My research in Karamoja, Uganda showed me up close how the absence of timely, reliable information leaves communities and decision-makers exposed when a season fails. That experience set the question behind most of my work since then: how do we give the people responsible for food security the evidence they need, in time to act, where ground data is scarce? Earth observation (EO) became my answer, which is why building operational, EO-based agricultural monitoring that African institutions can own and use through the Xylem Lab<\/a> and Institute I founded matters so much to me.<\/p>\n\n\n\n This cropland estimation study, led by my PhD Student Adebowale Adebayo, sits at the foundation of that arc, and it\u2019s very much a community effort. The Lab and Institute exist to give African researchers the platform, mentorship, and collaborations to lead science on the questions that matter most to the continent. My student ran much of the analysis, supported by a wider team of students and collaborators, and that model, with our people driving the work, is what I\u2019m proudest of. <\/p>\n\n\n\n Before you can forecast a harvest or warn of an incoming crisis, you have to answer a very basic question: how much cropland is there, and where? This underpins every production estimate, yet in many African countries, that figure is uncertain or out of date. Maps alone aren’t enough. A map can tell you where cropland is, but not how much there is with any confidence. Our work adds that confidence: by pairing map accuracy assessment with a stratified sampling design, it turns satellite maps into defensible area estimates \u2013 numbers that hold up next to official agricultural statistics across six countries. <\/p>\n\n\n\n For me, that loops back into Karamoja. The problem was always the same: decisions made without trustworthy information. This is the same problem, scaled up and made reliable, transparent, and locally usable evidence about the land that feeds people, so that food security decisions rest on something solid. And it\u2019s answered by a team and an institution built to keep doing exactly that, for years to come. <\/p>\n\n\n\n In plain terms, what problem does your study solve, and is counting pixels not enough?<\/strong><\/p>\n\n\n\n Governments need to know how much land is under cultivation, but ground surveys are slow and costly, so satellite imagery is increasingly used instead. The shortcut (pixel counting) is to add up every pixel that a classification model labeled cropland and call that the area. That treats the map as though it were error-free. It isn’t: every map both misclassifies real cropland (omission error) and labels other land cover types as cropland (commission error). The two only cancel out if they happen to be equal in size, which they almost never are, so what remains is a systematic bias, and a pixel count carries no uncertainty to warn you it is there. In our six countries, pixel counting deviated from the statistically defensible estimate by \u221217% to +43%. In Zambia, that is 9.0 Mha versus 6.3 Mha. Our framework returns a number with a defensible margin instead.<\/p>\n\n\n\n Your team tested seven land cover products across six African countries, and Digital Earth Africa was among the strongest performers. What made it stand out, and what has DE Africa’s role been in NASA Harvest’s wider work across the continent?<\/strong><\/p>\n\n\n\n DE Africa was the second-strongest product overall, with 15 top-two finishes across our metric\u2013country combinations led by Adebowale Adebayo, who spearheaded much of this rigorous analysis. What set it apart was how much real cropland it found: it had the highest average producer’s accuracy of anything we tested (0.73), meaning it missed the least actual cropland, ranking first in Rwanda, Malawi, and Zambia, and second in the other three countries. For a class as sparse as African cropland, misclassified cropland (omission error) is what drives survey costs up, so this is the metric that matters most. We put that down to two things: it is a dedicated cropland product at 10 m resolution, not a general land-cover map in which cropland is one class among ten, and its training data was collected by people who know these agricultural landscapes firsthand. Beyond this paper, DE Africa’s open, analysis-ready layers have become a default starting point for our work; they remove the step where every team has to build a continental baseline from scratch.<\/p>\n\n\n\n Can you walk us through how DE Africa’s data fit into your accuracy assessment and sampling design, from map to final area estimate?<\/strong><\/p>\n\n\n\n Three stages. First, we evaluated the DE Africa against reference points our team interpreted from high-resolution imagery. Those points were drawn from an independent stratification, so no product was scored on its own strata, which meant using estimators built for the case where the strata and the map classes differ. That yielded producers’ and users’ accuracy with standard errors. Second, we combined those accuracies with each country’s true cropland fraction to compute the sample size that each map would require to achieve a 10% coefficient of variation. DE Africa needed the fewest in Malawi and Zambia. Third, it became the stratification layer for a second, fully independent sample, from which we derived 3.63 \u00b1 0.29 Mha for Malawi and 6.31 \u00b1 0.93 Mha for Zambia.<\/p>\n\n\n One of your key findings is that no single global product is optimal, and that country-specific evaluation is essential. What does that mean in practice for a country trying to plan its next agricultural survey?<\/strong><\/p>\n\n\n\n It means don\u2019t select a map based on its published headline accuracy. Our best-performing map changed by country: GLAD in Kenya, Tanzania, and Uganda; DE Africa in Malawi and Zambia; an ensemble in Rwanda; and overall accuracy proved a poor guide: two products can score almost identically yet differ by more than 50% in the sample sizes they require. The practical move is to bring the evaluation in-house. A pilot of 50\u2013100 reference points, using a day of analyst time, is enough to estimate producers’ and users’ accuracy for two or three candidate products across a country’s territory, and to rank them on the only criterion that affects the budget. Where cropland is a minor class, producers’ accuracy dominates; where it dominates the landscape, users’ accuracy does.<\/p>\n\n\n\n Your framework cut required reference sample sizes by 20 to 40 percent. In real terms, what does that saving mean for a country with a limited survey budget?<\/strong><\/p>\n\n\n\n Take Kenya. Reaching a 10% coefficient of variation takes 919 reference points if you stratify on the least efficient map, and 616 if you stratify on the best. That is 303 fewer places an enumerator has to physically reach \u2014 roughly \u20ac7,600\u201310,600 at European field-survey rates, or 15 to 25 hours of analyst time if the points are interpreted from imagery instead. The design also names the specific locations, so a ministry knows exactly where to send its extension agents rather than spreading them across the country, and the estimate stays design-unbiased. Across our six countries, the reduction ran from 12% to 41%. Where no open product covers the season a country needs, the method works just as well on a map the country makes itself.<\/p>\n\n\n\n What is the biggest obstacle you still see to affordable, reliable cropland area estimation in data-scarce parts of Africa, and how could platforms like DE Africa help close that gap?<\/strong><\/p>\n\n\n\n Reference data, not maps. Maps are plentiful and free; high-quality labeled reference points are the binding constraint, and today almost every project collects its own, uses it once, and effectively discards it. The second obstacle is that we often cannot adjudicate disagreements; our Zambia estimate of 6.3 \u00b1 0.9 Mha sits well above official figures of 1.4\u20133.8 Mha, and because those figures carry no documented uncertainty, nobody can say which is closer to the truth. Platforms like DE Africa are well placed to address the first by hosting a shared, persistent, probability-based reference archive against which any product can be assessed, and by publishing per-country accuracy with every map release.<\/p>\n\n\n\n If you could ask Digital Earth Africa for one new feature or dataset to support this kind of work, what would it be?<\/strong><\/p>\n\n\n\n If I could ask for one thing, it would be a continental, openly shared reference data for cropland \u2013 a continuously updated pool of validated, labeled points that any team can contribute to and draw from. Reliable area estimates don\u2019t come from maps alone; they come from good reference data, which is always in short supply. Right now, everyone collects it in isolation, and then it\u2019s discarded. If Digital Earth Africa hosted it as a shared infrastructure with an in-built accuracy assessment tool, any country could produce defensible cropland statistics, not just well-resourced teams. It would make results comparable across borders. That\u2019s one missing piece that would move this work from one-off studies to something operational, long-term, consistent, and African-owned. <\/p>\n\n\n\n You lead a team that includes early-career researchers like Adebowale Adebayo, who ran much of this analysis. What would you say to a young African researcher who wants to use freely available satellite data to tackle food security questions in their own country?<\/strong><\/p>\n\n\n\n Three things. First of all, you already have what you need to start. The satellite data is free, the tools are open, and the questions in your own country are the ones you understand better than any outside expert. That combination is powerful enough, so don\u2019t wait for permission or a perfect setup to begin. <\/p>\n\n\n\n Secondly, learn the fundamentals properly. It\u2019s tempting to jump to a model because everyone is talking about it, but the value is in doing the unglamorous work well: understanding your data, being honest about your uncertainty, and validating your maps. My student Adebowale led much of this analysis, yet what made it strong wasn\u2019t just a fancy method; it was discipline.<\/p>\n\n\n\n And third, your local knowledge and life experience are assets, not afterthoughts. Knowing how farming actually works where you live makes remote sensing more meaningful. Collaborate with others, share what you make, and aim for work your own institutions can use. That\u2019s how this field becomes African-owned, one researcher at a time. <\/p>\n\n\n\n Read the study:<\/strong> Adebayo, A.D., Nakalembe, C., Skakun, S., Frimpong, D.B., Ginsburg, A., Gadhiya, Y.H., Adjei, M., Kerner, H. & Becker-Reshef, I. (2026). Optimizing satellite-based cropland area estimation through integrated map accuracy assessment and stratified sampling design across six African countries. International Journal of Applied Earth Observation and Geoinformation, 148, 105223. https:\/\/doi.org\/10.1016\/j.jag.2026.105223<\/a><\/em><\/p>\n\n\n\n <\/p>","protected":false},"excerpt":{"rendered":" Reliable cropland estimates are essential for agricultural planning and food security, yet many African countries lack up-to-date information on how much land is cultivated. In a study spanning six African countries, Catherine Nakalembe and her team assessed seven satellite-based cropland products, with Digital Earth Africa recording among the highest average producer’s accuracy. The findings highlight […]<\/p>\n","protected":false},"author":4,"featured_media":5801,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"footnotes":""},"categories":[14,10,12],"tags":[],"audience":[49],"class_list":["post-5800","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-change-makers","category-agriculture-food-security","category-use-cases","audience-academia"],"acf":[],"_links":{"self":[{"href":"https:\/\/digitalearthafrica.org\/fr\/wp-json\/wp\/v2\/posts\/5800","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/digitalearthafrica.org\/fr\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/digitalearthafrica.org\/fr\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/digitalearthafrica.org\/fr\/wp-json\/wp\/v2\/users\/4"}],"replies":[{"embeddable":true,"href":"https:\/\/digitalearthafrica.org\/fr\/wp-json\/wp\/v2\/comments?post=5800"}],"version-history":[{"count":4,"href":"https:\/\/digitalearthafrica.org\/fr\/wp-json\/wp\/v2\/posts\/5800\/revisions"}],"predecessor-version":[{"id":5814,"href":"https:\/\/digitalearthafrica.org\/fr\/wp-json\/wp\/v2\/posts\/5800\/revisions\/5814"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/digitalearthafrica.org\/fr\/wp-json\/wp\/v2\/media\/5801"}],"wp:attachment":[{"href":"https:\/\/digitalearthafrica.org\/fr\/wp-json\/wp\/v2\/media?parent=5800"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/digitalearthafrica.org\/fr\/wp-json\/wp\/v2\/categories?post=5800"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/digitalearthafrica.org\/fr\/wp-json\/wp\/v2\/tags?post=5800"},{"taxonomy":"audience","embeddable":true,"href":"https:\/\/digitalearthafrica.org\/fr\/wp-json\/wp\/v2\/audience?post=5800"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}![\"\"](\"https:\/\/digitalearthafrica.org\/wp-content\/uploads\/Study-Area.png\")
<\/figure>\n<\/div>\n\n\n