000Dx_ZCJB-3pi"><strong><em>The 3Blue1Brown Neural Network series on YouTube</em></strong></a></li><li><a href="https://towardsdatascience.com/multilayer-perceptron-explained-with-a-real-life-example-and-python-code-sentiment-analysis-cb408ee93141"><strong><em>MLP Explained with a Real-Life Example</em></strong></a></li><li><a href="https://towardsdatascience.com/multilayer-perceptron-explained-a-visual-guide-with-mini-2d-dataset-0ae8100c5d1c"><strong><em>MLP, Explained: A Visual Guide with Mini 2D Dataset</em></strong></a></li></ul><h4>Convolutional Neural Networks (CNNs)</h4><p>CNNs have been the most essential architecture for image processing for the past couple of decades (<a href="https://towardsdatascience.com/the-history-of-convolutional-neural-networks-for-image-classification-1989-today-5ea8a5c5fe20#:~:text=In%201989%2C%20Yann%20LeCun%20and,capable%20of%20processing%20information%20spatially!">they were introduced by Yann LeCun and his team in 1989</a>). This is because, unlike MLPs, they capture the spatial context of the data thanks to the <strong>convolution operation</strong>.</p><p><strong>A convolution operation</strong> refers to the operation of sliding a filter (also known as the kernel) across the input image and creating a feature map based on the dot product of the filter matrix with the image patch on which the filter is. This operation retains the spatial context of each pixel, which makes processing images more efficient, as more essential information is preserved from the data. The values of the filter are the weights of the network, and they are learnable through the training process, which similarly to the MLP, is done through a feedforward step and a backpropagation step. This structure of the CNN architecture makes it also a <strong>feedforward neural network.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/864/1*liQNHftUjNCWGek_84aaiA.gif" /><figcaption>An example of a convolution operation on a 5x5 px, one band image with a 3x3 filter.</figcaption></figure><p>The other key concepts to learn to fully understand CNNs are pooling layers, padding, and the networks’ hyperparameters, which are the stride and size of the filter and the stride, size, and type of the pooling block.</p><p><strong>Pooling layers</strong> are the other important building blocks of a CNN, which are responsible for the dimensionality reduction of the feature maps created by the convolution layer. There are different types of pooling layers, but <strong>max pooling </strong>is the most widely used as it preserves the parts of the image that were most activated.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*S1nbmk2opi1Km9SdJQdmIA.gif" /><figcaption><em>A visualization of how a max pooling layer works.</em></figcaption></figure><p><strong>Padding</strong> is a fundamental operation used to preserve the spatial resolution of an image in the resulting feature map. Because convolution inherently reduces the image size, padding helps maintain control over the feature map’s dimensions by adding one or more rows of pixels. While zero padding (filling with 0s) is the most common, other types of padding are used, such as reflection or replication padding.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/291/1*ug2cTwk0iM_7a8dGk9U7dg.png" /><figcaption>Zero padding applied with a size = 1 (1 row and 1 column of padding added to each side of the image).</figcaption></figure><p>An MLP (fully connected layer) is usually added at the end of a CNN model to perform the final classification task based on the extracted spatial features of the image.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-lrTo6-2lYzz04T4kHTFJA.png" /><figcaption>A simplified scheme of a simple CNN.</figcaption></figure><p>Since its introduction in 1989, the CNN architecture has been adapted in numerous ways, providing many widely used variants. If you’re curious about the evolution of this influential architecture, its history is very nicely described <a href="https://towardsdatascience.com/the-history-of-convolutional-neural-networks-for-image-classification-1989-today-5ea8a5c5fe20">in this post</a>. Below, I list some of the most widely used CNN-based architectures, so that when you come across their names, you’ll immediately recognize them as CNN models and understand their foundational principles.</p><ul><li><strong>Vanilla CNNs</strong> (<em>e.g. AlexNet, LeNet</em>): These are foundational CNN architectures that introduced the core concepts of convolutional layers, pooling, and fully connected layers for tasks like image classification.</li><li><strong>Deep CNNs</strong> (<em>e.g. VGGNet, GoogLeNet</em>): These models introduced significantly increased network depth to capture hierarchical features, using many stacked convolutional layers with small kernel sizes.</li><li><strong>Residual Networks</strong> (<em>ResNets</em>): These architectures address the challenge of training very deep networks by introducing shortcut connections, enabling gradients to flow more effectively.</li><li><strong>Other CNN-based models that are often utilized in GeoAI applications:</strong><br>- <strong><em>U-Nets</em></strong>: Specialize in semantic segmentation tasks with an encoder-decoder structure and skip connections;<br>- <strong><em>YOLO (You Only Look Once)</em></strong>: Optimized for real-time object detection by formulating it as a single regression problem;<br>- <strong><em>Feature Pyramid Networks (FPNs)</em></strong>: Designed for object detection, adapting a pyramid of features for detecting objects at multiple scales.</li></ul><p><strong>Use cases of CNN architectures applied in geospatial analysis:</strong></p><ul><li><a href="https://www.mdpi.com/2073-445X/12/4/879"><em>Land Cover Mapping</em></a><em> — </em>This study utilizes CNNs on Sentinel-2 imagery to map the land cover of the Rome region.</li><li><a href="https://medium.com/@tony.aloysius.77/implementing-yolov8-for-aerial-satellite-image-building-segmentation-and-converting-to-shape-files-899b2a8affa5"><em>Aerial Satellite Image Building Segmentation</em></a><em> — </em>A great practical example of how the YOLO architecture can be implemented to extract specific information. The post comes with the code implementation and practical walkthrough of the whole project.</li><li><a href="https://www.researchgate.net/publication/339878980_Change_Detection_of_Deforestation_in_the_Brazilian_Amazon_Using_Landsat_Data_and_Convolutional_Neural_Networks"><em>Change Detection of Deforestation in the Brazilian Amazon</em></a><em> — </em>This study implements the CNN architecture to map deforestation using Landsat satellite data.</li></ul><p><strong>Resources to learn more about CNNs:</strong></p><ul><li><a href="https://www.youtube.com/watch?v=pj9-rr1wDhM"><strong><em>CNN Explained and Visualized</em></strong></a><strong><em> [YouTube]</em></strong></li><li><a href="https://medium.com/analytics-vidhya/cnn-convolutional-neural-network-8d0a292b4498"><strong><em>CNN vs MLP for Image Classification</em></strong></a><strong><em> [Medium]</em></strong></li><li><a href="https://medium.com/@prathammodi001/convolutional-neural-networks-for-dummies-a-step-by-step-cnn-tutorial-e68f464d608f"><strong><em>Convolutional Neural Networks for Dummies</em></strong></a> <strong><em>[Medium]</em></strong></li><li><a href="https://www.simplilearn.com/tutorials/deep-learning-tutorial/convolutional-neural-network"><strong><em>Convolutional Neural Network Tutorial</em></strong></a><strong><em> [SimpliLearn]</em></strong></li></ul><h4>Transformers</h4><p>Despite being a relatively recent development, Transformers have quickly emerged as a powerful architecture for geospatial data. While they don’t capture spatial context in the same way as CNNs, their <strong>attention mechanism</strong> makes them one of the most efficient and promising machine learning technologies today. Originally introduced in the groundbreaking 2017 paper <a href="https://arxiv.org/abs/1706.03762"><em>“Attention is All You Need”</em></a><em> </em>for natural language processing (NLP), Transformers have rapidly evolved to tackle computer vision and image processing tasks. For a deeper understanding of the architecture, I recommend reading the original paper and <a href="https://jalammar.github.io/illustrated-transformer/">this insightful blog</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/466/1*gmyp6wJyp6OPxYz_7fOpRw.png" /><figcaption>The Transformer architecture. Taken from “<a href="https://arxiv.org/abs/1706.03762">Attention Is All You Need</a>“.</figcaption></figure><p>In the context of image processing, the Transformer is adapted as the <strong>Visual Transformer (ViT)</strong>. This adaptation of the original architecture was introduced in the paper <a href="https://arxiv.org/abs/2010.11929"><em>“An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale”</em></a> in 2020. <strong>ViT uses only the encoder part</strong> of the original Transformer architecture and works by dividing the input image into fixed-size patches, flattening them, and feeding these patch embeddings into the encoder. The encoder is built by attention blocks, based on the attention mechanism — the foundational concept of transformers. The <strong>attention mechanism</strong> enables the model to capture long-range dependencies and contextual information across the image, allowing for a global understanding.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/822/1*3zTapmgyHiUsapcTCaF94Q.png" /><figcaption>The Vision Transformer architecture. Taken from <a href="https://arxiv.org/abs/2010.11929"><em>“An Image is Worth 16x16 Words”</em></a>.</figcaption></figure><p>ViT played a significant role in the rise of <strong>foundation models</strong> and <strong>transfer learning</strong>, which are fastly developing concepts also in geospatial applications.</p><p><strong>Use cases of ViTs in geospatial applications:</strong></p><ul><li><a href="https://www.nature.com/articles/s41598-024-67186-4"><em>Transformer-based land use and land cover classification</em></a><em> — </em>A study utilizing transformer models for Land Use and Land Cover (LULC) classification.</li><li><a href="https://www.mdpi.com/2072-4292/13/21/4441"><em>Building Extraction from Remote Sensing Images</em></a><em> — </em>A research exploring the potential of using transformers for efficient building extraction from aerial imagery.</li><li><a href="https://github.com/amanbasu/wildfire-detection"><em>Enhanced Wildfire Detection and Characterization</em></a><em> — </em>A GitHub repository with a project aiming to detect and characterize wildfires using ViTs and Landsat-8 data. The repository contains all the code in Python, which can be a great resource to see how GeoAI projects can be organized.</li></ul><p><strong>Resources to learn more about Transformers:</strong></p><ul><li><a href="https://towardsdatascience.com/transformers-141e32e69591"><strong><em>How Transformers Work</em></strong></a><strong><em> [Medium]</em></strong></li><li><a href="https://jalammar.github.io/illustrated-transformer/"><strong><em>The Illustrated Transformer</em></strong></a></li><li><a href="https://towardsdatascience.com/vision-transformers-explained-a9d07147e4c8"><strong><em>Vision Transformers, Explained (series)</em></strong></a><strong><em> [Medium]</em></strong></li><li><a href="https://www.v7labs.com/blog/vision-transformer-guide"><strong><em>Vision Transformers: What It Is &amp; How It Works</em></strong></a><strong><em> [v7labs]</em></strong></li><li><a href="https://medium.com/correll-lab/building-a-vision-transformer-model-from-scratch-a3054f707cc6"><strong><em>Building a Vision Transformer Model from Scratch</em></strong></a><strong><em> [Medum]</em></strong></li></ul><h4>Recurrent Neural Networks</h4><p>The role of Recurrent Neural Networks (RNNs) is significant in the machine learning landscape, especially in NLP applications, however, it has a more niche role in geospatial analysis compared to the widely used CNN and Transformers frameworks. RNNs’ core applications are around temporal and sequential data analysis, which is relevant for spatiotemporal time series, such as meteorological forecasting, climate modeling, or land use change monitoring. The RNNs were introduced in the 1980s to handle sequential data, where the output from the previous step is used as input for the next step, allowing the model to maintain temporal context. Their core concept revolves around <strong>feedback loops</strong>, enabling them to process sequences of varying lengths.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1rV1QaiUWgBlwdTR0pdcpA.png" /><figcaption>Schematic rapresentation of a simple RNN.</figcaption></figure><p>Although not the most popular architecture for geospatial data analysis, RNNs are important to understand because they are widely used for tasks involving time series, speech recognition, and natural language processing. It’s valuable to know how RNNs work, what they can be used for, and some of the key models that have been developed based on the RNN architecture, such as <strong>LSTMs</strong> (Long Short-Term Memory) and <strong>GRUs</strong> (Gated Recurrent Units).</p><p><strong>Use cases of RNNs applied to geospatial applications:</strong></p><ul><li><a href="https://arxiv.org/pdf/2111.08900"><em>Application to Crop Yield Prediction Using a Graph-Based RNN</em></a><em> — </em>A study harnessing the RNN architecture for predicting crop yields in the US mainland based on data from 1981 to 2019. The model is trained on various factors, such as weather, land surface, and soil quality.</li><li><a href="https://www.mdpi.com/2073-445X/10/3/282"><em>Exploring the Sensitivity of RNNs for Forecasting Land Cover Change</em></a><em> — </em>The study evaluates the performance of LSTM models for forecasting land cover changes by conducting a sensitivity analysis involving hypothetical and real-world datasets.</li><li><a href="https://d1wqtxts1xzle7.cloudfront.net/32172913/Sauter_et_al_-_Spatio-temporal_prediction_of_snow_cover_in_the_Black_Forest-libre.pdf?1391553770=&amp;response-content-disposition=inline%3B+filename%3DSpatio_temporal_prediction_of_snow_cover.pdf&amp;Expires=1733691118&amp;Signature=JK19qK~1CWiKEyJvdvG5dvLBGYW9G5OgxEX~6RaLN751Wzd4c7ERSIIT-qwf6WV-3znCUqD4qmuCJtWavCsg0fQeNLyYOJ2alrzNupivJXJB45wVtyemGvZmiHExlD96niSMwGFxY42aN31Yda-A2VKLAHtQWCwB2hTWN7pIqBpfroYkkJPGiVefaCdtVN7khTce71YPkmgwnToWYaEJhDpyC6o7E7QTIsYL8ZjUBqpN6U789MajPVAD9TOp2yRvU9RLt~SihvKpKGu-fo-F040gfnUohF2XrpBSJNdFYOG7St-G76KwS288T4hgcg-7ILMwWq-fTzEm9UaOTY34FA__&amp;Key-Pair-Id=APKAJLOHF5GGSLRBV4ZA"><em>Prediction of snow cover in the Black Forest mountain range using RNNs</em></a><em> — </em>In this study, the authors utilized snow cover derived from MODIS satellite data with meteorological data to develop a snow depth forecasting system using an RNN approach.</li></ul><p><strong>Resources to learn more about RNNs:</strong></p><ul><li><a href="https://www.geeksforgeeks.org/introduction-to-recurrent-neural-network/"><strong><em>Introduction to RNNs</em></strong></a><strong><em> [Geeks4Geeks]</em></strong></li><li><a href="https://towardsdatascience.com/learn-how-recurrent-neural-networks-work-84e975feaaf7"><strong><em>How Recurrent Neural Networks Work</em></strong></a><strong><em> [Medium]</em></strong></li><li><a href="https://colah.github.io/posts/2015-08-Understanding-LSTMs/"><strong><em>Understanding LTSM Network</em></strong></a></li><li><a href="https://towardsdatascience.com/understanding-gru-networks-2ef37df6c9be"><strong><em>Understanding GRU Networks</em></strong></a><strong><em> [Medium]</em></strong></li></ul><h4>Generative models</h4><p>Generative models create a subfield of ML designed to generate new data samples based on the distribution of a given dataset. In geospatial analysis, they are used in data augmentation tasks, cloud removal, super-resolution enhancement, or anomaly detection. Their usage is less prominent than the usage of models from the CNN or Transformers family, but quite important for task-specific applications. The importance of generative models is evolving quickly with the introduction of transformers, which revolutionized their capabilities due to their scalability, multimodal inputs, and ability to model long-range dependencies. The most important generative models that a GeoAI practitioner should be aware of are:</p><ul><li>GANs (Generative Adversarial Networks) — The goal of GANs is to autonomously identify patterns in input data, which enables the model to produce new examples that resemble the original dataset. They are made up of two neural networks: <strong>a discriminator</strong> and <strong>a generator</strong>. To learn more check out <a href="https://www.geeksforgeeks.org/generative-adversarial-network-gan/">this post</a>.</li><li>VAEs (Variational Autoencoders) — A VAE is made up of an <strong>encoder </strong>and a <strong>decoder. </strong>The idea behind encoders in VAE is to generate encodings that are useful for reconstructing its input. The ideas behind VAEs are quite more complex, so if you’re interested in exploring them in more depth I encourage you to read more in posts like <a href="https://www.geeksforgeeks.org/variational-autoencoders/">this one</a>.</li></ul><p><strong>Use cases of generative models applied to geospatial analysis:</strong></p><ul><li><a href="https://arxiv.org/html/2402.14049v2"><em>Generative Adversarial Models for Extreme Geospatial Downscaling</em></a><em> — </em>This research paper utilizes a GAN architecture to generate high-resolution accurate climate datasets from very low-resolution inputs.</li><li><a href="https://towardsdatascience.com/erasing-clouds-from-satellite-imagery-using-gans-generative-adversarial-networks-2d7f8467ef2e"><em>Erasing Clouds from Satellite Imagery Using GANs </em></a>— This is a great resource showing how GANs are built and how exactly they work on the cloud removal task. This post has all the implementation and a great explanation of the concepts.</li><li><a href="https://www.tandfonline.com/doi/full/10.1080/01431161.2022.2048915"><em>Cloud Removal from Satellite Imagery Using Multispectral Edge-Filtered Conditional GANs</em></a><em> — </em>A scientific publication on the same topic as the previous example with a proposed GAN-based architecture using multispectral satellite data.</li></ul><p><strong>Resources to learn more about Generative models:</strong></p><ul><li><a href="https://medium.com/analytics-vidhya/an-introduction-to-generative-deep-learning-792e93d1c6d4"><strong><em>An Introduction to Generative Deep Learning</em></strong></a><strong><em> [Medium]</em></strong></li><li><a href="https://jonathan-hui.medium.com/gan-whats-generative-adversarial-networks-and-its-application-f39ed278ef09"><strong><em>GAN — What is Generative Adversarial Networks</em></strong></a><strong><em> [Medium]</em></strong></li><li><a href="https://towardsdatascience.com/intuitively-understanding-variational-autoencoders-1bfe67eb5daf"><strong><em>Intuitively Understanding Variational Autoencoders</em></strong></a><strong><em> [Medium]</em></strong></li></ul><h3>Summary</h3><p>In this two-part series on machine learning methods for GeoAI applications, I’ve outlined some of the most essential models shaping the field. My goal was to create a practical toolkit for anyone interested in building their own GeoAI solutions, complete with concise introductions, examples, and additional resources for each method. If you think I’ve missed any important approaches, have questions, or simply want to share your thoughts, feel free to leave a comment. <strong>Let’s continue learning together — see you in the next post!</strong> 🐸✨</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bc60ca96b266" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Traditional ML in GeoAI applications (The Roadmap to GeoAI — Part 3a)]]></title>
            <link>https://medium.com/@julialeonardi/traditional-ml-in-geoai-applications-the-roadmap-to-geoai-part-3a-ed670a1bb7fe?source=rss-7297bf51f197------2</link>
            <guid isPermaLink="false">https://medium.com/p/ed670a1bb7fe</guid>
            <category><![CDATA[geoinformatics]]></category>
            <category><![CDATA[geoai]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[remote-sensing]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Julia Leonardi]]></dc:creator>
            <pubDate>Tue, 10 Dec 2024 10:23:20 GMT</pubDate>
            <atom:updated>2024-12-10T10:26:46.676Z</atom:updated>
            <content:encoded><![CDATA[<h3>Traditional ML in GeoAI applications (The Roadmap to GeoAI — Part 3a)</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sI9QbzaveyZ7nKXbi5qXaA.png" /></figure><p>In the previous parts of this series, I introduced the field of GeoAI, its applications, and its most fundamental concepts—check them out if you missed them: <a href="https://medium.com/@julialeonardi/the-roadmap-of-geoai-part-1-e3fefa2575a9">Part 1</a>, <a href="https://medium.com/p/23d5167d93a6">Part 2</a>.</p><blockquote>This post is one half of a two-part tandem on the most popular GeoAI methods. While each stands alone, they complement each other well. I recommend reading this post first, f<a href="https://medium.com/@julialeonardi/deep-learning-in-geoai-applications-the-roadmap-to-geoai-part-3b-bc60ca96b266">ollowed by the other post</a>, for a fuller understanding.</blockquote><h3>What to expect from this post</h3><p>This post and <a href="https://medium.com/@julialeonardi/deep-learning-in-geoai-applications-the-roadmap-to-geoai-part-3b-bc60ca96b266">its deep learning counterpart</a> aim to create a toolkit with a selection of the most fundamental machine learning models and deep learning architectures used in the geospatial context. In this part, we’ll start with classical, “shallow” learning methods and progress into more complex, deep learning models in the corresponding article.</p><p>Below is a table of the content structure. For each topic, I provide a brief explanation of the method and examples of GeoAI use cases that utilize the specific algorithm.</p><ul><li><a href="#8fca"><strong>‘Shallow’ Learning Methods — A Quick Introduction</strong></a></li><li><a href="#d760"><strong>Clustering</strong></a><br>- <em>K-means</em></li><li><a href="#038e"><strong>Classification and Regression</strong></a><strong> </strong><br>- <em>K-Nearest Neighbors (KNN)<br>- Decision Tree (DT)<br>- Random Forest (RF)<br>- Support Vector Machine (SVM)</em></li><li><a href="#f906"><strong>Dimensionality Reduction</strong></a><br>- <em>Principal Component Analysis (PCA)<br>- Minimum Noise Fraction (MNF)</em></li><li><a href="#a353"><strong>Pre-gaming the Deep Learning methods</strong></a><br>- <em>Perceptron<br>- Single-layer Artificial Neural Network</em></li></ul><h3>‘Shallow’ Learning Methods — A Quick Introduction</h3><p>Traditional or ‘shallow’ machine learning methods refer to models that do not rely on artificial neural networks (ANNs) with multiple hidden layers. Although they receive less attention in the era of deep learning, traditional machine learning methods remain the foundation of the AI field. These methods are efficient, often state-of-the-art for specific tasks, and can outperform deep learning in scenarios with limited data or simpler problems where deep learning might be overkill. Below, I introduce a selection of the most important and widely used methods in geospatial data analysis.</p><h4>K-means</h4><p><strong>Clustering </strong>aims at grouping the data in an unsupervised manner and it’s widely used in geospatial data analysis. It’s often used to analyze the patterns in the data and detect outliers. K-means clustering is the most widely used clustering algorithm, and its mechanism is based on partitioning the data in <strong><em>k </em></strong>clusters, such that each data point belongs to the cluster with the nearest mean.</p><p>The algorithm first initializes <strong><em>k</em></strong> means (randomly or using a predefined rule), then assigns the data points to the closest mean, forming the <strong><em>k</em></strong> distinct clusters, and calculates the centroids of each of the newly formed clusters, which become the new mean — these steps are repeated until convergence has been reached.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/864/1*q2dr3YWjNlvHOwAIQ_hwOg.gif" /><figcaption>K-means visualization</figcaption></figure><p><strong>Use cases:</strong></p><ul><li><em>Land cover classification, by grouping pixels based on spectral similarity </em>as in <a href="https://towardsdatascience.com/unsupervised-land-cover-classification-with-python-8e3734dc3c6d">this paper</a>.</li></ul><blockquote><strong>Land cover classification is one of the fundamental tasks in remote sensing and will be the standard use case for most of these methods.</strong></blockquote><ul><li><em>Regional analysis, by clustering regions based on environmental, socioeconomic, or natural hazard susceptibility factors </em>such as <a href="https://www.mdpi.com/2076-3417/13/14/8123">in this study</a>, which uses geospatial data with k-means to model acoustic regions<em>;</em></li><li><em>Urban planning by identifying homogeneous zones in cities, such as residential, commercial, etc. </em>An exercise of this type is presented <a href="https://www.iaacblog.com/programs/k-means-clustering-for-urban-typologies/">here</a>, where k-means is used to extract different urban typologies based on different buildings’ characteristics.</li></ul><h4>K-Nearest Neighbors (KNN)</h4><p>K-nearest neighbor <strong>classification </strong>is a supervised and non-parametric classification algorithm, meaning that it needs a labeled dataset to train on and does not make any assumptions about the data distribution. It classifies data points based on their <strong>k</strong> closest neighbors.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/621/1*GtBBV_29MZpQnPQ5NFPjuA.png" /></figure><p>It’s based on the assumption that similar points can be found near one another, which coincides perfectly with the <strong>First Law of Geography</strong>:</p><blockquote>“Everything is related to everything else, but near things are more related than distant things”<em> </em><br><strong> — Waldo Tobler, 1969</strong></blockquote><p>This is why KNN is simple, yet powerful for geospatial analysis.</p><p><strong>Use cases:</strong></p><ul><li><em>Land cover classification of remote sensing imagery </em>as <a href="https://towardsdatascience.com/land-cover-classification-in-satellite-imagery-using-python-ae39dbf2929">in this example</a>.</li><li><em>Geospatial interpolation of missing values by imputation. </em>A good example of such a use case can be found <a href="https://www.linkedin.com/pulse/k-nearest-neighbors-imputation-understanding-geo-data-daniel-smith/">in this exercise</a>, where the population of missing ZIP codes in the data was interpolated using KNN.</li></ul><h4>Decision Tree (DT)</h4><p>Decision Tree is another non-parametric supervised learning method that can be used both for classification and regression. They split the data into subsets based on feature values to make classification or regression decisions. In geospatial analysis, they can be used either way depending on the data and task you have at hand. To learn more about the mechanism behind the DT, check out <a href="https://www.geeksforgeeks.org/decision-tree/">this article</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2Z-mm-700OLUPRIskAuULA.png" /><figcaption>A simplified example of a decision tree</figcaption></figure><p><strong>Use cases:</strong></p><ul><li><em>Habitat suitability analysis for a specific species based on environmental variables, such as type of land cover, vicinity of urban areas, etc.</em>, like in <a href="https://www.researchgate.net/publication/222574789_Habitat_suitability_modelling_for_red_deer_Cervus_elaphus_L_in_South-central_Slovenia_with_classification_trees">this study</a> on habitat suitability modeling for red deer in Slovenia.</li><li><em>Classification of regions into landslide risk categories based on the geomorphology of the area, vicinity of rivers, and other environmental factors </em>such as done <a href="https://onlinelibrary.wiley.com/doi/10.1155/2014/929768">in this study, </a>which specifies the importance of 21 factors causing landslides in the region of Penang Island, Malaysia.</li></ul><h4>Random Forest (RF)</h4><p>Random Forest is predominantly used as a classification algorithm in geospatial applications, though it can also be used in regression tasks. It is a supervised ensemble learning method, which means that it uses multiple weaker models to base its decision on. In this case, RF uses multiple Decision Trees to make decisions based on majority voting. Ensemble Learning is an important branch of ML, which is comprised of two fundamental concepts: <strong>Bagging </strong>and Boosting. The RF algorithm is based on the concept of Bagging, also known as bootstrap aggregation. The basic idea of the algorithm is based on sampling with replacement <strong><em>n </em></strong>random data and feature subsets from the training set and then training <strong><em>n </em></strong>independent DTs on those samples, thus creating an uncorrelated forest of decision trees.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0aDESHuilf9pQ19I34_r6A.png" /><figcaption>Random forest is built on many DTs trained with different subsets of the training data.</figcaption></figure><p>To learn more about RF <a href="https://towardsdatascience.com/random-forest-explained-a-visual-guide-with-code-examples-9f736a6e1b3c">check out this post</a>.</p><p><strong>Use cases:</strong></p><ul><li><em>Land cover classification. </em>Many resources and scientific articles use RF for land cover classification,<a href="https://medium.com/land-use-land-cover-classification-using-google/land-use-land-cover-classification-using-google-earth-engine-random-forest-algorithm-landsat-and-28b610f555de"> here I link an example of the use case that can be implemented in Google Earth Engine</a>.</li><li><em>Hazard susceptibility mapping. </em>An example is a mapping of landslide susceptibility such as <a href="https://www.tandfonline.com/doi/full/10.1080/17538947.2024.2346263#abstract">shown in this paper,</a> where the authors use RF with data including landslide-influencing factors and historical landslide records to create maps of areas especially prone to such events.</li><li><em>Modeling of soil properties</em>, as done <a href="https://www.researchgate.net/publication/343822164_Spatial_prediction_of_major_soil_properties_using_Random_Forest_techniques_-_A_case_study_in_semi-arid_tropics_of_South_India">in this article</a>, which maps the pH, electrical conductivity, and organic carbon of the soil in Bukkarayasamundram Mandal, India. The RF model was trained on many environmental factors such as land use, landform, and slope maps, the Digital Elevation Model (DEM) of the area, and spectral indices extracted from remote sensing data.</li></ul><h4>Support Vector Machine (SVM)</h4><p>Support Vector Machine is another powerful classifier often used in the geospatial field. The idea behind SVM is to find the hyperplane that best separates data into different classes. What makes SVMs so powerful is the concept of mapping the data into a higher-dimensional feature space using mathematical functions called <strong>kernels. </strong>The right kernel can make the data linearly separable in the mapped dimension. Explore <a href="https://towardsdatascience.com/support-vector-machine-explained-8bfef2f17e71">this article</a> to learn more!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/924/1*gGODGumo2e6oxkWJHB6Yxg.png" /><figcaption>A simple linear classification example of the SVM</figcaption></figure><p>In geospatial analysis for classification tasks SVM and RF are considered state-of-the-art methods.</p><p><strong>Use cases:</strong></p><ul><li><em>Monitoring the evolution of glacier mass changes </em>by automatically identifying the snow line on glaciers using SVM to classify ice and snow described <a href="https://www.mdpi.com/2072-4292/14/16/3868">in this article</a>.</li><li><em>Crop classification in agricultural regions </em>using hyperspectral data, as presented <a href="https://www.researchgate.net/publication/226556349_Support_Vector_Machines_for_Crop_Classification_Using_Hyperspectral_Data">in this work</a>, where 6 different crop types were classified using SVM with different kernel functions.</li></ul><blockquote><strong>I encourage you to use any of the above-mentioned use cases as inspiration to develop your small project to showcase in your GeoAI portfolio 🤗</strong></blockquote><h4>Dimensionality Reduction Methods</h4><p>With the rising amount of precise and rich data, especially in remote sensing, there’s a need to take into account the <a href="https://www.datacamp.com/blog/curse-of-dimensionality-machine-learning">curse of dimensionality</a>. Dimensionality reduction techniques aim to reduce the complexity of high-dimensional datasets by mapping the data to a lower-dimensional space, often transforming the original features into new ones. In remote sensing, especially with hyperspectral imagery containing hundreds of spectral bands, dimensionality reduction helps manage the computational challenges of processing all bands.</p><p>There’s a variety of dimensionality reduction methods, of which the most widely used in the geospatial context are:</p><ul><li><strong>PCA </strong>—Principal Component Analysis is the most well-known dimensionality reduction algorithm. It’s an unsupervised learning method, which finds a <strong>new</strong>, smaller set of features, retaining most of the sample’s information based on the variance in the original dataset. Learn more about the details behind PCA <a href="https://www.geeksforgeeks.org/principal-component-analysis-pca/">here</a>.</li><li><strong>MNF </strong>— Minimum Noise Fraction builds upon the concept of PCA with an additional noise reduction step, which is especially crucial for remote sensing data, making it a domain-specific dimensionality reduction method. You can find more information about MNF <a href="https://www.nv5geospatialsoftware.com/Support/Maintenance-Detail/2763">here</a>.</li></ul><h3>Pre-gaming the Deep Learning Methods</h3><p>Before exploring deep learning methods in the GeoAI field in the sister article, it’s important to introduce the foundational building blocks. These models, often categorized as “shallow” machine learning, are integral to understanding modern “deep” approaches. While their standalone applications are not of the most significant relevance anymore, they remain crucial in the context of deep learning, serving as the basis for more advanced architectures and methods.</p><h4>Perceptron</h4><p>The perceptron, often called the simplest form of an artificial neural network, is a linear, binary classifier. It was introduced by <strong>Frank Rosenblatt</strong> in the 1950s, and it laid the foundations of the ML field of artificial neural networks. It consists of input values, weights, and biases, a linear transformation using the inputs, weights, and biases, and a <strong>unit step activation function</strong>, which gives the output.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/739/1*8XiHErY_KnUWV3JIbdqdAw.png" /><figcaption>A simple visualization of the perceptron</figcaption></figure><p>To get a better understanding of what a perception is and how it works, check out these resources:</p><ul><li><a href="https://towardsdatascience.com/what-the-hell-is-perceptron-626217814f53">What the Hell is Perceptron?</a></li><li><a href="https://www.geeksforgeeks.org/what-is-perceptron-the-simplest-artificial-neural-network/">What is Perceptron | The Simplest Artificial neural network</a></li><li><a href="https://towardsdatascience.com/what-is-a-perceptron-basics-of-neural-networks-c4cfea20c590">What is a Perceptron? — Basics of Neural Networks</a></li></ul><h4>Single-layer ANN</h4><p>A single-layer artificial neural network is the natural extension of the concept of the perceptron — you can think of perceptrons as a specific case of a single-layer ANN. The difference lies in the flexibility of choice of the <strong>activation function</strong> (it can be a sigmoid, ReLU, or softmax). This simple architecture is the basic building block of all deep learning architectures, which connect the ANN layers to create complex models. This is why I believe it’s fundamental to understand how a single-layer, simple ANN works!</p><p>Before moving on, consider spending some time reading and thinking about the crucial concepts related to ANNs:</p><ul><li><em>ANN components: </em><a href="https://towardsdatascience.com/whats-the-role-of-weights-and-bias-in-a-neural-network-4cf7e9888a0f"><em>inputs, weights and biases</em></a><em>, </em><a href="https://towardsdatascience.com/activation-functions-neural-networks-1cbd9f8d91d6"><em>activation functions</em></a><em>;</em></li><li><a href="https://towardsdatascience.com/forward-propagation-in-neural-networks-simplified-math-and-code-version-bbcfef6f9250"><em>Forward propagation</em></a><em>;</em></li><li><a href="https://towardsdatascience.com/what-is-loss-function-1e2605aeb904"><em>Loss functions</em></a><em>;</em></li><li><a href="https://towardsdatascience.com/understanding-backpropagation-abcc509ca9d0"><em>Backward propagation</em></a><em>;</em></li><li><a href="https://towardsdatascience.com/optimizers-for-training-neural-network-59450d71caf6"><em>Optimization functions</em></a><em>.</em></li></ul><p>Here’s a good explanation of most of these terms: <a href="https://buse-koseoglu13.medium.com/one-layer-neural-network-from-scratch-classification-b6c71481f992">One Layer NN From Scratch</a>.</p><p>Once you master the fundamentals of how a single-layer ANN works with the concepts above you can expand that knowledge to bigger, more complex deep learning models. <strong>Happy learning — see you in the next post! 🐸✨</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ed670a1bb7fe" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Roadmap of GeoAI — Part 2]]></title>
            <link>https://medium.com/@julialeonardi/the-roadmap-of-geoai-part-2-23d5167d93a6?source=rss-7297bf51f197------2</link>
            <guid isPermaLink="false">https://medium.com/p/23d5167d93a6</guid>
            <category><![CDATA[geospatial-data-science]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[geoinformatics]]></category>
            <category><![CDATA[geoai]]></category>
            <dc:creator><![CDATA[Julia Leonardi]]></dc:creator>
            <pubDate>Tue, 22 Oct 2024 12:46:40 GMT</pubDate>
            <atom:updated>2024-10-22T12:46:40.084Z</atom:updated>
            <content:encoded><![CDATA[<h3>The Roadmap of GeoAI — Part 2</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sI9QbzaveyZ7nKXbi5qXaA.png" /></figure><p><a href="https://medium.com/@julialeonardi/the-roadmap-of-geoai-part-1-e3fefa2575a9">In the first part of this series</a>, I presented a broad overview of GeoAI with examples of how it could be applied in practice. In the end, I introduced a mindmap of terms related to the Geo and the AI fields, which will be the basis of today&#39;s post.</p><p>In this part, I’ll review the fundamental concepts you should learn depending on your learning path. This post will be more technical, introducing key concepts and demonstrating their relevance in GeoAI workflows.</p><h4>Table of Contents</h4><ul><li><strong>Geospatial Data: The Foundation of GeoAI<br> → Vector and Raster Data<br> → Remote Sensing Data</strong></li><li><strong>AI Fundamentals<br> → The Supervised, Unsupervised, and Reinforcement Learning Split<br> → GeoAI Tasks: Applying Intelligence to Geospatial Data</strong></li><li><strong>Closing Thoughts for Now</strong></li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*y2Y8WpFd-6LP315NXaSQwg.png" /><figcaption>Mindmap of concepts related to GeoAI.</figcaption></figure><h3>Geospatial Data: The Foundation of GeoAI</h3><p>Working with AI applications, including GeoAI, you need to know your data, and you need to know it well. Data is the most crucial component of any AI workflow — without it, even the best model architecture won’t learn anything. But it’s not just about having data; it’s about having good, clean data that aligns with your specific goals. If not, your model’s performance will suffer — ‘<a href="https://shelf.io/blog/garbage-in-garbage-out-ai-implementation/">Garbage In, Garbage Out</a>’.</p><p>While it would be ideal to find a curated, analysis-ready dataset that perfectly suits your needs, that’s rarely the case in the geospatial field. Often, you’ll need to build your dataset. Sure, building your AI sounds very fancy and cool, but the reality is that most of your time will be spent preparing the data your model will train on. And to create a great dataset, you really have to know your data. With that, let’s dive in!</p><h4>Vector and Raster Data</h4><p>Geospatial data includes a wide variety of data such as remote sensing imagery, cadastral maps, traffic data, or weather maps and it comes in different shapes and forms. The most important distinction being between <strong>vector </strong>and <strong>raster data</strong>. Understanding this difference is crucial, as it influences how you prepare your data for AI models.</p><ul><li><strong>Vector data</strong> is structured as points, lines, and polygons represented by a series of (x, y) coordinates. Vectors represent discrete data, meaning their file sizes are smaller and easier to manage. They can be used to store any real world features in the form of points, lines, or polygons such as place names, rivers, and boundaries respectively. The most commonly used file formats associated with vector data are <a href="https://en.wikipedia.org/wiki/Shapefile">Shapefile (*.shp)</a>, <a href="https://en.wikipedia.org/wiki/GeoJSON">GeoJSON (*.geojson)</a>, and <a href="https://en.wikipedia.org/wiki/GeoPackage">GeoPackage (*.gpkg)</a>, which is a file format for both vector and raster data.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/333/1*tbj2_QSrVMRTyfqnKIWlHw.png" /><figcaption>A vector polygon</figcaption></figure><ul><li><strong>Raster data</strong>, on the other hand, is structured in the form of a matrix of pixels, each of which represents a value for the pixel location. Rasters are used to represent continuous features, such as temperature, elevation, or land use. This form is used for remote sensing acquisitions (optical data, SAR data), digital elevation models (DEMs), or historical scanned maps. Raster data is usually stored in <a href="https://en.wikipedia.org/wiki/GeoTIFF">GeoTIFFs (*.tif)</a>, <a href="https://en.wikipedia.org/wiki/JPEG_2000">JPEG 2000 (*.jp2)</a>, or <a href="https://en.wikipedia.org/wiki/NetCDF">NetCDF (*.nc)</a>, to name a few.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/436/1*ZtdFacCbPoOZKx7FuBKiRA.png" /><figcaption>A raster layer</figcaption></figure><p>These two data types serve as the input for Machine Learning models in GeoAI workflows, whether it’s for land cover classification, environmental monitoring, or urban planning.</p><h4>Remote Sensing Data</h4><p>The rapid evolution of remote sensing data makes it one of the most interesting and significant types of geospatial data. It involves collecting data without making physical contact with the observed object or phenomenon, using various sensors.</p><p>Remote sensing acquisitions can be categorized based on the <strong>sensor type </strong>and the <strong>platform</strong> carrying it.</p><ul><li><strong>Platforms</strong> are where the data is acquired from. They can be spaceborne (from satellites), airborne (from aircraft or drones), or ground-based (such as a weather station). Each platform offers different advantages in terms of spatial coverage and resolution.</li><li><strong>Sensors</strong> can be either passive or active. <em>Passive sensors</em> measure the energy naturally available, i.e. the sun’s energy, that is reflected or re-emitted by the observed object. This means that such sensors can acquire information only during hours when that energy is available. <em>Active sensors</em>, on the other hand, emit their energy, which they then measure after it’s reflected from the observed target. This mean that they can acquire data independently on the naturally available energy.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*N0THS3w-VcKptucm7yGcIg.png" /><figcaption>Simplified visualization of the remote sensing data acquisitions</figcaption></figure><p>Remote sensing data is usually stored in raster formats with multiple layers of information, making it a core component in many geospatial applications, from environmental monitoring to disaster management.</p><h3>AI Fundamentals</h3><p>Knowing what we can expect from the “Geo” data, let’s explore how AI techniques interact with geospatial data. In GeoAI, the choice of machine learning methods depends on the task you’re addressing.</p><h4><strong>The Supervised, Unsupervised, and Reinforcement Learning Split</strong></h4><p>In general, machine learning methods can be classified into three main groups:</p><ul><li><strong>Supervised learning </strong>models are trained on labeled data, meaning that each training input is paired with its expected output (the “label”). These methods are widely used in geospatial tasks like land cover classification, where the type of land (e.g., forest, water, urban) is already known for certain images, and the goal is to learn to classify the same for new images.</li><li><strong>Unsupervised learning </strong>models work with unlabeled data, which means that it doesn’t receive the expected “solution” of the task. Such models attempt to find patterns in the input data itself. Clustering satellite images based on pixel similarities is an example of unsupervised learning in geospatial analysis.</li><li><strong>Reinforcement learning </strong>is a set of methods that learn to make decisions through trial and error, improving their actions based on feedback from the environment. The goal is to maximize rewards or positive outcomes over time. It is commonly used in training models for tasks like playing games such as chess or Go.</li></ul><p>In the context of GeoAI, reinforcement learning is the least used subfield of AI so in the next part of this post, I will mainly focus on tasks that are supervised, unsupervised, or both. The machine learning tasks that will be introduced are the most widely used in geospatial data analysis.</p><h4>GeoAI Tasks: Applying Intelligence to Geospatial Data</h4><p>Now that we know the types of machine learning methods, let’s move on to specific AI tasks that are commonly applied in the geospatial context.</p><p><strong>Classification</strong> is the most frequently performed task within the geospatial domain. It is a supervised learning task where the goal is to assign discrete labels to the data. In geospatial analysis, classification can be used to identify flood-prone areas or to classify land cover types (e.g., forest, urban, water) from satellite imagery. There are many types of classification: scene classification, pixel-based classification, object-based classification, and they each have slightly different aims. This distinction is beyond the scope of this post and will be covered in future ones.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/311/1*WsW4aztw366BJHn41__b8g.png" /><figcaption>Example of <strong>scene classification</strong> of a coastal area satellite image</figcaption></figure><p><strong>Regression </strong>is another supervised task, but instead of predicting labels, it predicts continuous values. In the context of GeoAI, this might involve predicting soil moisture levels, canopy height, or biomass from remote sensing data.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CyVMoXKXQBkzjoIeHRf_hw.png" /><figcaption>Example of soil moisture prediction based on a satellite image. Credit: <a href="https://medium.com/@mahyar.aboutalebi/is-it-possible-to-estimate-soil-moisture-from-satellite-images-if-yes-how-python-56f6aa73399e">https://medium.com/@mahyar.aboutalebi/is-it-possible-to-estimate-soil-moisture-from-satellite-images-if-yes-how-python-56f6aa73399e</a></figcaption></figure><p><strong>Segmentation </strong>is an image analysis task where the goal is to partition an image into meaningful regions. It is closely related to classification but focuses on grouping similar pixels. This task can be performed using both supervised and unsupervised methods. An example of segmentation on geospatial data is grouping satellite image pixels to differentiate between water bodies, forests, and urban areas. Segmenatation can also be differentiated in different types, such as semantic segmentation and instance segmentation. This disctinction is, again, going to be a part of a seperate post.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/386/1*3oCUNbA8FfEgWwbLDqd2Gg.png" /><figcaption>Example of <strong>semantic segmentation </strong>on a coastal area satellite image</figcaption></figure><p><strong>Change detection </strong>focuses on identifying differences between multiple geospatial datasets taken over time. This task is crucial in monitoring environmental changes such as deforestation, urban expansion, or disaster impacts.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uEl6DnjtUr4dIaXBz0uxfQ.png" /><figcaption>Example of <strong>landslide change detection</strong> from a multitemporal Sentinel-2 dataset</figcaption></figure><p><strong>Object detection</strong> aims at locating specific objects in remote sensing images. Unlike classification or segmentation, which label every pixel, object detection identifies and localizes distinct objects, like ships or buildings, using bounding boxes. In geospatial analysis, it can be used in detecting ships in satellite images for maritime monitoring.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/532/1*ju9hamtjCsIHFAA3tJlRtA.png" /><figcaption>Example of<strong> ships detection</strong> for maritime monitoring</figcaption></figure><p>Each AI task — whether it’s classification, regression, or change detection — can be approached using different machine learning methods. For example, classification often relies on supervised learning, while segmentation may use both supervised and unsupervised techniques. Change detection could leverage either supervised or unsupervised methods depending on the availability of labeled data. The key is selecting the right approach based on the data and problem at hand.</p><h3>Closing Thoughts for Now</h3><p>Each of the introduced concepts deserved its own blog post (or even a book!), but for now, we’ve established the foundation by introducing the core terminology from both the geospatial and AI domains. As you can see, mastering GeoAI involves navigating a wide array of concepts.</p><p>In the next post, we’ll dive deeper into the specific machine learning <strong>models </strong>most commonly used in GeoAI.</p><p>Stay tuned, and don’t hesitate to leave a comment if you have any questions or suggestions. Let’s learn together! ✨🐸</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=23d5167d93a6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Roadmap of GeoAI — Part 1]]></title>
            <link>https://medium.com/@julialeonardi/the-roadmap-of-geoai-part-1-e3fefa2575a9?source=rss-7297bf51f197------2</link>
            <guid isPermaLink="false">https://medium.com/p/e3fefa2575a9</guid>
            <category><![CDATA[earth-observation]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[geoai]]></category>
            <category><![CDATA[geoinformatics]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Julia Leonardi]]></dc:creator>
            <pubDate>Fri, 11 Oct 2024 21:20:32 GMT</pubDate>
            <atom:updated>2024-10-11T21:20:32.803Z</atom:updated>
            <content:encoded><![CDATA[<h3>The Roadmap of GeoAI — Part 1</h3><p>In this series of posts, I’ll introduce the field of GeoAI by outlining its core concepts, applications, and the main study paths for those looking to get into the field. This introductory post will provide a high-level overview, laying the foundations without diving into the finer details of the specific concepts — those will be explored in future posts. So, without further ado, let’s get into the intro!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sI9QbzaveyZ7nKXbi5qXaA.png" /></figure><h3>What is GeoAI?</h3><p>GeoAI is an emerging subfield at the intersection of Geoinformatics and Artificial Intelligence. It’s an exciting domain rapidly gaining relevance as spatial data — particularly remote sensing data — evolves and AI research advances at a fast pace. Since GeoAI is still a young term, there can be ambiguity about what exactly it encompasses and what the core concepts, applications, methods, and required skills are.</p><p>As I’m exploring the GeoAI landscape, I decided to organize the fundamental ideas in a roadmap across a series of posts (initially planned as just one post, but I quickly realized it would be quite a lengthy one 😜). The goal is to create a resource to refer back to when exploring this exciting emerging field, get inspiration, and organize thoughts.</p><h3>What is AI?</h3><p>To define GeoAI, it’s essential to first understand what we mean by “-AI” and what exactly the “Geo-” part of the equation represents. Let’s start with the first question — <strong>what is AI?</strong></p><p>AI, or Artificial Intelligence, is a field of research focusing on simulating human intelligence through the means of a machine. It’s a broad field that spans from philosophical and theoretical considerations to practical applications. When considering GeoAI we are mostly interested in the practical applications of AI, i.e. Machine Learning (ML).</p><p>ML is a set of statistical methods that teach the machine to learn specific tasks. These methods range from simple techniques like Linear Regression to more complex models, such as Decision Trees or SVMs. At the most advanced end, we have Deep Learning (DL), which is based on the Artificial Neural Network (ANN) architecture with multiple hidden layers. To learn more about the distinction above check out <a href="https://shyampatel1320.medium.com/artificial-intelligence-vs-machine-learning-vs-deep-learning-18bfcebc2aa2">this Medium post.</a></p><h3><strong>What is “Geo-” in GeoAI?</strong></h3><p>Now that we understand what the “-AI” stands for, <strong>let’s move on to the</strong> <strong>“Geo-”</strong>.</p><p><strong>“Geo-” </strong>stands for geospatial sciences, which cover a broad range of topics from geodetic applications through cartography and mapping to environmental analyses and weather forecasting. Essentially, it’s every type of data analysis involving spatial information, such as a coordinate system. So in GeoAI, the “Geo-” refers to the specific data type, while the “-AI” stands for the analysis method applied to said data.</p><p>At first glance, you might think: <em>‘Ah, GeoAI is simply AI applied to a specific data type’</em> or <em>‘ Well, isn’t it just normal geospatial data analysis but using AI methods?’ —</em> and to some extent, yes, you’d be right. However, it isn’t all that simple… Merging two broad fields creates new challenges, skills to master, and concepts to learn.</p><p>This field often attracts two types of experts: those with computer science or data science backgrounds, and those with geography or geospatial science backgrounds. Here a challenge arises: ML engineers often underestimate the complexity of geospatial data, while geospatial analysts underestimate the amount of knowledge and skill required to work with ML techniques and spatial data.</p><p>This creates two learning paths in GeoAI:</p><ul><li><strong>The geospatial path,</strong> where geographers learn to apply ML to their familiar spatial data.</li><li><strong>The data science path</strong>, where ML engineers and data scientists figure out how to integrate spatial data into well-known, to them, ML workflows.</li></ul><p>Both paths require a substantial amount of learning new concepts. To understand the GeoAI field expertise in both domains is needed to then deal with the complexity at their intersection.</p><p>Finally, to understand fully what GeoAI is about, it’s crucial to understand what are its real-world applications. <strong>So…</strong></p><h3><em>…what is GeoAI used for?</em></h3><p>The applications of GeoAI are limitless, depending on the type of data at hand — ranging from ‘.csv’ cadastral files with building locations to satellite imagery captured by SAR sensors. ML methods also enable a wide range of analyses to be performed on each data type — prediction, classification, segmentation, object detection, change detection… and the list of tasks that are available to a GeoAI analyst goes on. I highly recommend checking out <a href="https://github.com/satellite-image-deep-learning/techniques">the satellite image deep learning GitHub repository</a> by <a href="https://medium.com/u/5cfc523ed7a1">Robin Cole</a> for more examples of tasks performed, in this case specifically on remote sensing data.</p><p>I know it may be a lot of new terminology, and that’s perfectly okay. In future posts, I’ll break down concepts like SAR sensors or object detection analysis. For now, let me provide some examples to give context to what GeoAI is already being applied!</p><ul><li><strong>Landslide Detection and Prediction [1, 2]:</strong> Using satellite imagery of the Earth it’s possible to apply change detection or classification techniques to detect if a landslide has occurred. Thanks to automated detection, large landslide inventories can be created, which can be then used to predict areas susceptible to landslides. Such landslide susceptibility maps can be created using ML methods like Random Forest.</li><li><strong>Monitoring of Air Quality [3]: </strong>With data from ground measurement stations and satellite missions (such as <a href="https://sentinels.copernicus.eu/web/sentinel/missions/sentinel-5/overview">Sentinel-5</a>) it’s possible to monitor air quality and predict its values for upcoming days based on weather forecasts and other factors. These types of analyses help issue air quality warnings and understand trends, which can be then used to implement prevention strategies in cities.</li><li><strong>Cadastral Mapping [4]:</strong> ML is used in land registration systems to detect land borders from aerial and satellite images in areas of the world where there are no cadastral records of lands (such as several African countries). Using an ML model to detect physical borders between parcels cuts the cost and time of creating such land registries.</li><li><strong>Weather Forecasting [5]: </strong>A lot of efforts using GeoAI are concentrated on refining weather forecasting models using spatial data from ground monitoring stations and remote sensing acquisitions. ML models are first trained on historical data and then used for predicting continuous values from time series of spatial data.</li><li><strong>Detection of Illegal Dumping[6]: </strong>A big challenge in crowded world regions is waste management. GeoAI enables the detection of illegal waste dumping, which helps decision-makers detect areas in which such actions are performed and prevent further trash accumulation based on the information derived from the results of the analysis.</li><li><strong>GeoAI in Military Settings [7]: </strong>Given the last couple of years, when the war spread across many countries, including Ukraine, it’s important to understand the importance of GeoAI in military applications. With High-resolution remote sensing data, it’s possible to map the battlefield, detect tanks, find the safest routes to avoid enemy exposure, and provide rapid humanitarian assistance.</li></ul><h3>Summary</h3><p>In this post, I introduced the field of GeoAI and its main components. In the next ones, I’ll make sure to explore specific concepts from each field in detail, but for now, I’ll leave you with a <strong>GeoAI mind map</strong> to give you a broad overview of the ideas, technologies, and skills involved.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*y2Y8WpFd-6LP315NXaSQwg.png" /></figure><p>This is a high-overview, coarse visualization as to what I believe the fields’ landscape looks like — if you think something should be added or removed from the graphic, <strong>share your thoughts and ideas</strong>, and let’s learn together! See you in the next post, in which I’ll be introducing the core geospatial data types and the main distinction of machine learning methods — <strong>stay tuned 🐸!</strong></p><p><strong>[1]</strong><a href="https://www.sciencedirect.com/science/article/pii/S1674987120300542">https://www.sciencedirect.com/science/article/pii/S1674987120300542</a></p><p><strong>[2]</strong><a href="https://www.tandfonline.com/doi/full/10.1080/17538947.2024.2346263#abstract">https://www.tandfonline.com/doi/full/10.1080/17538947.2024.2346263#abstract</a></p><p><strong>[3]</strong><a href="https://www.sciencedirect.com/science/article/pii/S0034425723001608">https://www.sciencedirect.com/science/article/pii/S0034425723001608</a></p><p><strong>[4]</strong><a href="https://library.itc.utwente.nl/papers_2019/phd/crommelinck.pdf">https://library.itc.utwente.nl/papers_2019/phd/crommelinck.pdf</a></p><p><strong>[5]</strong><a href="https://sites.research.google/weatherbench/">https://sites.research.google/weatherbench/</a></p><p><strong>[6]</strong><a href="https://www.sciencedirect.com/science/article/pii/S1877050921011224">https://www.tandfonline.com/doi/full/10.1080/22797254.2019.1604082</a></p><p><strong>[7]</strong><a href="https://www.opastpublishers.com/open-access-articles/application-of-gis-and-artificial-intelligence-in-military-operations-prospects-and-challenges.pdf">https://www.researchgate.net/publication/384552788_Application_of_GIS_and_Artificial_Intelligence_in_Military_Operations_Prospects_and_Challenges</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e3fefa2575a9" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>