<?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[Aspectum - Medium]]></title>
        <description><![CDATA[Cloud based ecosystem for location data visualization and analysis. Improve your business performance by capitalizing your data intelligence - Medium]]></description>
        <link>https://medium.com/aspectum?source=rss----39e9b20cab0f---4</link>
        <image>
            <url>https://cdn-images-1.medium.com/proxy/1*TGH72Nnw24QL3iV9IOm4VA.png</url>
            <title>Aspectum - Medium</title>
            <link>https://medium.com/aspectum?source=rss----39e9b20cab0f---4</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 04:55:28 GMT</lastBuildDate>
        <atom:link href="https://medium.com/feed/aspectum" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Optimizing array formatting with filtering in JS]]></title>
            <link>https://medium.com/aspectum/optimizing-array-formatting-with-filtering-in-js-c89926b26e5c?source=rss----39e9b20cab0f---4</link>
            <guid isPermaLink="false">https://medium.com/p/c89926b26e5c</guid>
            <category><![CDATA[javascript]]></category>
            <category><![CDATA[performance]]></category>
            <category><![CDATA[filtering]]></category>
            <category><![CDATA[loopring]]></category>
            <dc:creator><![CDATA[Alex Devyatkov]]></dc:creator>
            <pubDate>Wed, 04 Dec 2019 15:12:26 GMT</pubDate>
            <atom:updated>2019-12-17T07:28:56.220Z</atom:updated>
            <content:encoded><![CDATA[<p>When you need to convert a small array of data from one format to another, you hardly ever think about performance because latest PCs with their computing capabilities perform such tasks in no time. However, when handling a vast amount of data even the slightest delay can cause an unpleasant surprise. That’s why such cases require particular deliberation and thorough research.</p><p>Let’s try to find the best way to convert array items from <em>ItemSource</em> to <em>ItemTarget.</em></p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/43ee7832f1b5d79f63585056da4ddf7f/href">https://medium.com/media/43ee7832f1b5d79f63585056da4ddf7f/href</a></iframe><p>Additional requirement: the resulting array must contain unique elements only.</p><p>Let’s compare the performance of different possible solutions. We will create a <a href="https://github.com/odevyatkov/filtersSpeedTest/blob/master/data/test.json">test data</a> array containing 531678 elements, including 8150 unique ones, and use it to determine execution time of every algorithm. In order to minimize the measure of inaccuracy, we will run each algorithm 100 times and then calculate the average execution time.</p><h3>Formatting</h3><p>The most obvious solution is to use</p><pre>items.map(…)</pre><p>However, let’s try to find the best iteration by using different loops. <a href="https://github.com/odevyatkov/filtersSpeedTest/blob/master/tests/simpleLoops.js">Test source</a>.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/b8fbe8f2422885b588bb1e784f65198f/href">https://medium.com/media/b8fbe8f2422885b588bb1e784f65198f/href</a></iframe><figure><img alt="" src="https://cdn-images-1.medium.com/max/733/1*jYz1F6JqVbcBZkchppRSsw.png" /></figure><p>Turns out, the fastest loop is <strong>for</strong>, while the “obvious” solution produces the worst results.</p><h3>Filtering by id</h3><p>Since we need to both format data and filter it by id, let’s create a new array that stores unique processed data. We will use the <strong>filter </strong>method to filter data of the <strong>map</strong> loop. <a href="https://github.com/odevyatkov/filtersSpeedTest/blob/master/tests/filterWithArray.js">Test source</a>.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/b04d043211769ed7c3e92949b411c191/href">https://medium.com/media/b04d043211769ed7c3e92949b411c191/href</a></iframe><figure><img alt="" src="https://cdn-images-1.medium.com/max/732/1*EcNDI_l8VvZgp4YxfNh17g.png" /></figure><p>As you see, the <strong>for</strong> loop produces the best results again. But we won’t stop here and will try to reach an even higher optimization level.</p><h3>Optimizing filtering by id</h3><p>Given that <strong>item.id</strong> is a unique value, let’s try using the <strong>Set </strong>object instead of an array. <a href="https://github.com/odevyatkov/filtersSpeedTest/blob/master/tests/filtersWithSet.js">Test source</a>.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/1c2cd365c1116fcad96cb35e794bde9e/href">https://medium.com/media/1c2cd365c1116fcad96cb35e794bde9e/href</a></iframe><figure><img alt="" src="https://cdn-images-1.medium.com/max/760/1*OvB4HGnys-way2d9i8Khlw.png" /></figure><p>Awesome!</p><p>Now, let’s save the formatted values to a key-value object and get the final result by running Object.values. <a href="https://github.com/odevyatkov/filtersSpeedTest/blob/master/tests/filterWithKeyValueObject.js">Test source</a>.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/4454ad99e5732b9fb2b85273ebd1e346/href">https://medium.com/media/4454ad99e5732b9fb2b85273ebd1e346/href</a></iframe><figure><img alt="" src="https://cdn-images-1.medium.com/max/767/1*o-7sO2XMCN6GcEJEWP0HDg.png" /></figure><p>The results show that the combination of <strong>for + ids as Set </strong>has the best running time.</p><h3>Summary</h3><p>The most effective way to format data is using <strong>for </strong>to iterate the array and using <strong>Set</strong> instead of an array that stores the unique processed elements. As a result, we get the combination that runs 60 times faster than <strong>map + filter</strong>! <a href="https://github.com/odevyatkov/filtersSpeedTest/blob/master/index.js">Test source</a>.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/0996c132d133e62c00880d138df4bf61/href">https://medium.com/media/0996c132d133e62c00880d138df4bf61/href</a></iframe><figure><img alt="" src="https://cdn-images-1.medium.com/max/764/1*1vVrgTROFagD50zEfJxUrw.png" /></figure><p>Filter fast and prosper!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c89926b26e5c" width="1" height="1" alt=""><hr><p><a href="https://medium.com/aspectum/optimizing-array-formatting-with-filtering-in-js-c89926b26e5c">Optimizing array formatting with filtering in JS</a> was originally published in <a href="https://medium.com/aspectum">Aspectum</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Geospatial Open Data. What’s it like?]]></title>
            <link>https://medium.com/aspectum/geospatial-open-data-whats-it-like-58445177118e?source=rss----39e9b20cab0f---4</link>
            <guid isPermaLink="false">https://medium.com/p/58445177118e</guid>
            <category><![CDATA[data-visualization]]></category>
            <category><![CDATA[data]]></category>
            <category><![CDATA[gis]]></category>
            <category><![CDATA[geospatial]]></category>
            <category><![CDATA[open-data]]></category>
            <dc:creator><![CDATA[Mykola Kozyr]]></dc:creator>
            <pubDate>Wed, 11 Sep 2019 20:20:20 GMT</pubDate>
            <atom:updated>2019-09-11T20:20:19.962Z</atom:updated>
            <content:encoded><![CDATA[<h3>Analyzing ArcGIS Hub Open Data</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DQhpa4zFeYIU1pt7d_0cuA.png" /></figure><p>I have always wondered what exactly data people usually work with. <strong>What could be a better way to describe data than by actually collecting and processing data about data?</strong></p><p>It seems to me that there is no better motivation for researching and writing this article than limitless <strong>curiosity</strong>. However, I’m also sure that information from it will be valuable and helpful to those who take part in developing data-related software or in planning storage capabilities.</p><h3>Data</h3><p>I have collected information about all ArcGIS Hub items: Layers, Maps, Documents, and Images. <strong>184406 objects altogether</strong>. The following information is available for all items:</p><ul><li>name</li><li>owner</li><li>date created</li><li>type of object</li><li>number of features</li><li>link to the source</li><li>geographical extent</li><li>geometry type</li><li>attributes names and data types</li></ul><p>In this article, I would like to present some of the most interesting insights discovered from the collected data. If you would like to dig deeper and explore it on your own, feel free to <a href="https://drive.google.com/open?id=128wZpIm_tM06AqCYtea15kFR6Fvl4HPD">check out the whole dataset here</a>.</p><p><em>For all of the visualizations below, except geographical extent, I have filtered and used information about items with </em>dateType <em>equals</em> “Layer” <em>(</em><strong><em>132151 features</em></strong><em>).</em></p><p>The dataset is collected on <strong>July 16, 2019.</strong></p><h3>Rows</h3><p>Probably the first thing to ask is how many features does a regular dataset have? Here is the distribution of a number of features per dataset.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SE6H8YAiqpv0i_MYpyXkRw.png" /><figcaption>Number of Features Distribution (10% outlier)</figcaption></figure><p>Among all the datasets, there are <strong>top 5 datasets by the number of features</strong>:</p><ol><li><a href="https://www.arcgis.com/home/item.html?id=bb69f10baf334d4c935a0fb23d758f38">Microsoft Building Footprints-Features</a> — 125 192 184</li><li><a href="https://www.arcgis.com/home/item.html?id=0f4a2790c72e4a0693328833f4a8b87c">CalTRANS LaneClosures</a> — 32 425 950</li><li><a href="https://www.arcgis.com/home/item.html?id=c08c3c44f0ae48c4a6016e6a68b508f5">WisDOT Waze Traffic Jam History</a> — 18 944 990</li><li><a href="https://www.arcgis.com/home/item.html?id=dedfdc7c280b4834ae1e2d03829b16ac">Recintos2016</a> — 16 056 367</li><li><a href="https://www.arcgis.com/home/item.html?id=92e27de06d1e4c3dbb0ea3fa1b8e6f59">Service Road</a> — 15 112 592</li></ol><h3>Attributes</h3><p>Let’s continue discovering insights by moving from information about rows to the information about columns. With no additional words — just questions I wanted to answer.</p><h4><strong>Question 1 — How many columns does a regular dataset have?</strong></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XTxcvbh0SUUIlliDebR5cQ.png" /><figcaption>Number of Attributes Distribution (1% outlier)</figcaption></figure><h4><strong>Question 2 — Which data types are the most common?</strong></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sXKIlCB4rPyo0ssKb0iYuQ.png" /><figcaption>Attributes Types Occurance</figcaption></figure><h4><strong>Question 3 — What are the most popular attribute names?</strong></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*d3FjRjI1-cZctzNMEOGBjg.png" /><figcaption>A word cloud with the most common attribute names (only letters and numbers, lowercase)</figcaption></figure><p>Seems like the Platform itself adds some column names to the datasets that users upload (<em>personally, I would never use ‘shapelength’ columnname in data)</em>. I’ve decided to avoid using them in the word cloud, so the following image shows the most popular column names, except the ones from the stoplist:</p><pre>stop = [&#39;objectid&#39;, &#39;fid, &#39;id, &#39;globalid&#39;, &#39;shape&#39;, &#39;shapelength&#39;, &#39;shapelen&#39;, &#39;shapestlength&#39;, &#39;shapearea&#39;, &#39;shapestarea&#39;]</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1e_lvoL9cAaklBVN8DxFWA.png" /></figure><h3>Geometry Types</h3><p>At this point, we are getting closer to the geospatial part.</p><p>I’ve had a hypothesis that the proportion between geometry types in data could not remain the same. The second hypothesis was that the points representation of data was much more popular than other geometry types.</p><p>As you can see below, I was totally wrong. The first chart shows normalized geometry types distribution over time.</p><p>Polygons geometries are the most popular over time. However, points and lines are quite popular as well.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*c1zCDqS0f8SRUIRA0AxtrQ.png" /><figcaption>Geometry Types Proportions in Datasets over Time</figcaption></figure><p>There are some strange patterns at the beginning of the chart due to a small number of datasets, so the next chart shows the <strong>distribution based on the absolute number of datasets uploaded at each time interval</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-47AMC0y0ZSqRV1N16tKRw.png" /><figcaption>Number of Datasets Uploaded by Geometry Types</figcaption></figure><h3>Extent</h3><p>Open data coverage is definitely one of the most interesting things to discover. In this article, we are presenting only the visual representation of it by adding the layers’ extents to the map. Here you can see screenshots from <a href="https://aspectum.com/">Aspectum</a> showing different regions based on the global visualization of open data layers extents.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2rNWNQZRL1BKC4oIClMakQ.png" /><figcaption>You may notice a big concentration of datasets in Western Europe. It is interesting that specific regions in Spain and France are well recognized here. Unfortunately, Ukraine, Belarus, Moldova, and Balkan countries do not contribute to Open Data Hub a lot</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*s_YtBjzWsIneLRQLpVYuwQ.png" /><figcaption>North America is definitely the champion for the number of open data on the portal. Just a quick reminder — the styling parameters are the same for all regions</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jja1h4NqGwh1JAVmQyfw_w.png" /><figcaption>Specific regions in Taiwan and Japan are well-recognized</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FAmHdjnbldQrb2jbddQc3Q.png" /><figcaption>New Zealand looks great here. Even though it is close to the extent of global datasets (180,-180 longitude line), it remains clear that the bunch of layers belong to New Zealand</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eoPkykrvJD7TROlB4l-HMw.png" /><figcaption>Kenya turned out to be one of the global leaders in Open Data Hub, and it is well-recognized here</figcaption></figure><h3>Contributors</h3><p>The last, but not the least important question to ask is who has contributed to ArcGIS Open Data Hub the most? This is a list of top 5 contributors per each category.</p><h4>Layers:</h4><ol><li>SouthVancouverIslandProsperityProject — 6496</li><li>plfran_umich — 1330</li><li>ONSGeography_data — 1182</li><li>idecvingt<strong> — </strong>1076</li><li>BIOS_Admin<strong> — </strong>1046</li></ol><h4>Maps:</h4><ol><li>nicolas.lopez.cundina — 685</li><li>uneplive_team — 278</li><li>esrichinahk — 253</li><li>Esri_BeLux_Content — 252</li><li>NIED_CRS@bosai — 199</li></ol><h4>Images:</h4><ol><li>lucy.dixon<strong> — </strong>270</li><li>sjanzen <em>and</em> mdimapdatacatalog — 210 <em>each</em></li><li>PKabatha — 110</li><li>SouthVancouverIslandProsperityProject — 106</li></ol><h4>Documents</h4><ol><li>PrefeituraRio — 3131</li><li>openMONCTONouvert — 2528</li><li>esrica_apps<strong> — </strong>1213</li><li>LakeCounty_Illinois — 1052</li><li>cohendri_sarco — 998</li></ol><h3>Summary</h3><p>Just as I have expected at the beginning of my research for this article, I found tons of valuable information. It helps me understand users of GIS SaaS better. It is the next step to define the target audience for the scale of the data used, as well as to set up appropriate default options for showing labels and styling parameters. So, is it about data-driven software development?</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=58445177118e" width="1" height="1" alt=""><hr><p><a href="https://medium.com/aspectum/geospatial-open-data-whats-it-like-58445177118e">Geospatial Open Data. What’s it like?</a> was originally published in <a href="https://medium.com/aspectum">Aspectum</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Acceleration for the Nearest Neighbor Search on Earth’s Surface Using Python]]></title>
            <link>https://medium.com/aspectum/acceleration-for-the-nearest-neighbor-search-on-earths-surface-using-python-513fc75984aa?source=rss----39e9b20cab0f---4</link>
            <guid isPermaLink="false">https://medium.com/p/513fc75984aa</guid>
            <category><![CDATA[python]]></category>
            <category><![CDATA[gis]]></category>
            <category><![CDATA[gpu]]></category>
            <category><![CDATA[geospatial]]></category>
            <dc:creator><![CDATA[Serhii Hulko]]></dc:creator>
            <pubDate>Tue, 20 Aug 2019 08:51:19 GMT</pubDate>
            <atom:updated>2019-08-20T09:28:13.217Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_4EQD7zWeP8wuCbkoIualQ.png" /><figcaption>Points connected to their nearest neighbors</figcaption></figure><h4>Application</h4><p>I’m currently working on a project which performs geospatial analysis called <a href="https://aspectum.com/">Aspectum</a>. Our team decided to increase the speed of algorithms for the analysis. Many of them rely on Nearest neighbor search (NNS). As speed increases in its computation, it affects the whole analysis time and allows to process more massive volumes of data. So, this article describes steps of improving that I’ve implemented in the project.</p><h4>Vincenty</h4><p>If you need to work with a specific area, say a city or a country, it’s easier to put the data in local projection and perform distance computation on a plane. However, if you want a universal tool for any area or scale (in my case, a web application), <a href="https://en.wikipedia.org/wiki/Vincenty%27s_formulae">Vincenty’s formula</a> is your choice.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/282/1*HtM-Cf-4Qd_xFLZzv_-T3g.jpeg" /><figcaption>Thaddeus Vincenty (1920–2002)</figcaption></figure><p>The formula uses the <a href="https://en.wikipedia.org/wiki/World_Geodetic_System#WGS84">WGS84 ellipsoid</a>, and the accuracy of calculations is limited only by the differences between the ellipsoid and the Earth’s surface. Despite being universal, the formula is far from trivial, and it contains the coverage cycle. It requires about 130 iterations to achieve a result accurate up to 1 mm. So, it takes more time to find a distance on an ellipsoid compared to a sphere or a plane.</p><h4>Specs</h4><p>All benchmarks below were done on a computer with CPU Intel Core i5–7500, 3.4GHz (4 cores) and GPU NVIDIA GeForce GTX 1060 (6GB). I used Python v3.6.8. The took a <a href="https://github.com/felytic/vincenty_cuda_nns_tests/blob/master/datasets/crimes_points.csv">dataset</a> with locations of crimes in the US. Computation with different data and hardware shows different timing, but the trends are similar.</p><h4>Pure python</h4><p>Let’s start with the easiest approach. Just use the formula from the <a href="https://pypi.org/project/vincenty/">Vincenty</a> package and calculate the distance between two points to find the nearest neighbor for each of them.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/40c2cbba11bb171eb867128ffdf2718b/href">https://medium.com/media/40c2cbba11bb171eb867128ffdf2718b/href</a></iframe><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MJkKqZXLPSgie2TQiE_G-Q.png" /></figure><h4>Sklearn</h4><p>A more sophisticated method is to use a <a href="https://en.wikipedia.org/wiki/Ball_tree">Ball Tree structure</a> for spatial indexing from the <a href="https://scikit-learn.org/stable/modules/generated/sklearn.neighbors.BallTree.html">Sklearn</a> package and <em>vincenty</em> as the metric.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/e31108714d131d10ddd07116002b38cc/href">https://medium.com/media/e31108714d131d10ddd07116002b38cc/href</a></iframe><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6TSe3HSR_p_N8X5MUTG6uw.png" /></figure><h4>Compilation with numba</h4><p>Another way to make your code faster is to compile. If some part of code runs many times (like <em>vincenty</em> in our case), it can be compiled once. This increases the speed. The <a href="https://numba.pydata.org/">Numba</a> package can help us with it. I’ve made some minor modifications in Vincenty package to fit Numba’s (also <a href="http://numba.pydata.org/numba-doc/latest/cuda/overview.html">CUDA</a>) requirements and put it into a <a href="https://pypi.org/project/cuda-friendly-vincenty/">cuda-friendly-vincenty</a> package.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/4d9d0f6ca1d10bcf9748cebd150472fc/href">https://medium.com/media/4d9d0f6ca1d10bcf9748cebd150472fc/href</a></iframe><p>Have a look at how a simple approach with compilation overcomes sophisticated ball tree structure:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*D1E4tWAu_ct9gFcqMfTYVw.png" /></figure><h4>Sklearn + numba</h4><p>The next obvious step is a combination of Numba compilation and Sklearn indexing.</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/063748173749354489e2a5593126c08b/href">https://medium.com/media/063748173749354489e2a5593126c08b/href</a></iframe><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ohDx5AU-cblckVw3D2lKjw.png" /></figure><h4>GPU</h4><p>Numba package also provides an interface for <a href="http://numba.pydata.org/numba-doc/latest/cuda/overview.html">CUDA</a> which allows you to run code on a GPU. <a href="https://blogs.nvidia.com/blog/2009/12/16/whats-the-difference-between-a-cpu-and-a-gpu/">GPUs</a> have many <a href="http://numba.pydata.org/numba-doc/latest/cuda/cudapysupported.html">limitations</a> in comparison with CPUs, but they are perfect for tons of simple calculations that can run simultaneously. Hopefully, the Vincenty formula is pure math, so that it can be easily run on graphic cards. Again, same simplest approach, but with CUDA:</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/09a7cf9803e3cff6acdb1126636e144c/href">https://medium.com/media/09a7cf9803e3cff6acdb1126636e144c/href</a></iframe><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VyWbBC-eyG6NmhL5kgLKJg.png" /></figure><h4>Ball Tree</h4><p>Let’s dig a little deeper and explain the ball structure a bit more.</p><p>The idea behind it is to put data points in a <a href="https://en.wikipedia.org/wiki/Binary_tree">binary tree</a> similar to a <a href="https://en.wikipedia.org/wiki/Binary_search_tree">search tree</a>. Every node has two parameters: <a href="https://en.wikipedia.org/wiki/Centroid">centroid</a> and radius. The centroid is a point on a plane and radius is the distance to the farthest point of a node. Let’s have a look at an example.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yGt7sXfyY6dzv4qLGUrrOw.gif" /><figcaption>First three levels of tree creation example</figcaption></figure><p>At level zero, we have all points and one circle to fit them.</p><p>Next, to create first-level nodes, a building algorithm finds a number of points that are spread in latitude or longitude. Then it splits them by coordinates (in our case, it’s longitude).</p><p>Finally, it creates nodes for both subsets.</p><p>Then everything repeats until we get nodes with the number of points less than a particular parameter. It is usually called leaf size. Multiple tests showed than optimal leaf size is about 5.</p><p>After building a tree, we can start searching for each point`s neighbors. At first, the algorithm finds the closest neighbor inside point’s home node using the simple approach described before. Next, it walks around the tree. It’s easy to calculate the distance from any point to a node. it’s the distance to its centroid minus radius. If this distance is smaller than the minimal one, the algorithm goes deeper inside the node and looks for the neighbors within; otherwise, it skips the node and all its children.</p><p>This approach is much faster because it reduces the number of points that must be checked.</p><h4>Ball Tree on GPU</h4><p>My task was to put the ball tree structure inside a GPU. The challenge was to implement this structure using only arrays of fixed size and to be able to walk over this pseudo-tree without any recursion or additional memory. You can find the result in the <a href="https://pypi.org/project/vincenty-cuda-nns/"><strong>vincenty-cuda-nns</strong></a> package. Here is an example of its usage:</p><iframe src="" width="0" height="0" frameborder="0" scrolling="no"><a href="https://medium.com/media/3c173d3c74ecaed6f25930c40a83a173/href">https://medium.com/media/3c173d3c74ecaed6f25930c40a83a173/href</a></iframe><p>And this is a comparison with other approaches:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*liW_v086Hug2ZF6cOZLarw.png" /></figure><h4>P. S.</h4><p>These techniques have domain-specific flaws because I’m not a GPU or a geospatial guru. So, if you have an expertise in this field, I would be grateful for some constructive feedback.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=513fc75984aa" width="1" height="1" alt=""><hr><p><a href="https://medium.com/aspectum/acceleration-for-the-nearest-neighbor-search-on-earths-surface-using-python-513fc75984aa">Acceleration for the Nearest Neighbor Search on Earth’s Surface Using Python</a> was originally published in <a href="https://medium.com/aspectum">Aspectum</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Spatial Pattern Analysis]]></title>
            <link>https://medium.com/aspectum/spatial-pattern-analysis-e27b745bc891?source=rss----39e9b20cab0f---4</link>
            <guid isPermaLink="false">https://medium.com/p/e27b745bc891</guid>
            <category><![CDATA[gis]]></category>
            <category><![CDATA[data-visualization]]></category>
            <category><![CDATA[research]]></category>
            <category><![CDATA[open-data]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Mykola Kozyr]]></dc:creator>
            <pubDate>Tue, 21 May 2019 10:59:30 GMT</pubDate>
            <atom:updated>2019-08-20T10:23:08.439Z</atom:updated>
            <content:encoded><![CDATA[<h3>Data enrichment matters. Business and Education collaboration</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UCLRvIW-9ZbnbPGByK38yg.png" /></figure><p>What do we expect to see while visualizing location data by adding it to the map for the first time? I’m pretty much sure our goal is to discover whether there are any spatial patterns: regions with higher and lower intensity of features, clusters and outliers, common directions, irregular trajectories, territories with higher or lower values, etc.<em> </em>As soon as we explore patterns we are ready to determine features significance, build predictive models and <em>call ourselves data scientists</em>.</p><p>The approaches used to discover patterns are worth a separate article. Currently, I’m working on it and going to start it with the title <em>“Stop using heatmaps…”</em>. So let’s imagine the researcher is aware of pattern analysis algorithms, and ready to go deep into the world of GIS.</p><p>In this article, I will mention point pattern analysis, show the results of pattern analysis algorithms and issues the non-GIS professional would face, and describe the collaboration between business and science in developing the solution for the current software incompetence.</p><h3>Point Pattern Analysis. Intensity Clusters &amp; Outliers</h3><p>Let’s start with a simple, but a real-world case: researcher has got locations of thefts <em>(car accidents, emergency calls, tweets, photos, birds’ nests or beavers’ lodges. Absolutely any type of location data)</em> in a city <em>(park, sea, country… no matter what)</em>. Researcher’s task is to explore if there are any spatial patterns and distinguish them if they are. He or she has built the grid over the area of distribution using the specific cell size to avoid randomness. At the next step, the researcher has counted the number of features inside every cell and finally run the Clusters &amp; Outliers Analysis and/or Hotspots &amp; Coldspots Analysis, it depends on the actual task.</p><p>In addition, he or she has run NNA and Quadrat Analysis, so the researcher is pretty much sure now the objects have clustered distribution and results show these clusters, as well as outliers.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yvYpuj4_v28qEGbltxyiJQ.png" /><figcaption><a href="https://vision.eos.com/maps/shared/0c5c4f07-a787-4cc1-bc7b-5ddce79ac66c">Thefts Pattern Analysis. Screenshot from Aspectum</a></figcaption></figure><p>I would say the researcher is pretty much satisfied with the result. It is good enough to start exploring outliers, however, the area covered by clusters of high and low values is insanely huge. For people who know the city, high value cluster looks obvious — these are residential and business areas. The cluster with low values (blue part) corresponds with forests, river, and industrial areas.</p><p>At this point <strong>to get more precise results</strong> for clustered areas the researcher decides to take into account urban classes while analyzing the data. It may help to find clusters within similar territories: comparing features distribution within residential areas separately from green or water zones.</p><p>So, imagine the researcher did not use the Aspectum and faced all the GIS-related issues: shapefiles, projections, clumsy desktop software, areas of distribution vector data, grid size formula, appropriate algorithms via plugins or separate extensions <em>(btw, in case of using Aspectum it’s just about the raw data).</em></p><p>Now he or she has to build the urban land classification vector dataset. But how? Where to find data? What zones should be distinguished? What are the parameters to classify this zone? How to automate this for further tasks?</p><h3>The Project. When business and education enrich</h3><p>We believe it is the task of the software development companies to deliver a product able to produce a valuable result with no need to teach the target audience to work. <em>And in case of GIS, it usually takes 6 years of high school studying.</em></p><p>Aspectum decided to initiate a project to generate urban landuse data for all the settlements in the world. Such an ambitious goal became interesting to our friends and partners from <a href="https://www.facebook.com/NoosphereEngineering/"><strong>Rivne Noosphere Engineering School</strong></a>. The team of students with software engineering, GIS, urban studies background, and first of all, passion was organized under the supervision of lecturers.</p><p>Since there is no correct approach for the task, we started to develop by iterations, constantly reviewing and discussing the results. This article will introduce the core idea of the implementation of the project, and show the results now available for testing.</p><p>In the following articles, we will share approaches have been tested, results which did not work, ideas for improvements and supplemental features implemented.</p><h4>The Concept</h4><p>There are huge open source projects dealing with geospatial data. Our goal is to merge the data and the ways to process it to get the most accurate and complete information about the certain area in settlements all around the world. Just to clarify — we are dealing not with landuse types, it is about <em>the idea of similarity of the </em><strong><em>influence</em></strong><em> on the phenomena and processes taking place in these areas. </em>So as well as residential areas could be absolutely different across the city, we have to assign them different classes. On the other hand, we are not dealing with the zoning approach — there is no need to combine retail and business areas in one group, as well as it is not obvious if it makes much sense to separate education and medical areas in terms of influencing on spatial patterns in these areas.</p><h4>Current results</h4><p>We are now testing the MVP version of the product, delivering information about the 4 classes inside an area of data distribution. Go to Aspectum, add your data and run Urban Land Classification analysis. It may take a while to process and we’ve still got a number of limitations, but we are constantly improving the algorithm.</p><p><a href="https://vision.eos.com/maps/shared/dec00246-53e3-46fb-a6b2-94a3f775f333">Check the map with the classes at local and regional level</a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*X_ssaOuN5FM04gy6Z3MF-g.png" /><figcaption>Classes in local and regional levels. Screenshot from Aspectum</figcaption></figure><p>There have been a number of developments made during the task, and we are happy to share them. Take a look at our <a href="https://github.com/eos-vision">GitHub page</a>, we are constantly updating and adding projects there. <a href="https://github.com/eos-vision/osm2geojson">osm2geojson</a>, our latest update, is a pure Python solution to process OpenStreetMap data. Follow Aspectum not to miss the next articles.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e27b745bc891" width="1" height="1" alt=""><hr><p><a href="https://medium.com/aspectum/spatial-pattern-analysis-e27b745bc891">Spatial Pattern Analysis</a> was originally published in <a href="https://medium.com/aspectum">Aspectum</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Discovering potential settlements’ areas in Argentina]]></title>
            <link>https://medium.com/aspectum/discovering-potential-settlements-areas-in-argentina-eb73d9bd3bc1?source=rss----39e9b20cab0f---4</link>
            <guid isPermaLink="false">https://medium.com/p/eb73d9bd3bc1</guid>
            <category><![CDATA[gis]]></category>
            <category><![CDATA[openstreetmap]]></category>
            <category><![CDATA[remote-sensing]]></category>
            <category><![CDATA[open-data]]></category>
            <category><![CDATA[mapping]]></category>
            <dc:creator><![CDATA[Malyuk Maksim]]></dc:creator>
            <pubDate>Wed, 08 May 2019 11:21:00 GMT</pubDate>
            <atom:updated>2019-05-08T14:04:50.459Z</atom:updated>
            <content:encoded><![CDATA[<p>Let’s imagine there is a dataset with all the settlements in the world. You’ve got boundaries, name, population, administrative subordination. The dataset is being constantly updated, saving the information about the previous data. How beneficial it could be?</p><ul><li><em>Quick and targeted emergency response</em></li><li><em>Exploring and modeling of worldwide urban growth</em></li><li><em>Analyzing human footprint on climate change</em></li><li><em>Creating regional planning strategies in developing countries</em></li><li><em>Huge potential for data fusion in the fields of statistics, IoT, location intelligence, autonomous driving, AI and ML.</em></li></ul><p>Obviously, the best place to organize such data would be the OpenStreetMap. Using OSM history and changesets databases researchers and developers would have access to history and the modernity of humanity’s progress.</p><p><a href="https://eos.com/eos-vision-retail/">Vision Team</a> has challenged itself to lead the way in developing the Global Settlements Dataset.</p><p>This article describes the process of predicting potential settlements in Argentina. We have analyzed OpenStreetMap data, VIIRS night imagery, MaxMind World Cities, and developed a ready-to-use data to start a targeted campaign for adding settlements boundaries to OSM.</p><p>For information about the current state of settlements in Argentina see our <a href="https://twitter.com/EOS_Vision/status/1067133310630010881">post here</a>.</p><ul><li>Image pre-processing</li><li>Intersection</li><li>Pitfalls</li><li>Outcome</li><li>Use case</li></ul><h3><strong>IMAGE PRE-PROCESSING</strong></h3><p>We used data from the <strong>VIIRS </strong>spacecraft as the basis for the search project on potential settlements, missing on OpenStreetMap, as its night mode enables to capture in high accuracy objects emitting light. The spatial resolution of its sensor is 750 m.</p><p>The satellite image pre-processing procedure comprises 2 stages:</p><ul><li><em>Creating a binary mask on the threshold;</em></li><li><em>Vectorization of the outcome result.</em></li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*K_K9Dadt5XCyPjmJ" /><figcaption>The image pre-processing was made with the gdal (gdal_calc and gdal_polygonize) tools.</figcaption></figure><p><strong>INTERSECTION</strong></p><p>Settlements in Argentina are mapped with a tag <strong>relation[“admin_level”=”8”]<em>.</em></strong></p><p>For the further workflow, the data is uploaded via <a href="http://overpass-turbo.eu/"><em>overpass-turbo.eu</em></a><em>.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*5PN5q3wSZqJs4v8D" /><figcaption><em>The binary mask should eliminate all area settlements already mapped on OSM.</em></figcaption></figure><p>Next step is intersection exclusion:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*E9p0wbmcdlMsOm45" /></figure><p><strong>PITFALLS</strong></p><p>The final mask showed two large clusters of light source, which turned not to include the settlements. During analysis, we discovered that the first source of the light was an oil derricks concentration cluster and the second — a power plant.</p><p>Similar to settlements, we have eliminated intersections with the objects, having the following tags:</p><ul><li>“industrial” = ”oil”</li><li>“power” = ”generator”</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ohWX26RwV80rgAor" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*WHOnY3io_-FxJg6r" /></figure><p>While processing the data, we found out that many settlements are mapped partially, meaning areas are mapped without the boundaries of the cities. We have highlighted such cases in the project.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4eL5P0JNhV5NaDc4PhvwBQ.jpeg" /><figcaption>1.Contains node[“admin_level”=”8”]; 2.Contains[“land use”=”residential”]; 3. No valuable intersection;</figcaption></figure><p><strong>Blue color</strong> means that the settlement, which we got on the basis of the satellite image has a point with the name of the city on OpenStreetMap. This means, that <em>most likely</em> this point is a real populated area.</p><p><strong>Red color</strong> means a potential settlement which does not intersect with any object from OpenStreetMap.</p><p><strong>Red outline</strong> means that the object has got intersections with residential areas [“land use” = ”residential”], but at the same time, has no mapped borders.</p><p>We included the settlements’ points from MaxMind, and excluded all the objects within a radius of 450 m from the points on <a href="https://www.openstreetmap.org/#map"><em>OpenStreetMap</em></a>.</p><p><strong>USE CASE</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*YlRhBf_zpqaRDo77" /></figure><p>Here is an example of how to apply the project data in searching settlements and mapping on <a href="https://www.openstreetmap.org/">OpenStreetMap</a>.</p><p>So, we can see a wasteland on the map to the right</p><p>At the same time, in this area we found a potential area settlement and the point from MaxMind</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*1jPF0-XRTkEOyVuq" /></figure><p>After adding the WMS layer from Sentinel-2 satellite onto the given territory, one can see that the following area really has the settlement.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GAVd-VfsZLBhIhMcLNbc8g.gif" /></figure><p>Our team is currently analyzing existing data on OpenStreetMap and other global open data projects. I’m preparing a dataset with potential settlements in the South America and moving towards the global coverage.</p><p><strong>Check the </strong><a href="https://vision.eos.com/maps/shared/79e6cabf-0679-4662-948c-db17d6d85c6e"><strong>result</strong></a></p><p>We are encouraging the OSM community to join us in this initiative.</p><p>#GlobalSettlementsDataset is a huge step in the open data concept.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=eb73d9bd3bc1" width="1" height="1" alt=""><hr><p><a href="https://medium.com/aspectum/discovering-potential-settlements-areas-in-argentina-eb73d9bd3bc1">Discovering potential settlements’ areas in Argentina</a> was originally published in <a href="https://medium.com/aspectum">Aspectum</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How good is Open Data?]]></title>
            <link>https://medium.com/aspectum/how-good-is-open-data-5b51d4cb3e83?source=rss----39e9b20cab0f---4</link>
            <guid isPermaLink="false">https://medium.com/p/5b51d4cb3e83</guid>
            <category><![CDATA[openstreetmap]]></category>
            <category><![CDATA[gis]]></category>
            <category><![CDATA[mapping]]></category>
            <category><![CDATA[open-data]]></category>
            <category><![CDATA[geocoding]]></category>
            <dc:creator><![CDATA[Mykola Kozyr]]></dc:creator>
            <pubDate>Wed, 27 Mar 2019 14:40:12 GMT</pubDate>
            <atom:updated>2019-08-20T10:21:18.549Z</atom:updated>
            <content:encoded><![CDATA[<h3>Episode 1. OpenStreetMap Nominatim Geocoder</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tHpL2xNHoD4qj6SGB8BSFg.png" /></figure><p>Developing <a href="https://aspectum.com/">Aspectum</a> as a tool for not only GIS-professionals it was a crucial task to adopt services transforming geospatial data in the text format to geometry. The idea supporting this is to keep users’ psyche safe by not letting them work with the Shapefile only. Supporting CSV, XLS, XLSX formats with geospatial data as latitude an longitude were not enough since it is still a common way to store geospatial data as addresses. So, we were ready to go deep into Geocoding.</p><p>There are plenty of geocoding services. As a first iteration, we have decided to try Nominatim — the geocoding tool based on OpenStreetMap data. While developing the technical description of the task, the first question appeared was the following:</p><blockquote>Could Nominatim be used for <strong>all the countries</strong>?</blockquote><p><em>I could not find such information in the documentation, as well as on </em><a href="https://gis.stackexchange.com/q/288366/75547"><em>GIS StackExchange</em></a><em> (spoilers alert). </em>In addition to the coverage question, we have wanted to evaluate the quality of geocoding by accuracy, precision, preferable input data formats etc.</p><p>In this article, I would mention Nominatim data coverage, Reverse vs Direct Geocoding, and additional features appeared during this small research.</p><p><em>No OpenStreetMap was harmed in the making of this research. Everything run on the OSM replication.</em></p><h3>Coverage</h3><p>It seems to be a simple task — check if the geocoder works for all the countries: find some sample of addresses for all the countries and run the geocoder. Okay, it seems easy. Let’s just find the data. Believe me or not — it is not what you can find by just googling it. I was sure GIS-geeks met similar tasks, so I went to Twitter, asked the community, and got the brilliant answer:</p><h3>Dr. Michele M Tobias on Twitter</h3><p>@MykolaKozyr @mdsumner You could generate random lat/long coordinates and then reverse geocode them. #gischat You could probably do that in #rstats or #python</p><p>That’s easy! Next steps were simple: Natural Earth (<a href="https://www.naturalearthdata.com/corrections/ticket.php?track=4DXWZN66S4"><em>I’m still mad at you, guys</em></a>), QGIS (random 100 points per country), geopy. Now I was ready to answer the original question. Running direct geocoding from addresses dataset showed <strong>Nominatim is available for all the countries</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*puyLjHTzSvUEa6nYm3dLVg.png" /><figcaption><a href="https://vision.eos.com/maps/shared/e78d7186-6453-4d1c-be33-cf3d6d8493bb">Direct Geocoding &amp; Countries Data. Screenshot from Aspectum workspace</a></figcaption></figure><h3>Precision</h3><p>During the previous task I have spotted an interesting thing — reverse geocoding returns not only the <strong>address</strong> but also <strong>latitude and longitude:</strong></p><pre>{  <br>   &quot;type&quot;:&quot;FeatureCollection&quot;,<br>   &quot;licence&quot;:&quot;Data © OpenStreetMap contributors, ODbL 1.0. <a href="https://osm.org/copyright">https://osm.org/copyright</a>&quot;,<br>   &quot;features&quot;:[  <br>      {  <br>         &quot;type&quot;:&quot;Feature&quot;,<br>         &quot;properties&quot;:{  <br>            &quot;place_id&quot;:&quot;111094940&quot;,<br>            &quot;osm_type&quot;:&quot;way&quot;,<br>            &quot;osm_id&quot;:&quot;154201801&quot;,<br>            &quot;place_rank&quot;:&quot;26&quot;,<br>            &quot;category&quot;:&quot;highway&quot;,<br>            &quot;type&quot;:&quot;unclassified&quot;,<br>            &quot;importance&quot;:&quot;0.1&quot;,<br>            &quot;addresstype&quot;:&quot;road&quot;,<br>            &quot;name&quot;:null,<br>            &quot;display_name&quot;:&quot;Palianychyntsi, Fastiv Raion, Kyiv Oblast, 08542, Ukraine&quot;,<br>     <strong>       &quot;address&quot;:{  <br>               &quot;village&quot;:&quot;Palianychyntsi&quot;,<br>               &quot;county&quot;:&quot;Fastiv Raion&quot;,<br>               &quot;state&quot;:&quot;Kyiv Oblast&quot;,<br>               &quot;postcode&quot;:&quot;08542&quot;,<br>               &quot;country&quot;:&quot;Ukraine&quot;,<br>               &quot;country_code&quot;:&quot;ua&quot;<br>            }</strong><br>         },<br>         &quot;bbox&quot;:[  <br>            29.9982289,<br>            49.99599,<br>            30.004028,<br>            50.0029167<br>         ],<br>         &quot;geometry&quot;:{  <br>            &quot;type&quot;:&quot;Point&quot;,<br><strong>            &quot;coordinates&quot;:[  <br>               30.000565205347,<br>               50.0005604121134<br>            ]</strong><br>         }<br>      }<br>   ]<br>}</pre><p>I made an assumption this is the coordinates of the address returned, but I have noticed a thing — the point with address “AAT, Antarctica” has been returned from Reverse Geocoding, but could not be recognized by direct geocoding.</p><p>To check how different results are, I have connected coordinates from reverse geocoding with coordinates from direct geocoding. Results were absolutely various:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YR2ZDP8Q8rT4mfouKqWsXg.png" /><figcaption><a href="https://vision.eos.com/maps/shared/ecfaa9cb-e622-4e95-89de-6c4f8472baa6">Reverse vs Direct Geocoding. <em>Screenshot from Aspectum workspace.</em></a></figcaption></figure><p>Okay, it becomes more and more intriguing. Do reverse or direct geocoding return inappropriate results? Hold on, what if we go further, and <em>run reverse geocoding based on coordinates returned from direct geocoding based on addresses returned from reverse geocoding based on random coordinates!</em> It looks fun, isn’t it?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MFVNejvPY2AJe1E0WHm6NQ.png" /><figcaption><a href="https://vision.eos.com/maps/shared/579e5bf3-6f80-4a79-9bb0-0890be87ea43">Random-Reverse-Direct-Reverse Geocoding Results</a></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/904/1*CUyp4g43SgyNUc9SuCZe9A.png" /></figure><p>You may explore the map and find out that <strong>Direct Geocoding results are usually more precise</strong> than the ones taken from Reverse Geocoding. Does it look like a scientific breakthrough? Not sure, but it looks nice, isn’t it?</p><p>For me personally, the most interesting things on the map became “spider webs” usually showing the nearest address from the random coordinates. <em>That was a little spoiler for the next episode of “How good is Open Data”?</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=5b51d4cb3e83" width="1" height="1" alt=""><hr><p><a href="https://medium.com/aspectum/how-good-is-open-data-5b51d4cb3e83">How good is Open Data?</a> was originally published in <a href="https://medium.com/aspectum">Aspectum</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>