<?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 Alice Heiman on Medium]]></title>
        <description><![CDATA[Stories by Alice Heiman on Medium]]></description>
        <link>https://medium.com/@aliceheimanxyz?source=rss-838957cd1468------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*T6EVWoVhtefnOoL3mk5WAw.jpeg</url>
            <title>Stories by Alice Heiman on Medium</title>
            <link>https://medium.com/@aliceheimanxyz?source=rss-838957cd1468------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 10:16:16 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@aliceheimanxyz/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[7 Lessons on AI for Science from Nobel Prize Dialogue London 2026]]></title>
            <link>https://medium.com/@aliceheimanxyz/7-lessons-on-ai-for-science-from-nobel-prize-dialogue-london-2026-412c6c75e424?source=rss-838957cd1468------2</link>
            <guid isPermaLink="false">https://medium.com/p/412c6c75e424</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[research]]></category>
            <category><![CDATA[nobel-prize]]></category>
            <dc:creator><![CDATA[Alice Heiman]]></dc:creator>
            <pubDate>Tue, 23 Jun 2026 15:31:00 GMT</pubDate>
            <atom:updated>2026-06-23T15:31:00.910Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LDJrRX9tkKTzJxjMN-tfXg.png" /></figure><p>I recently attended <a href="https://www.nobelprize.org/events/nobel-prize-dialogue/london-2026/">Nobel Prize Dialogue London</a> featuring Demis Hassabis, Paul Nurse, and Alison Noble.</p><p>The focus: AI for Science.</p><p>In the <a href="https://www.youtube.com/watch?v=zVrA-_WP9Mw&amp;source_ve_path=MjM4NTE&amp;embeds_referring_euri=https%3A%2F%2Fwww.nobelprize.org%2F">90 minute long session</a>, Alison Noble presented the <a href="https://royalsociety.org/news-resources/projects/science-in-the-age-of-ai/">Royal Society’s findings on how researchers are using AI</a>, Demis Hassabis explained his vision for building the ultimate meta-tool for science, and Paul Nurse shared his enthusiasm for potentially finally understanding the cell (at some abstraction layer).</p><p>Here are my 7takeaways from the conversation.</p><h3>Lesson #1: Let’s use AI for true Team Science</h3><p>We need interdisciplinary teams, but collaboration between different domains is often suboptimal since they tend to speak different languages.</p><p>Similar to how AI-enabled translation software can translate between Hindi and French, AI can now facilitate true cross-cutting work by enabling translation between entire fields.</p><ul><li><em>For technical people,</em> AI makes it easier to quickly learn new problem domains and application areas.</li><li><em>For domain experts</em>, AI makes it easier to understand, apply, and adapt technical tools to their field and downstream tasks.</li></ul><p>These tools are becoming accurate enough to provide real value in science. This makes it crucial to practice collaborating well with a wide range of people.</p><h3>Lesson #2: ”The Cat is out of the Hat”: It’s all hands on deck</h3><p>Not long ago, there was a significant movement to stop the development of AI.</p><p>Today, it’s moving faster than ever. Hassabis argues that the alarm bell has sounded. Society has been warned: AI is coming, and it will continually become more capable.</p><p>For a long time, AI research was dismissed as hype (remember, the field is over 70 years old!), but even previous critics are now embracing the technology.</p><p>The AI train will continue to boom. This means that, rather than debating whether it will come, we should assume that it will come and then think about the society we want and need post-AGI.</p><p>This is a responsibility for all corners of society, including economists (e.g., shifts in the workforce, universal income), artists, social scientists, and philosophers (e.g., purpose, the nature of understanding).</p><h3>Lesson #3: Match the problem with the appropriate method</h3><p>AI is one method among many (along with pure statistics and qualitative surveys, for example) for solving certain types of problems.</p><p>Demis Hassabis identified three main requirements for problems that are appropriate for AI:</p><ol><li>A huge combinatorial space (e.g., the countless possible board configurations of Go or the billions of possible proteins).</li><li>A good objective function (there must be something to optimize, such as a win condition or lower free energy).</li><li>Data: Either experimental or synthetic, such as generating data with a less capable model and improving iteratively.</li></ol><p>Many problems don’t require AI in the first place. On the other hand, many problems in fields such as medicine are very well suited for AI methods.</p><h3>Lesson #4: Engineering is ahead of science in computer science</h3><p>The science of AI lags behind model engineering.</p><p>For instance, we need to spend time understanding these tools scientifically: delving into model analysis, understanding them, examining the black boxes, and monitoring these systems.</p><p>Topics such as interpretability and explainability present good opportunities for academia to contribute without requiring the substantial computing budgets of companies and major research labs. Academia may also be better suited for this evaluative work, as it would be akin to AI companies “marking their own homework” (Demis Hassabis).</p><p>Furthermore, open-source models are often only six to twelve months behind frontier models, enabling even academia to produce relevant insights.</p><h3>Lesson #5: There are still many worries to tackle</h3><p>Demis lists these in roughly the order of their probable immediate impact.</p><ol><li>Bad actors using technologies for harmful purposes, such as disinformation and biological weapons.</li><li>Technical AGI risks (not having enough guardrails or benevolent intelligence, so the technology takes off without considering humans).</li><li>The economics of AI: equitably sharing the potential benefits of AI and not contributing to the widening wealth gap.</li><li>Philosophical: Meaning and purpose in the age of AGI.</li></ol><h3>Lesson #6: New tech tools must either be simple to use or automated</h3><p>For your AI for science tools to make a difference, scientists must be able to use them.</p><p>This means you can’t build tools that require users to spend most of their time debugging. The guideline should therefore be to either 1) distribute simple tools that are easy to use and debug, or 2) if the tools are complex, automate them so that users do not need to understand how the system works. One way to automate the complexities of large-scale model hosting and inference is to provide an API and do the computations for users.</p><p>In AI for science, collaborating with potential users is especially crucial.</p><h3>Lesson #7: Develop deep creative thinking</h3><p>Hassabis argues that the bottleneck is not technology; rather, it is creative thinking.</p><p>The hope is that AI tools will free up time for us to engage in more creative and in-depth thinking instead of rote work. Deep thought, deep work, and deep care for your research are all part of being a good scientist. Nurse echoed this point by emphasizing the importance of quality over quantity. According to Nurse, many papers today are just “fancy papers” that apply AI but lack purpose or interesting, useful results.</p><p>Hopefully, AI will free us from the endless stream of information so we can have the space to be creative.</p><h3>Conclusion</h3><p>Science is perhaps the most exciting application of AI, and it is a field that is genuinely being revolutionized by the technology.</p><p>Hopefully, the hype surrounding an AI utopia or dystopia can be regulated by adopting a researcher’s mindset. This mindset involves maintaining an insatiable curiosity about new technology and fields while remaining humble about our current understanding of the universe.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=412c6c75e424" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Book Notes: The Folklore of Trees by Aidan Meighan]]></title>
            <link>https://medium.com/@aliceheimanxyz/book-notes-the-folklore-of-trees-by-aidan-meighan-048579fdf0b0?source=rss-838957cd1468------2</link>
            <guid isPermaLink="false">https://medium.com/p/048579fdf0b0</guid>
            <category><![CDATA[books]]></category>
            <category><![CDATA[folklore]]></category>
            <category><![CDATA[nature]]></category>
            <dc:creator><![CDATA[Alice Heiman]]></dc:creator>
            <pubDate>Tue, 02 Jun 2026 19:42:01 GMT</pubDate>
            <atom:updated>2026-06-02T19:42:01.218Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/640/0*ympyhYw4Veri-EhC" /></figure><p>Most people have some kind of connection to trees, whether it’s the backyard tree that provides your family with fresh fruit, your favorite childhood forest, or the tree under which you buried a loved one.</p><p>Perhaps the connection to and mystery surrounding trees was stronger before, when most humans lived off the land and slept among them.<br>Trees, stories, forests, and culture are naturally intertwined. However, I didn’t realize just how deeply ingrained trees and their influence are across cultures.</p><p>This book describes 36 trees from different regions, sizes, and shapes, as well as their influence on folklore. Here are the three trees that stood out to me!</p><h4><strong>The Coconut — Cocos Nucifera</strong></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*YXJthClNTEdmYAxm" /><figcaption>Photo by <a href="https://unsplash.com/@zibik?utm_source=medium&amp;utm_medium=referral">zibik</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>The coconut is considered a divine gift because it provides both food and water. Its neat, portable packaging made long-distance seafaring possible. This also meant that coconuts could be traded globally. Coconuts appear in the folklore of regions close to water, such as the Philippines, Samoa, and Hawaii.</p><h4>The Walnut— Juglans Regia</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*3vfbNtdFkOLsLrU8" /><figcaption>Photo by <a href="https://unsplash.com/@jogara_0?utm_source=medium&amp;utm_medium=referral">JOGphotos</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>The walnut tree is an example of how trees and their folklore have influenced our language. In Italian, “noce” means “walnut,” while “nocere” means “to harm.” Why are these words so closely linked? In Benevento, Italy, there was supposedly a walnut tree around which all the witches in Europe would gather and dance. Despite attempts to remove the tree, the rituals continued, cementing Benevento’s reputation as a witch town and the connection between walnuts and evil. Who knew!</p><h4>The Coast Redwood — Sequoia Sempervirens</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*jEMTe9Kn1s4buaEx" /><figcaption>Photo by <a href="https://unsplash.com/@ryanwanx?utm_source=medium&amp;utm_medium=referral">Ryan Wan</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>The Redwoods are known as California’s giants. They are the tallest (growing upwards 100m) and heaviest (each weighing more than ten blue whales) trees on earth and can live as long as 2,000 years. As such, they are often described as steady, kind, and full of wisdom. One of my favorite tales from the book originated from the Yurok- a Californian Indigenous people. In it, an injured bird seeks shelter in the forest, and after many rejections from other trees finds a safe harbor among the redwoods.</p><h4>Conclusion</h4><p>I mostly enjoy reading up on the science behind nature, but this book was a great window into how nature affects our culture and language. The book also introduced me to many trees I didn’t know a lot about before!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=048579fdf0b0" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[5 Lessons from the 2026 ESIIL Innovation Summit]]></title>
            <link>https://medium.com/@aliceheimanxyz/5-lessons-from-the-2026-esiil-innovation-summit-fff1aae5865e?source=rss-838957cd1468------2</link>
            <guid isPermaLink="false">https://medium.com/p/fff1aae5865e</guid>
            <category><![CDATA[geography]]></category>
            <category><![CDATA[climate-change]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[gis]]></category>
            <category><![CDATA[sustainability]]></category>
            <dc:creator><![CDATA[Alice Heiman]]></dc:creator>
            <pubDate>Tue, 02 Jun 2026 18:55:55 GMT</pubDate>
            <atom:updated>2026-06-02T18:55:55.107Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*qZ6tbaN0ukWOqDzi" /><figcaption>Photo by <a href="https://unsplash.com/@intricateexplorer?utm_source=medium&amp;utm_medium=referral">Intricate Explorer</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>Boulder, Colorado, is home to some of the country’s leading geography and earth science research laboratories and institutions.</p><p>In addition to the interdisciplinary teams at the University of Colorado Boulder, it is also home to the headquarters of <a href="https://esiil.org/">ESIIL</a>, an NSF institution that provides the computing infrastructure, resources, and training necessary for large-scale environmental data science. I had the great opportunity to take part in this year’s summit, titled “<a href="https://cu-esiil.github.io/Innovation-Summit-2026/">AI for Sustainability: Translating Environmental Data into Decisions</a>.” The event brought together the environmental and computer science worlds, and it was incredibly insightful.</p><p>Below are five key takeaways from the event at the intersection of computer science and environmental decision-making.</p><h3>Lesson #1: The translation problem is the problem</h3><p>We currently have many technologies that perform well on specific benchmarks.</p><p>But what ultimately matters is how these technologies actually impact people’s lives. In environmental science, this means making better decisions about how to perform the energy transition within a municipality, managing a protected forest more effectively, and deciding which marine protected areas would provide the greatest benefits. I would summarize the <em>translation problem</em> as converting raw machine learning model outputs into <strong>relevant</strong> information that downstream stakeholders can <strong>act upon</strong>.</p><p>Although we have a lot of data, and AI promises to “find patterns that humans cannot see,” we are still figuring out how to translate data into patterns and further into actionable information.</p><h3>Lesson #2: There are many interdisciplinary themes to explore</h3><p>One thing I learned from the summit is that there are many topics at the <strong>intersection of AI and the environment that we need to work on</strong>.</p><p>Here are some of my personal highlights from the <a href="https://cu-esiil.github.io/Innovation-Summit-2026/science_jam/#leadership-program-invitees">ESIIL Innovation Summit 2026 Science Jam</a>:</p><p><strong>Ecosystem Dynamics &amp; Conservation</strong></p><ul><li>How can we unify genetic, species, and ecosystem-level biodiversity data to detect ecosystem thresholds and inform decision-making?</li></ul><p><strong>Food Systems &amp; Agriculture</strong></p><ul><li>How can AI enable real-time monitoring of crop health and food security risks using remote sensing and field data?</li></ul><p><strong>Energy Systems &amp; Climate Mitigation</strong></p><ul><li>How might AI methods be used to evaluate and optimize tradeoffs between energy infrastructure development and environmental impact?</li></ul><p><strong>Earth Observations &amp; Digital Twins</strong></p><ul><li>What are the data opportunities to advance Earth embeddings and improve predictions and decisions across regions and over time?</li></ul><p><strong>Community &amp; Knowledge Co-Production</strong></p><ul><li>How can AI systems be co-designed with stakeholders to support place-based, culturally relevant environmental solutions?</li></ul><p><strong>AI Education for Sustainability</strong></p><ul><li>How can we prepare the students of today to use emerging AI technology and how can AI enhance science communication?</li></ul><p>One of the reasons I enjoy these questions is because they require interdisciplinary teams.</p><h3>Lesson #3: AI usage is fraught with tensions</h3><p>Like most technologies, AI is not black or white.</p><p>Most people must face the many tensions that arise when using AI. For example,</p><ul><li>the tension between <em>wanting</em> to use AI and <em>having</em> to use AI (e.g., for work requirements, to get investments, or to safeguard a future career).</li><li>the tension between <em>using tools</em> that allow for faster work for the environment vs. <em>their impact</em> on the environment</li><li>the tension between reaching more people’s opinions with AI vs. taking the time to meet with people one-on-one.</li><li>etc.</li></ul><p>In many cases, there are no clear right or wrong answers, and people will fall somewhere on a continuum regarding each of these tensions. What I find important is our ability to be <em>intentional</em> with our usage, recognize the tensions, and weigh them against each other to determine what is most important to us, our work, and the people around us.</p><h3>Lesson #4: Prediction and causal inference are not the same</h3><p>Prediction and causal inference are fundamentally different tasks.</p><p>While prediction attempts to optimize the accuracy of the <em>output</em>, causal inference attempts to optimize the understanding of the <em>inputs</em>. In many environmental decision-making contexts, causal inference is more valuable because it reveals which factors and scenarios produce the desired outcome and helps inform necessary decisions. For example, does having a protected forest in the area actually decrease local deforestation rates? Machine learning is a different paradigm that does not directly answer these causal questions.</p><p>Therefore, it is important to understand ML’s capabilities and limitations. ML can be used in the causal inference pipeline (e.g., in preparing data layers), but not as a replacement for the analysis itself.</p><h3>Lesson #5: Don’t trade a deep life for a convenient one</h3><p>AI can definitely make things faster, from retrieving information to writing content.</p><p>However, as we now know, faster does not necessarily mean better. The final keynote speaker made it clear that personal relationships matter most for a fulfilling and meaningful life. Yes, you can use AI tools, but talk to people first. Learn about your community from the people in it first, and then use web sources such as Wikipedia and LLMs.</p><p>Also, cultivate the habit of giving back.</p><p>After using extractive tools, how can you give back some of what was taken? This could be done by being present with a friend, volunteering, planting trees, donating to charities, or fundraising for community issues.</p><p>In essence, don’t trade the deep life for the convenient one.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fff1aae5865e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[4 Things I Learned from a Brazillian Reforestation NGO: A Peek into the Restoration of the…]]></title>
            <link>https://medium.com/@aliceheimanxyz/4-things-i-learned-from-a-brazillian-reforestation-ngo-a-peek-into-the-restoration-of-the-2938f916d8a8?source=rss-838957cd1468------2</link>
            <guid isPermaLink="false">https://medium.com/p/2938f916d8a8</guid>
            <category><![CDATA[gis]]></category>
            <category><![CDATA[travel]]></category>
            <category><![CDATA[conservation]]></category>
            <category><![CDATA[brazil]]></category>
            <dc:creator><![CDATA[Alice Heiman]]></dc:creator>
            <pubDate>Tue, 26 May 2026 14:18:43 GMT</pubDate>
            <atom:updated>2026-05-26T14:18:43.878Z</atom:updated>
            <content:encoded><![CDATA[<h3>4 Things I Learned from a Brazillian Reforestation NGO: A Peek into the Restoration of the Brazillian Atlantic Forest</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/900/0*cO2S8LZ-f12V6uXT" /><figcaption>Credit: SOS Mata Atlântica, <a href="https://www.youtube.com/user/SOSMata">YouTube</a></figcaption></figure><p>Many people think of the Amazon rainforest when they hear about Brazil.</p><p>But in fact, <a href="https://www.wwf.org.br/en/our_work/in_the_atlantic_forest/">70% of Brazilians</a> live in the Atlantic Forest biome, not the Amazon. Stretching along the east coast of Brazil, the Atlantic Forest is home to a staggering array of biodiversity. It is also crucial for supplying ecosystem services, such as clean air and water, to the surrounding populations.</p><p>However, only <a href="https://www.wwf.org.br/en/our_work/in_the_atlantic_forest/">10–12% of the forest remains</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4zrYSCRBsrd2-TZi_jtb4w.png" /><figcaption>NASA and Miguelrangeljr, Public domain, via Wikimedia Commons</figcaption></figure><p>The conservation and restoration of the Atlantic Forest, through the <a href="https://www.decadeonrestoration.org/trinational-atlantic-forest-pact">Trinational Atlantic Forest Pact</a>, is designated a United Nations Flagship Project. One of the participating non-profit organizations is <a href="https://www.sosma.org.br/">SOS Mata Atlântica</a> (<em>eng</em>. SOS Atlantic Forest). SOS MA is one of the largest domestic reforestation organizations, and has planted over 43 million native trees.</p><p>As part of Stanford University’s Natural Capital Project, my research team and I helped SOS Mata Atlântica model the impact of various restoration scenarios on water quality to help them prioritize their efforts.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*9r7b6boqoZssI7ZQk-fU6A.png" /><figcaption>Some of the input data layers prepared for the analysis.</figcaption></figure><p>I recently had the opportunity to visit them in person at their headquarters in Itu, São Paulo. Here are some of the things I learnt.</p><h3>Learning #1: With the right knowledge, trees can regrow very quickly.</h3><p>Just 20 years ago, the land surrounding the SOS MA headquarters was essentially pasture land.</p><p>Now, 20 years later, there is a dense canopy of native Atlantic forest trees. This type of active restoration allows for a more strategic and specific approach (i.e. focusing on connecting forest fragments with biological corridors) and enables faster growth. However, these forests may not include the full biodiversity seen in primary forests, as there are limits to the number of different species that can be cultivated and planted.</p><p>So the best time to plant a tree is truly today.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*-VDlYLSekNNbnnRaT7dH-w.png" /></figure><h3>Learning #2: Growing trees for restoration purposes is very different from growing trees for other uses.</h3><p>One reason the trees grow so quickly is that they are prepared in a very specific way in an area known as a nursery.</p><p>The seeds are first activated to skip their “sleep phase”, then cultivated in soil, then manually transferred to small plant pots in a greenhouse (a warm environment), then into a shade house (to simulate the shade that young saplings would get in a real forest), and finally into full sunlight. The most important thing is to encourage solid root growth.</p><p>One of my main takeaways is that the seedlings are kept quite small before being planted, as this makes them easier to stack and transport to remote areas.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*Y1GZkAvy4kEELntVpDo3jw.png" /><figcaption>A seedling with a nicely developed root system.</figcaption></figure><h3>Learning #3: Education is crucial</h3><p>You not only have to do the work, but also ensure that others understand why your work is important.</p><p>Therefore, outreach and education are core pillars of the operation. I was impressed by every team member’s sense of purpose and understanding of how their role contributes to SOS MA’s mission. This is also why NGOs often need large media and communications teams. NGOs are short on staff, money, and time. In order to keep operations going, they often rely on external funding.</p><p>It is crucial that the NGO demonstrate why funding is needed, how it leads to impact, and why someone should fund future projects.</p><h3>Learning #4: People are everything</h3><p>At the end of the day, people are the most important thing.</p><p>Volunteers routinely measure water quality in rivers, team members work extra hours when the truck suddenly breaks down, and artists create music and art inspired by their deep connection to the land. Data is important, but the most important nudge is often not statistics, but stories. Building relationships builds trust and a shared sense of mission.</p><p>After all, behind every brand and partnership are real people who can feel, hope, and take action.</p><h3>Conclusion</h3><p>The Atlantic Forest is a fascinating place, but it’s far from something we can take for granted.</p><p>I’m inspired by the people in organizations like SOS Mata Atlântica who work tirelessly every day to restore the forest, one tree, one relationship, and one message at a time.</p><p>Thank you to the entire team for hosting me! I had a wonderful time.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*GYONJjkpQCmu_LbIm69bOg.png" /></figure><p><em>If you are interested in learning more about the Brazilian Atlantic Forest and how technology is being used to conserve and restore it, please check out my five-part article series, The Invisible Forest, here: </em><a href="https://the-invisible-forest.vercel.app/.*"><em>https://the-invisible-forest.vercel.app/.</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2938f916d8a8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The ICLR 2026 Keynote That Changed How I Think About AI Development]]></title>
            <link>https://medium.com/@aliceheimanxyz/the-iclr-2026-keynote-that-changed-how-i-think-about-ai-development-430b613a0851?source=rss-838957cd1468------2</link>
            <guid isPermaLink="false">https://medium.com/p/430b613a0851</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[science]]></category>
            <category><![CDATA[scientific-method]]></category>
            <category><![CDATA[learning]]></category>
            <dc:creator><![CDATA[Alice Heiman]]></dc:creator>
            <pubDate>Tue, 19 May 2026 18:01:00 GMT</pubDate>
            <atom:updated>2026-05-19T18:01:00.605Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*sQquwckZzq_RyXz8" /><figcaption>Photo by <a href="https://unsplash.com/@bkaraivanov?utm_source=medium&amp;utm_medium=referral">Bozhin Karaivanov</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>In 2012, the two AI researchers Yann LeCun and Yoshua Bengio founded the International Conference on Learning Representations (ICLR).</p><p>ICLR has since grown exponentially to become one of the largest and most influential AI conferences. In 2026, the conference received 19,525 paper submissions compared to 3,014 submissions just five years prior. Many seminal works that significantly enabled better machine learning algorithms were first published at ICLR, such as:</p><ul><li>2013: Efficient Estimation of Word Representations in Vector Space (Word2Vec) — <a href="https://arxiv.org/abs/1301.3781">https://arxiv.org/abs/1301.3781</a></li><li>2017: Adam: A Method for Stochastic Optimization — <a href="https://arxiv.org/abs/1412.6980">https://arxiv.org/abs/1412.6980</a></li><li>2021: An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale (Vision Transformer) — <a href="https://arxiv.org/abs/2010.11929">https://arxiv.org/abs/2010.11929</a></li><li>2021: Low-Rank Adaptation of Large Language Models (LoRA) — <a href="https://arxiv.org/abs/2106.09685">https://arxiv.org/abs/2106.09685</a></li></ul><p>The conference is also one of the most global in its rotation, including destinations such as Brazil, Singapore, Austria, Canada, and Rwanda.</p><p>At the conference itself, it is a packed schedule with thousands of participants. Keynote speakers, poster sessions, live presentations, socials, workshops, and a large hall with industry sponsors and events. Given its scale and influence, I was a bit unsure about the overall focus: would it be the industry focus of closed, concentrated development of larger black-box models, or more towards an open science community focus with interpretable models?</p><p>I was positively surprised, and the keynote that struck me most was from Percy Liang.</p><h3>From Closed-Source to Open Development</h3><p><a href="https://cs.stanford.edu/~pliang/">Percy Liang</a> is a Professor of Computer Science at Stanford University with a focus on natural language processing, and he held a unique <a href="https://blog.iclr.cc/2026/04/17/announcing-the-iclr-2026-keynotes/">keynote</a> at ICLR 2026.</p><p>Instead of simply presenting the model results, he took us on the <em>journey</em> of actually training the model itself (as part of project Marin). As an academic lab, they only had one model run with limited resources to make the most of. Professor Liang went over how they, in real-time, had to debug the model training (that of course did not converge as expected) and put out any fires (such as numerical issues) on the fly. They ended up with a model that was really good given their resource constraints.</p><p>Yet, Professor Liang was deeply unsatisfied.</p><p>Constantly having to twisting the knobs for the right setting for each model run felt unsatisfying. First, it is a waste of resources since you have to do many computationally expensive parameter searches. Second, many of these runs and experiments are hidden and therefore not reproducible or able to be scrutinized by the broader community. And perhaps the biggest hit to our ego, it means we have very limited understanding of how these systems actually behave.</p><p>That is why Professor Percy Liang heavily advocated for Open Development (with the ability to predict model experiment outcomes in advance) as the grand challenge of ICLR 2026.</p><h3>What is Project Marin?</h3><p>Marin is the open lab Liang and his team at Stanford built to put open development into practice. According to the official Marin announcement, it is an open community lab in which the research and development of models is completely transparent from the start. The team leverages GitHub to organize code, experiments, issues, hypotheses, and other project efforts. (Hall, et al., 2025)</p><p>Concretely, each experiment is tracked by a GitHub issue, which “serve as form of mini-<a href="https://en.wikipedia.org/wiki/Preregistration_(science)">preregistration</a>, which declares the hypotheses and goals upfront.” Anyone can submit a pull request to run an experiment, anyone can review it, and once approved, the experiment is launched and anyone can watch it execute live. (Hall, et al., 2025)</p><h3>Open Source vs. Open Development</h3><p>So, how is this different from traditional open source development?</p><p>In the field of machine learning, these terms differ:</p><ul><li><strong>Open-weight models</strong> (e.g., Llama, DeepSeek, Gemma): the trained model weights are released, but the code and data used to produce the model are not released. These are <strong>not</strong> open-source models.</li><li><strong>Open-source models</strong> (e.g., EleutherAI, AI2, LLM360): teams release not only the model weights, but also the code and data used to produce the model, so others can build their own models on top.</li><li><strong>Open development</strong> (Marin): goes a step further. The research and development of models is completely transparent. Experiments are preregistered as GitHub issues, declared in code, reviewed in public, executed live for anyone to watch, and negative results are kept.</li></ul><p>As stated by the Marin team: “This is a new way to do research that’s more scientific and more inclusive. We normalize making mistakes. Any negative results are in the open. There is no cherry picking. There is no ambiguity about what was run.” (Hall, et al., 2025)</p><h3>Why this changed my perspective</h3><p>I was encouraged by this keynote, since it aimed to bring more of a scientific workflow into AI development. Just like a model for physics allows us to take input variables, such as the initial velocity and the mass, to predict the trajectory of a ball, we should develop a similar model of understanding of AI models. Stating a hypothesis <em>prior</em> to the experiment is crucially important to display you understand what you are doing.</p><p>And it also showed that by aiming to understand the underlying principles of AI systems, you can achieve impressive results with far fewer resources.</p><h3>References</h3><ul><li>ICLR 2026 Media Fact Sheet. Available at <a href="https://media.iclr.cc/Conferences/ICLR2026/ICLR2026_Fact_Sheet.pdf">https://media.iclr.cc/Conferences/ICLR2026/ICLR2026_Fact_Sheet.pdf</a></li><li>International Conference on Learning Representations. <a href="http://wikipedia.org/">Wikipedia.org</a>. Accessed 2026–05–19 from <a href="https://en.wikipedia.org/wiki/International_Conference_on_Learning_Representations">https://en.wikipedia.org/wiki/International_Conference_on_Learning_Representations</a></li><li>“Announcing the ICLR 2026 keynotes”, <a href="http://blog.iclr.cc/">blog.iclr.cc</a>. Accessed 2026–05–19 from <a href="https://blog.iclr.cc/2026/04/17/announcing-the-iclr-2026-keynotes/">https://blog.iclr.cc/2026/04/17/announcing-the-iclr-2026-keynotes/</a></li><li>Saskia Hoving. “Why are proceedings so important in computer science?”. <em>Springer Nature</em>. Accessed 2026–05–19 from <a href="https://www.springernature.com/gp/librarians/the-link/ebooks-blogpost/why-are-proceedings-so-important-in-computer-science/17801220">https://www.springernature.com/gp/librarians/the-link/ebooks-blogpost/why-are-proceedings-so-important-in-computer-science/17801220</a></li><li>David Hall, et al., “Introducing Marin: An Open Lab for Building Foundation Models”. Marin Community Blog. Accessed 2026–05–19 from <a href="https://marin.community/blog/2025/05/19/announcement/">https://marin.community/blog/2025/05/19/announcement/</a></li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=430b613a0851" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[3 Points We Need to Address to Advance AI for Sustainability]]></title>
            <link>https://medium.com/@aliceheimanxyz/3-points-we-need-to-address-to-advance-in-ai-for-sustainability-730dde12c534?source=rss-838957cd1468------2</link>
            <guid isPermaLink="false">https://medium.com/p/730dde12c534</guid>
            <category><![CDATA[decision-making]]></category>
            <category><![CDATA[sustainability]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[environment]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Alice Heiman]]></dc:creator>
            <pubDate>Wed, 13 May 2026 03:37:30 GMT</pubDate>
            <atom:updated>2026-05-13T03:37:51.240Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*rhheCqhTJoe1F1jB" /><figcaption>Photo by <a href="https://unsplash.com/@gabigi?utm_source=medium&amp;utm_medium=referral">Gabriela</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>AI for sustainability is a growing field that evokes both enthusiasm and skepticism.</p><p>AI may enable us to understand the petabytes of Earth observation data collected daily, manage large-scale energy grids, and accelerate the discovery of carbon-neutral materials. However, most AI models have not been trained to handle the complexities and uncertainties of the natural world (unlike the structured realm of code, math, and language). Even solutions to so-called “solved problems” in the AI community often fail when transitioned from the clean, abstract digital world to the messy real physical world.</p><p>This raises the question: What topics do we need to address to make AI useful in the sustainability domain?</p><h3>Point #1: Ensure Sustainability Professionals Without Technical Expertise Can Validate Model Outputs</h3><p>Companies often market coding tools as a way to democratize technical knowledge and implementation.</p><p>However, while these tools are great for demonstrations, a developer is still needed to validate the outputs and bring the proof of concept into production. Therefore, environmental NGOs with tight budgets still have to contract expensive consultants to build their software solutions. How can someone without technical expertise feel confident that the generated code is secure and follows best practices?</p><p>To truly empower environmental practitioners with AI capabilities, we must ensure they can correctly and efficiently validate and interact with the technology.</p><h3>Point #2: Develop AI Methods to Harmonize Data Across Scales and Methods</h3><p>The AI community always demands more data, but we should also be smarter about the data we already have.</p><p>Unlike the clean, labeled, and structured datasets often used to train machine learning models, environmental data is messy. This data includes field studies in Brazil, where water samples were taken from polluted rivers in glass bottles; extensive natural history collections; and near real-time global satellite images. How do we use these datasets to inform each other? This is especially difficult when data is collected at vastly different temporal and spatial scales using different methods and sensors across diverse ecosystems.</p><p>The data is there, but we need better methods to efficiently standardize and utilize it. This will allow all sustainability work to build on each other without the need to redo work.</p><h3>Point #3: Ensure AI Supports A Clear Data-to-Decision Pipeline</h3><p>AI can make many predictions, but not all are useful for environmental action.</p><p>When it comes to AI development, most prestige is centered on the technical implementation. What novel model architecture did you invent? What new loss function did you optimize? How many flops were you able to squeeze out of the GPUs for training? However, none of this matters if your model optimizes the wrong outcome, or if decision-makers don’t understand the outputs.</p><p>We need to transition from AI-<em>first</em> models to AI-<em>supported</em> decision-making systems.</p><p>This means intentionally and carefully considering the entire decision pipeline, in which the technical aspect is only a small part of the puzzle. Who are the stakeholders? Who has the power to make decisions? What information is needed, and how much, to make the decision or policy change? Where can AI be intentionally deployed to support this process, and where should it not be used</p><p>Instead of building AI models that predict various things we hope could be useful, we need tighter integration and a feedback loop between developers, stakeholders, decision-makers, and people executing actions on the ground.</p><h3>Conclusion</h3><p>In short, AI <em>does</em> hold a lot of potential in the sustainability domain. However, it is time to move from potential to practice.</p><p>This requires ensuring that AI tools can be used and trusted by practitioners with real domain knowledge. It means developing AI methods specifically suited to the environmental domain, where data is extremely heterogeneous.</p><p>It also means understanding that AI is not <em>the</em> solution, only part of it.</p><h3>Credits</h3><p>These thoughts were developed from fruitful conversations at the machine learning conference ICLR2026, the Brazillian reforestation NGO SOS Mata Atlântica, and with participants at the 2026 ESIIL Innovation Summit.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=730dde12c534" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[3 Steps to Use Alpha Earth Embeddings for LULC Change Detection]]></title>
            <link>https://medium.com/@aliceheimanxyz/3-steps-to-use-alpha-earth-embeddings-for-lulc-change-detection-151483cc7bc4?source=rss-838957cd1468------2</link>
            <guid isPermaLink="false">https://medium.com/p/151483cc7bc4</guid>
            <category><![CDATA[remote-sensing]]></category>
            <category><![CDATA[google]]></category>
            <category><![CDATA[gis]]></category>
            <category><![CDATA[forestry]]></category>
            <category><![CDATA[change-detection]]></category>
            <dc:creator><![CDATA[Alice Heiman]]></dc:creator>
            <pubDate>Fri, 17 Apr 2026 13:35:25 GMT</pubDate>
            <atom:updated>2026-04-17T13:35:25.120Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*JyW7fN7Zv2V9m0uY" /><figcaption>Photo by <a href="https://unsplash.com/@rmvisuals?utm_source=medium&amp;utm_medium=referral">Renaldo Matamoro</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>One use case of Earth Foundation Model embeddings is change detection.</p><p>The Google DeepMind AlphaEarth Foundation model, for instance, projects every point on earth into an embedding space represented by a 64-dimensional embedding sphere. This sphere learns the meaning of each point, meaning that points closer together on this 64-dimensional sphere should be more similar than points further apart. This also means that embeddings from the same geographic location can change sphere location over time, for instance, as that location experiences land-use change. By comparing points at two points in time, we can identify areas that have remained relatively untouched versus those that have changed.</p><p>In this article, we’ll run a small experiment to identify larger regions that may have undergone drastic human-made changes in a region around Rio de Janeiro, Brazil.</p><h3>Step #0: Define region of interest</h3><p>Before we can begin, we need to define a region of interest.</p><p>Although you can do the following calculations on a global scale, it’s often faster and simpler to start on a smaller region. In Google Earth Engine, it is pretty straightforward since you can literally draw a polygon straight on the map. Alternatively, you can load in your own polygon or region of interest.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gia93Hq1jU5BwRxqsVf9eA.png" /></figure><h3>Step #1: Load Alpha Earth Embeddings and compute similary</h3><p>Next, we need to load the alpha-earth embeddings for our start and end years. In this example, we use the embeddings from 2017 and 2024.</p><pre>var col = ee.ImageCollection(&#39;GOOGLE/SATELLITE_EMBEDDING/V1/ANNUAL&#39;);<br><br>function getYearEmbedding(year) {<br>  return col<br>    .filter(ee.Filter.date(<br>      ee.Date.fromYMD(year, 1, 1),<br>      ee.Date.fromYMD(year, 12, 31)<br>    ))<br>    .filter(ee.Filter.bounds(aoi))<br>    .mosaic()<br>    .clip(aoi);<br>}<br>var emb2017 = getYearEmbedding(2017);<br>var emb2024 = getYearEmbedding(2024);</pre><p>We can then compute the cosine similarity between the two time points. Since the alpha embeddings are already unit vectors, we can compute cosine similarity using a dot product.</p><pre>// Compute Cosine Similarity<br>var cosineSimilarity = emb2017.multiply(emb2024).reduce(ee.Reducer.sum());<br><br>// Visualize on the map<br>var simVis = {<br>  min: -1, max: 1,<br>  palette: [&#39;#d73027&#39;,&#39;#f46d43&#39;,&#39;#fdae61&#39;,&#39;#fee08b&#39;,<br>            &#39;#ffffbf&#39;,&#39;#d9ef8b&#39;,&#39;#a6d96a&#39;,&#39;#66bd63&#39;,&#39;#1a9850&#39;],<br>};<br>Map.addLayer(cosineSimilarity, simVis, &#39;Cosine Similarity 2017→2024&#39;);</pre><h3>Step #2: Apply Thresholding</h3><p>Now you can manually inspect the map to find regions with larger changes.</p><p>In the above visualization palette, brighter regions indicate greater change. We can then click around to find a suitable threshold. In this example, I decided to set a threshold of 0.6. That means that any points with a cosine similarity of 0.6 or higher between their 2017 and 2024 images are masked out.</p><p>Put another way, we only keep pixels that are less than 0.6 similar.</p><pre>var changeMask = cosineSimilarity.lt(0.6); // Threshold of 0.6<br><br>Map.addLayer(<br>  changeMask.selfMask(),<br>  {palette: [&#39;orange&#39;]},<br>  &#39;Changed pixels (below threshold)&#39;<br>);</pre><p>We can see these visualizations below:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/794/1*T98cnfGGRYoNNzsY14tv1A.png" /></figure><h3>Step #3: Vectorize and filter</h3><p>Finally, we can connect the pixels into coherent areas to filter changes by area and structure.</p><p>The process of connecting the individual pixels in a group is called vectorization. After vectorization, we get <em>polygons</em> rather than <em>individual pixels</em>. This allows us to compute the polygon’s area and derive characteristics of its shape. In this example, I filtered for polygons that are greater than 5 hectares and that resemble a rectangle. To perform this filtering, we compute the polygon’s area and compare it to the bounding box area. A perfect rectangle would have a one-to-one ratio of these areas.</p><p>In practice, filtering for rectangle-ish areas usually indicates artificial transitions.</p><p>Applying this filter allowed me to quickly find case studies of rapid land-use changes between 2017 and 2024, most of which were indeed projects of new construction.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/853/1*nGk5bsOT9TiALQB6FwEPgw.png" /></figure><p>I also compared the identified regions with the land-use land-cover maps produced by MapBiomas to see if we picked up a similar trend. What I realized was that a lot of these urban changes were already classified as the “urban” land cover class throughout the entire period by MapBiomas. But there were also cases where this similarity algorithm identified recent deforestation (i.e., going from “forest” land cover class to “urban” land cover class).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/479/1*Lsj-JHdtiY0HVt0W-UEpRg.png" /></figure><h3>Conclusion</h3><p>In this article, we can spot changes in land use using cosine similarity. Vectorizing a thresholded map enables us to filter further and categorize the detected changes.</p><p>However, this type of change detection doesn’t always tell us what land use cover we change from or to.</p><p>Therefore, the algorithm proposed in this article would require additional information to, for instance, distinguish between deforestation and reforestation. Also, I noticed the similarity map showed large changes in areas of forest plantations. Thus, these regions would also need to be mapped to prevent them from being mistaken for deforestation.</p><p>Hopefully, this article helped you get up and running with change detection using Alpha Earth embeddings!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=151483cc7bc4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How do Alpha Earth Embeddings for stable and changing LULC vary over time?]]></title>
            <link>https://medium.com/@aliceheimanxyz/how-do-alpha-earth-embeddings-for-stable-and-changing-lulc-vary-over-time-9aedfd727f0c?source=rss-838957cd1468------2</link>
            <guid isPermaLink="false">https://medium.com/p/9aedfd727f0c</guid>
            <category><![CDATA[google-earth-engine]]></category>
            <category><![CDATA[earth-foundation-models]]></category>
            <category><![CDATA[gis]]></category>
            <category><![CDATA[embedding]]></category>
            <category><![CDATA[remote-sensing]]></category>
            <dc:creator><![CDATA[Alice Heiman]]></dc:creator>
            <pubDate>Thu, 16 Apr 2026 20:29:36 GMT</pubDate>
            <atom:updated>2026-04-16T20:29:36.904Z</atom:updated>
            <content:encoded><![CDATA[<p>Earth Foundation Models and their embeddings are exciting because they can integrate a wide range of data types into a single representation.</p><p>For instance, Google DeepMind’s AlphaEarth model produces 64-dimensional Satellite Embeddings that group points on Earth not just by location, but also by land cover type and seasonal characteristics. However, these embeddings are not immediately interpretable and thus require experimentation to understand how they behave.</p><p>In this article, I’ll explore the temporal dynamics of Alpha Earth Embeddings in the Atlantic Forest of Brazil: in short, how do these embeddings evolve for point locations with various characteristics?</p><h3>Step #1: Extracting the Points</h3><p>Using Google Earth Engine, I sampled lat-lon points with stable LULC (same class for 10+ years) and changing LULC (4 years: LULC A → LULC B) according to the MapBiomas annual land use land cover maps from Collection 10.</p><p>I used Forest, Agriculture, and Urban as the stable classes, and then focused on the Forest→Urban change in the visualizations below. I picked 100 random points for each category in the state of Sāo Paulo. Finally, I extracted the Alpha Earth embeddings for each point across the available years 2017 to 2024.</p><p>Finally, I ended up with a file for each point, for each year, with the Alpha Earth embedding bands A00→A63.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zXGh4fNWw7Uxj-uzHpIi6w.png" /></figure><h3>Step #2: Visualizing the Embedding Space</h3><p>We can then fit a UMAP to project the 64-dimensional spherical embedding space onto 2- and 3-dimensional spaces, respectively. This helps us visualize any groupings in the embedding space. The graph below is a snapshot of the embedding space in 2024.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LD2GbWOkS01_REX1xEStSQ.png" /></figure><p>As we can see, we do have clusters, but there is a fair bit of noise. The urban class is the least noisy group. Interestingly, agriculture sits between the forest and urban classes, which makes sense conceptually.</p><h3>Step #3: Tracking Embedding Change Over Time</h3><p>To visualize how the embeddings of a single point change over time, we can use 2017 as the “anchor” and measure how much any subsequent temporal embeddings for the same point compare spatially.</p><p>We can use angular drift, which is the angle between two unit vectors in embedding space:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zP7OqmFkz-IAP1NhKF8Dhw.png" /></figure><p>Where v₀ is the 2017 embedding and vₜ is the embedding at year t. Since all embeddings are L2-normalized to the unit hypersphere, the dot product directly equals the cosine similarity, and arccos converts that to a geometric angle in degrees. A drift of 0° means the embedding is identical to 2017; 90° means completely orthogonal (no shared structure); 180° means maximally opposite.</p><p>We first apply this angle metric to the stable classes:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*t5GpP7i8tn4PABxFHqGHrw.png" /></figure><p>For stable LULC, we see that the angular drift stays around 10–25 degrees and remains relatively flat over time. The AlphaEarth Foundation’s embedding space is supposed to be consistent over time. Therefore, any temporal drift should,, in theory,, be due to an actual changes in the point, such as forest growth, crop changes, or new buildings.</p><p>We can compare these to known changing LULC to see the relative difference in drift. For instance, below we plot points converting from forest to urban. MapBiomas classifies the points as forest years 2017, 2018, 2019, and 2020, and then as urban years 2021, 2022, 2023, and 2024.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*K9M8NfhZCkIKlve9-DxjMg.png" /></figure><p>Here, we can see a much clearer drift for the same point between 2017 and 2024. What is interesting is that we do see the steepest mean slope (largest angular velocity/change velocity) between 2020 and 2021, right when the conversion is supposed to have happened according to MapBiomas.</p><p>We can sample a point showing a significant change and track its trajectory in embedding space from 2017 to 2024:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ZF2SOsaKvvEEApBEvj_EDw.png" /></figure><p>Here, we can see how the point starts deep in the forest cluster (green) and moves toward and into the urban cluster (red). Thus, we can see how the embedding space reveals how points change across different land use and land cover classifications!</p><h3>Conclusion</h3><p>This article presents an initial exploration of using Earth Foundation model embeddings to analyze land-use and land-cover trajectories. Knowing the baseline temporal drift of reference points can help create thresholds for automatic change detection and even characterize the speed of land conversions.</p><p>In a future post, I’ll keep building on this concept to see if we can get meaningful metrics to quantify how well reforestation projects in the Atlantic Forest are progressing!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9aedfd727f0c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to download Alpha Earth Embeddings]]></title>
            <link>https://medium.com/@aliceheimanxyz/how-to-download-alpha-earth-embeddings-d2e8a1c2c058?source=rss-838957cd1468------2</link>
            <guid isPermaLink="false">https://medium.com/p/d2e8a1c2c058</guid>
            <dc:creator><![CDATA[Alice Heiman]]></dc:creator>
            <pubDate>Wed, 15 Apr 2026 21:14:24 GMT</pubDate>
            <atom:updated>2026-04-15T21:14:24.847Z</atom:updated>
            <content:encoded><![CDATA[<p>This scripts allows you to upload CSV with points and extract yearly Alpha Earth Embeddings: <a href="https://code.earthengine.google.com/afec2f1b1a7f8b372275dd41ab05c577">https://code.earthengine.google.com/afec2f1b1a7f8b372275dd41ab05c577</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d2e8a1c2c058" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to Sample Reference LULC Points from MapBiomas Brazil]]></title>
            <link>https://medium.com/@aliceheimanxyz/how-to-sample-reference-lulc-points-from-mapbiomas-brazil-7ef1b59a81ca?source=rss-838957cd1468------2</link>
            <guid isPermaLink="false">https://medium.com/p/7ef1b59a81ca</guid>
            <category><![CDATA[satellite-imagery]]></category>
            <category><![CDATA[mapbiomas]]></category>
            <category><![CDATA[statistics]]></category>
            <category><![CDATA[remote-sensing]]></category>
            <category><![CDATA[brazil]]></category>
            <dc:creator><![CDATA[Alice Heiman]]></dc:creator>
            <pubDate>Tue, 14 Apr 2026 20:46:58 GMT</pubDate>
            <atom:updated>2026-04-15T09:39:25.768Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*6lNeefJtZ-nn7VSD" /><figcaption>Photo by <a href="https://unsplash.com/@hdbernd?utm_source=medium&amp;utm_medium=referral">Bernd 📷 Dittrich</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>MapBiomas is a platform that creates and distributes land use maps of South America.</p><p>Their mapping efforts from 1985 and onward allow scientists, NGOs, policy makers, and hobbyists to analyze vegetation change. One useful thing to have is to sample points with a specific LULC code for a given year for later analysis. However, there can be a learning curve to create these custom scripts.</p><p>In this post, I’ll show you how to sample points for a specific LULC code using MapBiomas Collection 10 in Google Earth Engine.</p><p>→ You can find the full interactive script here: <a href="https://code.earthengine.google.com/c79ed7fa9c5317e600bde2ff6a111aee">https://code.earthengine.google.com/c79ed7fa9c5317e600bde2ff6a111aee</a></p><h3>Step #1: Load the Google Earth Engine Script</h3><p>I have prepared the Google Earth Engine script here: <a href="https://code.earthengine.google.com/c79ed7fa9c5317e600bde2ff6a111aee">https://code.earthengine.google.com/c79ed7fa9c5317e600bde2ff6a111aee</a></p><p>You will need to log into a Google account and can then run the code to save the CSV export of points (lat, lon, id) to your drive.</p><h3>Step #2: Configure the Script</h3><p>At the top, you can configure the script to detail what LULC code you want to sample points from, what year to sample from, and how many points.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/932/1*jkuChCmtj6lTNA3E-E6YuQ.png" /><figcaption>Example of Config from the script.</figcaption></figure><p>This is also where you can specify the state to use and what biome to use. The default is to sample from the state of Sao Paulo and in the Atlantic Forest. But you can change the ROI parameter to sample from any custom region you would like (for instance by drawing your own polygon).</p><h3>Step #3: Run the Script</h3><p>Finally, press run to visually see your sample visually.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-OptqrSNio6-4sN5gCpTFg.png" /><figcaption>A few sampled points in the specified region</figcaption></figure><p>This also creates a “Task” to save your points. To save the sampled points to a CSV in your Google Drive, you must go to the “Tasks” tab and hit “Run”.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/672/1*_oCizDe40w3O0u8HYZeJAQ.png" /><figcaption>The task shows up in the Tasks tab.</figcaption></figure><p>After the job finishes, you can click “Open in Drive” to see the export folder and download the CSV from your Google Drive.</p><h4>(Extra: Define LULC Trajectories)</h4><p>It is often necessary to define LULC trajectories over time.</p><p>For instance, we might want to sample pixels that have been forest <em>for the past 10 years</em>, or a pixel that <em>was </em>pasture but <em>became </em>forest and <em>stayed forest </em>for the next 4 years. We can apply all those conditions using masks.</p><p>I created a Google Earth Engine script for this use case available here: <a href="https://code.earthengine.google.com/7f83d8905ff71abde60854dae0c8c511">https://code.earthengine.google.com/7f83d8905ff71abde60854dae0c8c511</a></p><h3>Conclusion</h3><p>And that is it! Hopefully these scripts can help you sample points for analysis in MapBiomas!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7ef1b59a81ca" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>