<?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 andyaketch on Medium]]></title>
        <description><![CDATA[Stories by andyaketch on Medium]]></description>
        <link>https://medium.com/@andyaketch?source=rss-cd5c922bb751------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*-wiiv1VeTxlU1FiO</url>
            <title>Stories by andyaketch on Medium</title>
            <link>https://medium.com/@andyaketch?source=rss-cd5c922bb751------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 12:40:52 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@andyaketch/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[Building Kwa Ground]]></title>
            <link>https://medium.com/@andyaketch/building-kwa-ground-a5086ec0b5a8?source=rss-cd5c922bb751------2</link>
            <guid isPermaLink="false">https://medium.com/p/a5086ec0b5a8</guid>
            <category><![CDATA[api]]></category>
            <category><![CDATA[geoapify]]></category>
            <category><![CDATA[maps]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[overpass-api]]></category>
            <dc:creator><![CDATA[andyaketch]]></dc:creator>
            <pubDate>Mon, 02 Nov 2026 08:58:52 GMT</pubDate>
            <atom:updated>2026-11-02T08:58:52.277Z</atom:updated>
            <content:encoded><![CDATA[<p><em>A map is only as true as its last check</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/857/1*xpikj_MLUwpce26UCoER-g.png" /><figcaption>Aerial View of Nairobi City (Courtesy of Google Maps)</figcaption></figure><p>I was looking for a jersey shop in Nairobi’s bustling Central Business District. Google Maps said it was right there, with a pin, a name, and a confident blue line leading me to it. I walked to the spot and found nothing. After asking around, I found the shop about 800m away.</p><p>It was a small problem, but it stayed with me. I work in GIS, so I know maps are never perfect. What bothered me was that I, someone who could have fixed that pin, had no easy way to do it. And if I couldn’t, an ordinary person had no chance.</p><p>That’s where Kwa Ground started.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TyezZaKve9KFthg4UhbhZA.png" /><figcaption>Kwa Ground Interface</figcaption></figure><h3>The idea</h3><p><em>Kwa</em> means “at” in Swahili, so Kwa Ground is literally “on the ground&quot;. The idea is simple: the people standing at a place know best where it is. Kwa Ground is a web app where you can view a building or business on the map, see where the data places it, and propose a correction by moving it to its real location.</p><p>No shapefiles, no desktop GIS software, no coordinate reference systems to wrestle with. Just a map and a pin.</p><h3>The stack</h3><p>I kept it deliberately small:</p><ul><li><strong>HTML/CSS</strong> for frontend development</li><li><strong>Javascript</strong> for both frontend and backend</li><li><strong>Leaflet</strong> for the interactive map</li><li><strong>JSON Server</strong> to store and serve corrections during development</li><li><strong>Geoapify</strong> as the place data source</li></ul><p>Small tools meant I could focus on the actual problem instead of fighting the setup. I also gave the interface a restrained three-colour palette (deep navy, warm off-white and teal) so the map stays the star and nothing else competes for attention.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eHyQT7jiM-Bfa6zj_K2v3w.png" /><figcaption>Kwa Ground functionality.</figcaption></figure><h3>The hard part: finding the right data</h3><p>This was where I spent most of my time.</p><p>For an app about correcting locations, you first need a reliable source of locations to correct. I started with Overpass, the query API for OpenStreetMap data. It’s powerful, and for a GIS person the appeal is obvious: open data, flexible queries, and full control.</p><p>But it came with real friction. Mostly it was reliability. Overpass runs on a shared public server, and mine kept turning me away. One page said the server was at capacity and every query worker was busy. Another error said my input contained only whitespace, which looked like the query was reaching the server empty. On other tabs I got a 403 and a 404. I never confirmed the cause of each one, and I couldn’t tell whether I was sending something wrong or the server was simply overloaded. When I looked into it, overload on the public instances turned out to be a known, recurring problem, and the official manual says apps that depend on them are better off running their own. That wasn’t realistic for this project. Users pick any area at runtime, so every load would be a live request to a server I couldn’t count on.</p><p>I also went looking for an API built for this exact case, one that returns places in a form that’s easy to display and correct. There wasn’t a perfect one. Geoapify’s Places API came closest, and in the end I settled on it and committed to using it as the only source of locations. You give it a bounding box and a category, and it returns GeoJSON with a name, an address, coordinates and a place ID for each result. There is no query language to learn, and in my browser test it answered in about 30 seconds. It isn’t perfect either. Some results have no name, and some addresses are messy, but that is the kind of error this app is meant to let people fix.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TmABV4s1N1DXQjaUrl_F4A.jpeg" /><figcaption>Photo by <a href="https://unsplash.com/@bluestonex_apphaus?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Bluestonex</a> on <a href="https://unsplash.com/photos/sticky-notes-with-words-and-drawings-on-wooden-table-gLxNxONfRz0?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Unsplash</a></figcaption></figure><p>The lesson I took from this is that the data source shapes the whole app. I assumed the hard work would be the map interface. It was choosing and understanding the data underneath.</p><h3>What’s next?</h3><p>Right now Kwa Ground proves the idea. To become useful, it needs four things.</p><p>The first is <strong>validation</strong>, so corrections can be checked rather than blindly trusted. The form already checks that input is well formed: a sensible name, coordinates in range, and a status from the allowed list. It can’t tell whether a correction is true. Someone can drag a pin to the wrong spot, and the app will accept it. A better version would treat each correction as a claim that needs support, such as a second person confirming it or a photo taken on site. A correction would count as confirmed only once that support agreed.</p><p>The second is <strong>user accounts</strong>, so changes are traceable. At the moment anyone with the link can create, edit or delete records, and nothing records who did what. With accounts, every change would carry a name and a time. Mistakes would be easier to trace and undo, and there would be a way to deal with abuse, which an open API can’t offer.</p><p>The third is <strong>a real database</strong> to replace JSON Server. JSON Server keeps everything in one file, which is fine for a demo but poor for several people editing at once. A real database handles many people writing at the same time and can keep a history of changes instead of overwriting them. It can also answer spatial questions directly, such as which buildings are within 200 metres of where someone is standing. PostGIS would suit that.</p><p>The fourth is <strong>a smartphone-friendly design</strong>, because the people standing at the place are holding a phone, not sitting at a laptop. The layout already stacks on narrow screens, and there is a “Use my location” button, but I designed it for a desktop first. A phone version needs bigger touch targets and a form that doesn’t hide the map when the keyboard opens. It also has to cope with weak signal, and a phone’s GPS reading is only approximate.</p><p><a href="https://andyaketch.github.io/Kwa-Ground/">https://andyaketch.github.io/Kwa-Ground/</a></p><p>If you’ve ever walked to a pin that wasn’t there, this one is for you.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a5086ec0b5a8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why the Most Useful AI in Geospatial Might Be the Boring One]]></title>
            <link>https://medium.com/@andyaketch/why-the-most-useful-ai-in-geospatial-might-be-the-boring-one-8131c0b889cf?source=rss-cd5c922bb751------2</link>
            <guid isPermaLink="false">https://medium.com/p/8131c0b889cf</guid>
            <category><![CDATA[building-construction]]></category>
            <category><![CDATA[surveying]]></category>
            <category><![CDATA[gis]]></category>
            <category><![CDATA[trimble]]></category>
            <dc:creator><![CDATA[andyaketch]]></dc:creator>
            <pubDate>Mon, 05 Oct 2026 07:29:44 GMT</pubDate>
            <atom:updated>2026-10-05T07:29:44.411Z</atom:updated>
            <content:encoded><![CDATA[<p><em>At this year’s Esri User Conference, a Trimble executive made a case that goes against the AI hype. His argument was that in infrastructure work, the best AI is the one that gives the same answer every time. The reporting comes from xyHt’s coverage of Trimble’s Unity Suite and SiteVision. These are Trimble’s claims, not independent findings, and I’ve tried to treat them that way.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*y4zjqTyGZLgGBnPvhe00Xw.jpeg" /><figcaption>Photo by <a href="https://unsplash.com/@nasimuddin872?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Nasim Uddin</a> on <a href="https://unsplash.com/photos/a-yellow-surveying-instrument-stands-outdoors-MK2ZSaI0W2k?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Unsplash</a></figcaption></figure><h3>Repeatability matters more than cleverness</h3><p>If you ask any general-purpose chatbot the same question two times, there will always be two answers that will differ a little bit. It may be acceptable in the case of writing an email, but not in deciding which bridge needs inspection or if an order should be issued. The position of Trimble, according to the source cited above, is that one wrong answer in asset management means millions lost; thus, the company prefers to release AI features late but ready.</p><p>This position is quite useful for everyone who joins GIS now. GeoAI articles are mostly oriented on what the model is capable of doing, but the main issue is if the same answer will come tomorrow and if you will be able to prove its legitimacy to the person who will need to approve it.</p><h3>One database instead of two</h3><p>It would be the most obvious case in the article but also the least obvious one. The GIS manager from the city saw during the demonstration how his authority’s authoritative GIS data and data management system could peacefully coexist. Before this point in time, any change on either side had to be done manually on the other, while the authority had to manage three GIS databases at once.</p><p>Anyone with the experience of merging mismatching layers knows about the problem. Duplicates become desynchronized, accumulating errors in the process. All the practical skills in the sphere have much more value than the flashy analysis.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ujV6ojJrsvu2Z7CiAuBNkQ.jpeg" /><figcaption>Photo by <a href="https://unsplash.com/@sunburned_surveyor?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Scott Blake</a> on <a href="https://unsplash.com/photos/macro-shot-of-gray-and-yellow-surveyor-1GW45VDkFI0?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Unsplash</a></figcaption></figure><h3>Precision needs context, and context needs precision</h3><p>The Trimble representative spoke of the relationship between Trimble and Esri as follows: Esri offers spatial awareness, while Trimble offers precise location, but both elements cannot exist alone. He explained using the example of a flooded highway where the storm drains are buried and invisible to us. With precise knowledge of their locations, a team can reach the location directly.</p><p>Both components are necessary. With the knowledge of the exact location of a point, one can do nothing without placing it on an analysable map, and vice versa.</p><h3>Convenience beats millimetres, up to a point</h3><p>In SiteVision, we discussed how to generate georeferenced point clouds using the lidar on the latest generation of the iPhone or iPad together with the GNSS receiver but not with the help of a drone or scanner. This method has been called by their customers “the human drone&quot;. It works great in dense vegetation and even monitors changes from day to day, such as the stockpile levels, and he mentioned that what is more important for users is mobility and speed rather than precise centimetric measurements.</p><p>What I liked was his directness in talking about limitations. The upper limit of accuracy depends on GNSS technology, and he said that the project which requires tolerances of an inch or better does not fall into the scope of this software.</p><h3>Guidance, not gospel</h3><p>Trimble is now developing integrations where data about utilities from other sources is brought to the field view, allowing crews to determine the potential locations of utilities prior to excavation. It is essential to note that the company specifies the fact that it is purely for guidance and not a replacement of the confirmation of the existence of a utility.</p><h3>What I take from it</h3><p>This is not about some particular software package; rather, it is about how value is being increasingly realised in the relationships among tools: clean data, accurate inferences, and consistency of results. This is what should inform our education of students and beginning analysts and what skills they need to develop: reproducible pipelines, sound data management practices, and accurate performance assessment.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8131c0b889cf" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Your Machine Learning Model Needs to Explain Itself .]]></title>
            <link>https://medium.com/@andyaketch/why-your-ai-model-needs-to-explain-itself-a-plain-english-guide-to-shap-cb46fcb18296?source=rss-cd5c922bb751------2</link>
            <guid isPermaLink="false">https://medium.com/p/cb46fcb18296</guid>
            <dc:creator><![CDATA[andyaketch]]></dc:creator>
            <pubDate>Mon, 31 Aug 2026 08:21:01 GMT</pubDate>
            <atom:updated>2026-09-15T07:20:06.327Z</atom:updated>
            <content:encoded><![CDATA[<h3>Why Your Machine Learning Model Needs to Explain Itself: A Plain-English Guide to SHAP</h3><p>Although machine learning algorithms have proven themselves to be very good at prediction, they are typically poor at justifying those predictions. If you query an algorithm about its reasons for rejecting a loan application, detecting a fraudulent transaction, or estimating the probability of a readmission of a patient, the answer will likely be a deafening silence.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*p9oFleAqnS9gM-MpsUCg-A.jpeg" /><figcaption>Photo by <a href="https://unsplash.com/@markuswinkler?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Markus Winkler</a> on <a href="https://unsplash.com/photos/a-close-up-of-a-typewriter-with-a-paper-reading-machine-learning-f57lx37DCM4?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Unsplash</a></figcaption></figure><p>You will get a number, but the justification will remain locked up in thousands, or even millions, of hidden parameters which cannot be read by anyone directly. This is the infamous “black box” problem, and this is a serious issue. The hospital cannot take action based on an unexplained risk score; a bank cannot reject a credit application without a justified reason, and researchers cannot trust any results that cannot be questioned. SHAP is one of the most widely used tools for prying that box open.</p><h3>What SHAP is</h3><p>The acronym SHAP stands for SHapley Additive exPlanations. It can sound rather long and intimidating, yet the basic idea behind it comes from a more ancient field of science: game theory.</p><p>During his studies in the 1950s, an economist, Lloyd Shapley, invented a system of distributing payouts between players of a cooperative game according to their individual contribution. In case four people cooperated and won a prize, who would get how much money, and why? The method involved considering all groups of players one by one, estimating the increase in payoff that each group produced, and finding an average of all those contributions.</p><p>Similarly to Shapley’s approach, SHAP looks at every input variable in a machine learning model as a player, with the output being a prize or payoff. Just like the original, it asks a simple question: how much would this particular feature increase the output if considered with all possible combinations of other features for this specific output? The result is a SHAP value for each feature of every prediction. Positive value indicates the increase in prediction caused by this feature, while a negative one indicates a decrease. Summing up all SHAP values for a single prediction with a baseline gives the exact output. While other methods give you a list of features that were important for the model in general, SHAP will tell you which features were important in this particular case for this particular person or row. And this is the main idea behind it.</p><p>Let’s take a loan application as an example again. The model might heavily depend on income in its training data. But, say, for one particular applicant, the crucial feature was not income but lack of credit history. A global explanation saying “the model relies heavily on income” is correct, but it is of no use to this particular applicant. It can help him understand that his lack of credit history had the most impact on the result of the prediction. This is a difference between statistics and answers.</p><p>This is also the reason why SHAP is now practically a must for regulated industries like healthcare, finance, or insurance, where the pressure to justify decisions made by automated systems grows. “The model said so” won’t help in an appeal.The SHAP method describes what happened with the model but does not prove its correctness. If the model learnt an incorrect pattern during training, the SHAP values will reflect this learning regardless of any flaws. Think of SHAP as a diagnostic method rather than a cure.</p><h3>A concrete example</h3><p>Now picture a model which predicts the prices of houses and aims at predicting the price of one specific house for $340,000. Its base estimate, what the model would predict without any data at all, equals $280,000. And now SHAP tells us how much each feature contributes to the target value:</p><p>- Location contributes $45,000<br>- Size contributes $20,000<br>- The age of the house contributes $(-8,000)<br>- The number of bathrooms contributes $3,000</p><p>When you sum up all those contributions to the base estimate, you get $340,000. Each dollar of the target value is explained by this method. That’s the kind of explanation a homeowner, an appraiser, or a sceptical buyer can actually work with.</p><h3>What SHAP doesn’t do</h3><p>It is very computationally expensive. The calculation of exact Shapley values involves numerous combinations of features. This becomes increasingly difficult to calculate as the number of features increases. That’s why many real-world applications use approximate algorithms such as TreeSHAP for tree-based models and KernelSHAP for others, which sacrifice some accuracy for faster calculations.</p><p>Moreover, there is an issue of misinterpretation of SHAP values. The SHAP values do not determine the cause-and-effect relationships between features and outcomes. It shows contribution in terms of a specific model that was trained on specific data. Correlated variables create another problem. When two features increase at the same time, such as the square footage of a house and the number of rooms within it, the algorithm fails to separate them during training, as does SHAP. Rather than attributing the entire impact on the price to one feature, SHAP divides it among both. While the size of the house may increase its price due to both square footage and the number of rooms, both of which essentially refer to the same characteristic of the house, it makes the feature appear not as significant as it really is.</p><h3>The bigger picture</h3><p>Explainability of AI systems is all about resolving a basic conflict, which is that AI models have become much better and more complex but less understandable at the same time. While SHAP cannot resolve this conflict itself, it does provide people dealing with AI, be they scientists, regulators, or end users, with a unique opportunity of asking “Why?” and getting an answer based on the actual workings of the model rather than assumptions made about its behaviour.</p><p>This is a small but proper step forward. Tools for explainability do not trust models; they allow inspecting them.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cb46fcb18296" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[GeoAI Has a Geography Problem]]></title>
            <link>https://medium.com/@andyaketch/geoai-has-a-geography-problem-73d333a41845?source=rss-cd5c922bb751------2</link>
            <guid isPermaLink="false">https://medium.com/p/73d333a41845</guid>
            <category><![CDATA[data-center]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[sustainability]]></category>
            <category><![CDATA[geoai]]></category>
            <dc:creator><![CDATA[andyaketch]]></dc:creator>
            <pubDate>Mon, 17 Aug 2026 08:08:25 GMT</pubDate>
            <atom:updated>2026-08-17T08:08:25.141Z</atom:updated>
            <content:encoded><![CDATA[<p>The Spatial Balance: Why GeoAI’s Carbon Problem Isn’t Just About Energy</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*5vEvvqwQG4XL_sITgW1z5g.jpeg" /><figcaption>Photo by <a href="https://unsplash.com/@nasa?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">NASA</a> on <a href="https://unsplash.com/photos/photo-of-outer-space-Q1p7bh3SHj8?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Unsplash</a></figcaption></figure><p>Most people have never thought about it. A Google search uses roughly 0.5 grams of CO₂ and draws about 0.3 Wh of energy. An average request made to generative AI services requires closer to 2.9 Wh, almost ten times more than a regular search engine lookup. Multiply that across hundreds of millions of daily requests, and it’s easy to see why data centres, once an invisible piece of internet infrastructure, are now showing up in climate policy debates.</p><p>For most of the AI world, this is already an active conversation. But for GeoAI, whose entire purpose lies in analysing remote sensing, LiDAR, trajectory data, and other forms of geospatial inputs? That’s the gap worth talking about.</p><h3>Two questions, one framework</h3><p>There’s a useful way to split this problem in half.</p><p>Geography for AI sustainability asks where AI physically runs. If you take two otherwise identical models, trained on precisely the same task, but run them on electricity drawn from two grids with radically different carbon intensities, they’ll produce vastly divergent results. The range could be anywhere between 200 and 755 gCO₂/kWh depending on your grid’s fuel sources, and even within individual grids, running the training workload at noon versus midnight makes an 8% difference. Training happens around the clock.</p><p>Sustainability in GeoAI flips that question back onto itself: What’s the planetary cost of constructing and deploying this particular form of AI? Models trained on text data benefit from economies of scale because all text data looks rather similar . But GeoAI models deal with satellite imagery, radar, point clouds, trajectories, and datasets that never stop updating, making scaling far more difficult.</p><h3>The part nobody puts on a slide: water</h3><p>Carbon gets most of the attention, but water consumption is the quieter crisis. Data centres rely on evaporative cooling, and the efficiency of that process, measured by something called &#39;Water Usage Effectiveness&#39; (WUE), can vary tenfold depending on where a facility sits. Facilities in arid regions fare the worst.</p><p>Google’s data centres in The Dalles, Oregon, are the go-to example here. Water use there has nearly tripled in five years, and the facilities now account for roughly a quarter of the entire town’s water consumption. That’s not an abstract sustainability metric anymore. It’s a local resource fight.</p><p>There’s also a genuinely counterintuitive tension buried in here: abundant solar energy, which sounds like an unambiguous climate win, often correlates with hotter, more arid conditions, which means more cooling demand, which means more water. Clean energy and water scarcity can pull in opposite directions.</p><h3><strong>Whose emissions, whose benefit?</strong></h3><p>Another interesting idea in the sustainability-for-GeoAI conversation isn’t a new efficiency trick. It’s a fairness question.</p><p>Train up a global pandemic such as a COVID-19 forecasting model. Yes, there’s going to be emissions involved, but the benefit is obviously global too; its training footprint is arguably justified. Now train a traffic prediction model for LA out of a data centre with high water scarcity in India? Suddenly people start having second thoughts. Why? Because the populations paying the price for training that thing don’t seem to be reaping any reward.</p><p>That’s the logic behind a proposed metric called the Spatial Benefit-Cost Ratio (SBCR). It compares the population that benefits from a model against the population bearing its resource costs, weighted by how “close” those two groups are, interactively speaking. A ratio near 1 means training and benefit are roughly co-located. A ratio near 0 means the costs are quietly landing somewhere far away. It’s a simple idea, but a useful one: GeoAI’s use cases are so often local that a tool like this, matching cost to benefit geographically, is overdue.</p><h3><strong>Why GeoAI can’t just borrow AI’s playbook</strong></h3><p>It would be convenient if GeoAI could just adopt the sustainability tooling built for large language models. It can’t, not entirely. Two things set it apart.</p><p>Geospatial datasets move. Cities grow, coastlines adapt, and farmland moves based on seasonal considerations. The land cover classification system that worked perfectly last year may now be considered stale, meaning more frequent retrainings for GeoAI models than for a generic language model. Across the lifecycle of any single GeoAI system, this leads to higher power usage, not lower.</p><p>GeoAI is also multimodal in a way text models aren’t. One model might need to reconcile optical imagery, SAR, LiDAR point clouds, and trajectory data, each with its own geometric and semantic structure. Building a single foundation model that handles all of that well (a “Geo-FM”) is a much harder problem than building a text-only foundation model. Early examples like Prithvi-100M for remote sensing are still narrow compared to their LLM counterparts.</p><h3><strong>The infrastructure squeeze nobody’s watching</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hT3-MK-G3-LyMB6tvAY2HA.jpeg" /><figcaption>Photo by <a href="https://unsplash.com/@ericstoynov?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Eric Stoynov</a> on <a href="https://unsplash.com/photos/two-server-racks-filled-with-electronic-components-and-wires-k27hkqXuveo?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Unsplash</a></figcaption></figure><p>Zoom out from individual models and there’s a bigger structural story: What’s becoming clear now is that AI’s expansion rate is starting to exceed its ability to be supported by physical grid systems.</p><p>Current projections suggest Africa’s data centers will increase from around 0.4 gigawatts (GW) today to 2.2 GW by 2030. This is estimated to require investments in the range of $10 to $20 billion dollars. Again, those numbers aren’t speculative at all ;they have the very real possibility of changing national-scale energy planning. And the cautionary tales are already written: Ireland’s data centers hit 22% of the country’s metered electricity by 2024, prompting new grid connection restrictions, and Amsterdam has simply banned new data centers until 2030. Regional power-sharing through mechanisms like the West African Power Pool is one proposed fix, and it’s exactly the kind of geographic thinking this conversation needs more of.</p><h3>What actually helps</h3><p>None of this is intended as an argument against GeoAI. What it really is, is a case for approaching GeoAI development with more intentionality behind it.</p><p>Retrieval-Augmented Generation (RAG): Combine a small, pre-trained generative model with an external storehouse of knowledge rather than rebuilding your giant pre-trained model from the ground up with each update.</p><p>Extreme quantisation: BitNet’s 1.58-bit parameters represent massive reductions in storage requirements, energy costs, and computational needs for negligible losses in performance quality.</p><p>Transparency indexes: The Foundation Model Transparency Index scores development teams based on the completeness of documentation regarding their methods for training AI systems.</p><p>Power-sharing regions, such as the West African Power Pool, may also prove useful.</p><h3>The bottom line</h3><p>Geospatial data has always been about where things are. It’s a little ironic, then, that the AI models built to analyse that data have mostly ignored where they themselves physically live: which grid they’re drawing from and which community bears the water cost.</p><p>The spatial balance is a reminder to map GeoAI’s own footprint with the same rigour we bring to everything else we map.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=73d333a41845" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Fifteen Years of Data About Kiambu’s Ecological Health]]></title>
            <link>https://medium.com/@andyaketch/fifteen-years-of-data-about-kiambus-ecological-health-9fd3fe32ccf1?source=rss-cd5c922bb751------2</link>
            <guid isPermaLink="false">https://medium.com/p/9fd3fe32ccf1</guid>
            <category><![CDATA[kiambu]]></category>
            <category><![CDATA[ecology-and-environment]]></category>
            <category><![CDATA[remote-sensing]]></category>
            <category><![CDATA[kenya]]></category>
            <category><![CDATA[ecology]]></category>
            <dc:creator><![CDATA[andyaketch]]></dc:creator>
            <pubDate>Mon, 03 Aug 2026 12:48:12 GMT</pubDate>
            <atom:updated>2026-08-03T12:48:12.811Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ln2369CxaPEnHDlhoCjUuA.jpeg" /><figcaption>Photo by <a href="https://unsplash.com/@iamr_r_r?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Iryna Marienko</a> on <a href="https://unsplash.com/photos/a-river-running-through-a-lush-green-valley-zizYAD6Pm2Y?utm_source=unsplash&amp;utm_medium=referral&amp;utm_content=creditCopyText">Unsplash</a></figcaption></figure><p>Kiambu County borders Nairobi on its northern side, and that alone has made it one of the more closely watched landscapes in Kenya. A recent remote sensing study I did sets out to measure something harder to pin down than land use alone: ecological quality itself, using a tool called RSEI, comparing the county in 2010 against the county in 2025.</p><p>Here’s what I found.</p><h3>What RSEI Is Measuring</h3><p>RSEI stands for Remote Sensing Ecological Index. It pulls together four signals from satellite imagery: how much vegetation is present through the Normalised Difference Vegetation Index (NDVI); how much moisture the land is holding through Tasselled Cap – Wetness (TCT-W); how hot the surface runs through Land Surface Temperature (LST); and how much bare or built ground is exposed through the Normalised Difference Built-up and Soil Index (NDBSI). Those four get combined into a single score between 0 and 1, where anything close to 1 means the land is in a strong ecological condition and anything near 0 translates to a poor ecological condition. This is achieved by use of the first principal component (PC1).</p><p>The point of doing this is that any single indicator such as greenness through NDVI alone can be misleading. A patch of land can look vegetated on a basic index and still be under real stress from heat or dryness. RSEI is an attempt to catch that.</p><h3>Back in 2010</h3><p>In 2010 the county’s average RSEI score was 0.613. Under the grading system the study used, that put nearly half the county, 48.7%, in the “Good” band, with another 18.2% reaching &quot;Excellent&quot;. Most of what remained (about 29.7%) sat in &quot;Moderate&quot;, and only a thin slice, 3.4%, dropped to &quot;Fair&quot;. Nothing scored &quot;poor&quot;.</p><p>That’s a landscape doing reasonably well. More than two thirds of the county were sitting at good or better.</p><h3>Fast forward to 2025</h3><p>By 2025 the average had fallen to 0.561, lower than where it started fifteen years earlier. The distribution shifted with it. “Good” land dropped to 22.4%. “Excellent” disappeared entirely; not a single pixel in the county qualified. “Moderate” swelled to 54.2%, and “Fair” grew from that earlier 3.4% up to 23.4%.</p><p>The scores also spread out more. In 2010 the county was fairly consistent in its ecological condition. By 2025 the gap between the best and worst parts of Kiambu had widened, meaning healthy pockets and degraded ones sat closer together on the map than before.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*w3dKIoj55oW47TPDSHKuIw.png" /><figcaption>Fifteen years of ecological shift: Side-by-side comparison illustrating environmental change across Kiambu County from 2010 to 2025.</figcaption></figure><h3>Where the Quality Sits</h3><p>The study also broke RSEI down by what kind of land it was measuring, and one pattern held steady across both years: forested areas scored highest while bare, exposed ground scored lowest. In 2010 forested land averaged around 0.861. By 2025 that had slipped to 0.783, still comfortably the healthiest land type in the county, but showing wear. Bare soil went the other way and went further. It averaged around 0.400 in 2010 and fell to 0.303 by 2025, the steepest decline of any land condition in the study.</p><p>Built-up land sat near the bottom in both years too, which is not a surprising place for it to land. Everything else, land that kept some vegetation on it, tended to hold a moderate to good score in both periods.</p><h3>Behind the scenes</h3><p>I also tried to work out what was actually driving these scores, using a modelling technique called SHapley Additive exPlanations that ranks how much each factor contributes to the overall ecological health. Rainfall came out on top in both years by a wide margin, which tracks with how much the four RSEI components depend on moisture. Distance from the forest mattered too; land closer to the forest tended to score better, and land further away tended not to.</p><p>That part of the story didn’t change much between 2010 and 2025. What did shift, and it’s worth a brief mention, is that built-up density became a noticeably stronger factor in 2025 than it had been in 2010, suggesting that as construction expanded, it started pulling more weight in determining a location’s ecological score than it used to.</p><h3>The bigger picture</h3><p>Kiambu’s ecological trajectory over this period wasn’t a straight line down. It was a peak followed by a retreat. The county in 2010 was in better ecological shape than the county in 2025, and the loss wasn’t spread evenly. Forests held up better than most land types but still lost ground. Bare soil, already the weakest performer, degraded further and fastest.</p><p>What the numbers leave you with is a fairly plain read: the parts of Kiambu that kept vegetation cover kept their ecological quality too, more or less, and the parts that didn’t fell the furthest. Fifteen years earlier the county looked to be moving in the right direction. By 2025 a real share of that ground had been given back.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9fd3fe32ccf1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[5 Geospatial Analysis Applications Riding the 2026 Wave]]></title>
            <link>https://medium.com/@andyaketch/5-geospatial-analysis-applications-riding-the-2026-wave-46abd029cc0c?source=rss-cd5c922bb751------2</link>
            <guid isPermaLink="false">https://medium.com/p/46abd029cc0c</guid>
            <category><![CDATA[geospatial-intelligence]]></category>
            <category><![CDATA[geoai]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[gis]]></category>
            <dc:creator><![CDATA[andyaketch]]></dc:creator>
            <pubDate>Mon, 27 Jul 2026 13:03:54 GMT</pubDate>
            <atom:updated>2026-07-27T13:03:54.637Z</atom:updated>
            <content:encoded><![CDATA[<p>The use of geospatial analytics has been silently evolving from being a niche cartography skill to becoming one of the decision-making components in non-map-related fields, such as insurance, banking, agriculture, and emergency response. The difference in 2026 is not about the underlying concept of using location as the driving factor but about how quickly location intelligence can be created. Below are five examples that demonstrate what is really happening in the geospatial world this year.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1RHmJ3-5Kp4MvO040bXiMA.jpeg" /><figcaption>Image courtesy of Jo Lin (Unsplash)</figcaption></figure><h3>1. Automated Feature Extraction</h3><p>The digitisation of roads, buildings and other infrastructures from satellite or aerial imagery was previously one of the most time-consuming parts of GIS workflows. It was done manually through tracing the outlines from one tile to another, which could take quite a lot of time. This is increasingly being handed off to deep learning models trained to pick out these features directly from imagery, cutting extraction time from days to hours.</p><p>The foundation models such as Prithvi, Clay, SatMAE, and GeoFM-type models are trained using large amounts of satellite and aerial imagery. As a result of the wide training, it became unnecessary to create new classifiers from scratch for each new region. Instead, the general-purpose model is fine-tuned with the local training set, and this allows for achieving positive results for various tasks, from land cover mapping to the analysis of the post-disaster damages to infrastructure. All of these were previously requiring specific machine learning skills but now are getting closer and closer to standard GIS deliverables.</p><h3>2. Climate and Disaster Risk Analytics</h3><p>While flood zones, wildfires, droughts, and heat have been scientific research topics, they are now becoming integral components that shape the decision-making process in various businesses. Insurers use these technologies to assess risks at the building level, while real estate developers use them to analyse potential locations. Agricultural and logistics companies rely on this technology to forecast possible disruptions and act in order to prevent them from happening or minimise the impact. The distinguishing feature in 2026 is automation, as imagery-based hazard detection becomes one of the response strategies for government agencies when there is a disaster going on, rather than a way to conduct a post-disaster assessment.</p><h3>3. Digital Twins and 3D Reality Capture</h3><p>3D GIS and digital twins are moving beyond simple visualisation into a realm where they are used as means of simulation. They allow engineers and planners to run through multiple “what if” scenarios related to infrastructure, utilities, and city planning before expending any resources on them. In the rapidly changing world, novel visualisation technologies are being developed to be used as part of the geospatial workflow process. One of the most interesting technologies that have been introduced recently is called &#39;Gaussian Splatting&#39; and allows photorealistic rendering of the scene. This offers photorealistic scene rendering that complements traditional point clouds and meshes with photorealistic imagery without the cost of full photogrammetric surveys.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*0QCyUg7F_QYl7chuWWGvDw.jpeg" /><figcaption>Image courtesy of Kenny Murgor (Unsplash)</figcaption></figure><h3>4. Predictive GIS for Public-Sector Risk Forecasting</h3><p>There have been more efforts made by governmental and business enterprises to train models on large amounts of spatial data over many years in order to be able not only to map events such as flooding, subsidence, and coastal erosion but also to predict their occurrence. At the same time, both the pre-event and post-event analysis are expected to work as an audit trail for future planning and regulation. Yet, one of the main caveats is prediction quality depends heavily on data quality.</p><h3>5. Open, Cloud-Native Geospatial Data</h3><p>The constant and continuous growth of high-resolution imagery available through satellite imagery providers such as Sentinel and Landsat and other openly available government datasets, together with OpenStreetMap data, has greatly reduced the entry barrier into geospatial analytics. Together with cloud-based technologies and an API-driven approach to providing services and tools, this means that small teams and individual users are able to perform geospatial analysis that was earlier possible only within the confines of an enterprise environment, with its costly data licensing agreements and expensive infrastructure. This is also why a big part of GeoAI and predictive analytics have spread so quickly outside large institutions.</p><p>Geospatial analysis in 2026 is less about producing a map and more about producing a decision: a risk score, a forecast, a simulated outcome. The tools have gotten cheaper and faster; the value has shifted toward the judgement of the person interpreting them.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=46abd029cc0c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Left Behind on Foot: Mapping Healthcare Access in Kenya’s Forgotten Drylands]]></title>
            <link>https://medium.com/@andyaketch/left-behind-on-foot-mapping-healthcare-access-in-kenyas-forgotten-drylands-6115305d2d70?source=rss-cd5c922bb751------2</link>
            <guid isPermaLink="false">https://medium.com/p/6115305d2d70</guid>
            <category><![CDATA[healthcare-accessibility]]></category>
            <category><![CDATA[postgis]]></category>
            <category><![CDATA[gis]]></category>
            <dc:creator><![CDATA[andyaketch]]></dc:creator>
            <pubDate>Mon, 20 Jul 2026 10:04:16 GMT</pubDate>
            <atom:updated>2026-09-15T01:08:07.006Z</atom:updated>
            <content:encoded><![CDATA[<p><em>How I used PostgreSQL, PostGIS, and pgRouting to measure how far 123 health facilities can actually reach ,and what the numbers revealed about life without a vehicle in arid Kenya.</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*j-Rpdjx_RPMd48mAxelgOA.jpeg" /></figure><p>Imagine you’re ailing. Not the emergency type of ailment, but the kind that needs a doctor. You live in Mandera County, in Kenya’s north-eastern drylands. The nearest health facility is about 40 kilometres away. You don’t own a vehicle. The road, if it exists at all on a map, is an unpaved track that disappears deep into scrubland.</p><p>How long does it take you to get there?</p><p>This is the key question that motivated my research. Not just an idea you see on the 9PM news in theory, not with straight-line distance, but through actual road network analysis: tracing the paths a vehicle or a person on foot would realistically take to reach a health facility across Kenya’s Arid and Semi-Arid Lands (ASAL) region.</p><p>The findings were bleaker than I expected.</p><h3>Area of Interest</h3><p>I selected four Kenyan ASAL counties: Garissa, Wajir, Mandera, and Isiolo, which cover roughly a third of the country’s land area but hold some of its most underserved populations. The terrain is harsh, the climate harsher, the road network sparse, and health infrastructure stretched thin beyond its capacity across immense distances.</p><p>This is not a region where you assume healthcare is easy to reach. But assumption is not analysis. I wanted facts, numbers.</p><h3>The Data</h3><p>The analysis draws on three datasets:</p><ul><li><strong>Road network</strong>: 5,914 road features covering the ASAL region, sourced from Kenya’s national road dataset as well as OpenStreetMaps datasets. Road classes range from Type A highways to unclassified tracks, with surface conditions recorded from Excellent to Very Poor.</li><li><strong>Health facilities</strong>: 123 facilities pulled from OpenStreetMap via Overpass Turbo. Facilities span all four counties.</li><li><strong>Administrative boundaries</strong>: County boundaries for spatial context.</li></ul><p>The road data was at a national scale and therefore had to be clipped on QGIS 4.01 before subsequent steps. All data was loaded into PostgreSQL with the PostGIS and pgRouting extensions. Coordinates were projected to UTM Zone 37S (EPSG:32737) for accurate distance calculations in metres.</p><h3>The Method</h3><p>The core of this project is network routing<strong>. </strong>That is treating the road system as a mathematical graph where intersections are nodes and road segments are edges with travel costs.</p><h3>Building the Network</h3><p>Roads from the national data source frequently appear in a MultiLineString format, that is, bundled road segments that comprise a single feature. However, pgRouting needs individual LineString features with one start and one endpoint. This required the first step to split 5,914 multi-part features into 9,493 road edge features using the ST_Dump function of PostGIS.</p><p>From those edges, pgRouting’s pgr_extractVertices identified 7,737 unique network nodes, i.e., every point where roads meet, branch, or end.</p><h3>Calculating Travel Costs</h3><p>For this first model, two travel speeds were applied uniformly across all road types:</p><ul><li><strong>Walking: 5 km/h</strong> which is a standard assumption for pedestrian travel</li><li><strong>Driving: 50 km/h</strong> which is a conservative average for mixed road conditions in the ASAL region</li></ul><p>The travel cost per edge was calculated as the length of the road segment (in metres) divided by speed (in metres/second), resulting in the cost in seconds per road edge. For example, the cost of traversing a 1,000-metre road segment equals 720 seconds when walking and 72 seconds when driving.</p><p>Flat speed assumptions are a deliberate simplification. A very poor-condition earth track cannot realistically be driven at 50 km/h; similarly, good roads also encourage higher speeds within the given speed limits. Refining costs by road surface condition and class is planned for the next iteration.</p><h3>Snapping Facilities to the Network</h3><p>Hospital footprints were converted to centroids using ST_Centroid. All 123 facility points were then snapped to their nearest road network vertex using ST_Distance nearest-neighbour matching.</p><p>The snapping distance is crucial. The furthest facility from its nearest road vertex was 2,800 metres translating to roughly 33 minutes of walking before even reaching the road network. In my model, such access is not considered, and hence travel time is slightly overestimated for remote facilities.</p><h3>Running the Routing</h3><p>pgr_drivingDistance was run for all 123 facilities across both travel modes at a maximum threshold of 4 hours. Rather than running multiple executions based on various time threshold values, I executed a single run up to the highest threshold value (i.e., 4 hours) and then filtered the output while generating isochrones giving four time cuts per routing execution per facility per transport mode.</p><p>Using PostgreSQL’s CROSS JOIN LATERAL, all 246 routing runs (123 facilities × 2 modes) were executed in a single SQL statement, storing every reachable vertex and its accumulated travel time.</p><h3>Building Isochrones</h3><p>Reachable vertices for each facility, mode, and time threshold were collected and passed to ST_ConcaveHull PostGIS, a function that draws a tight polygon around a set of points. The result: 984 individual isochrone polygons (123 facilities × 2 modes × 4 thresholds).</p><p>These individual isochrones were merged using the ST_Unionoperation for generating eight regional isochrones showing total reachable area across all facilities for each mode and threshold combination.</p><h3>The Findings</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Qs3VT_ES63Q0dldsn3rICQ.png" /><figcaption>On the left map is the walkable coverage while on the right map is the driveable coverage for 123 health facilities in Garissa, Wajir, Mandera and Isiolo Counties. The shades represent different thresholds in terms of time which are 30 minutes, 1 hour, 2 hours and 4 hours. It clearly shows the difference in terms of coverage in that the walkable coverage is so small and almost negligible compared to the driveable coverage that almost covers all the mapped roads in 4 hours. Note: The polygons are just the concave hulls around reachable roads vertices and therefore have gaps between roads.</figcaption></figure><h3>The Headline</h3><p><strong>Driving covers 46 times more area than walking within 4 hours.</strong></p><p>Within 4 hours of travel from any health facility in the region:</p><ul><li><strong>Walking reaches 2,282 km²</strong></li><li><strong>Driving reaches 105,744 km²</strong></li></ul><p>Walking coverage is 2.2% of driving coverage. For communities without vehicle access, which in ASAL counties is a significant proportion of the population, the effective healthcare catchment area collapses to almost nothing.</p><h3>The 30-Minute Picture</h3><p>The 30-minute threshold reveals the most acute gap:</p><ul><li><strong>Walking 30 minutes</strong> from all 123 facilities combined covers <strong>34.7 km²</strong></li><li><strong>Driving 30 minutes</strong> covers <strong>4,077.9 km²</strong> which translates to nearly 120 times more</li></ul><p>For someone walking, a 30-minute radius from a health facility covers an area roughly equivalent to a small town. For someone with a vehicle, it covers a region.</p><h3>Isolated Facilities</h3><p>Ten facilities effectively cannot be reached via the mapped road network at all. The routing algorithm could find only a single reachable vertex from these facilities meaning they sit in network dead-ends or in areas where roads exist on the ground but have never been digitized.</p><p>These are not just hard-to-reach facilities. They are, from a network routing perspective, unreachable:</p><ul><li>Umma Medical Centre</li><li>Dunto Dispensary</li><li>Iftin Sub County</li><li>Manyalo Roadside Medical Clinic</li><li>Wargadud Dispensary</li><li>Burmayo Mandera Dispensary</li><li>Dilmanyale Dispensary</li><li>Hungai Dispensary</li><li>Bute Hospital</li><li>Mugurnanyore Simba Clinic</li></ul><p>Whether this reflects incomplete road digitization or genuine infrastructure absence is a question the data alone cannot answer, but it flags exactly where ground-truthing is needed most.</p><h3>What This Means</h3><p>The images show an area where being able to get a ride to your destination isn’t a nicety but something that makes the difference between accessing healthcare and not. It’s an example of the geography of a mother travelling to a health facility with a sick child in Wajir County in comparison with a family that uses a car; it’s more than just slow travel but a completely different geography of the facilities accessible to her.</p><p>This method provides the ability to see the issue from a perspective impossible to convey through anecdotes. What are the areas that have a lack of healthcare coverage? What are the facilities that cover the largest number of people within a 30-minute radius? What communities find themselves outside all isochrones?</p><h3>Technical Stack</h3><p>Database-PostgreSQL 15 + PostGIS + pgRouting 4.0.1</p><p>Road topologypgr_extractVertices, ST_Dump, ST_Equals</p><p>Routingpgr_drivingDistance via CROSS JOIN LATERAL</p><p>IsochronesST_ConcaveHull, ST_Union</p><p>Web map Leaflet.js, CartoDB dark basemap</p><p>Projection : EPSG:32737</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6115305d2d70" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>