What is the value of having a goal if you can’t measure it? This axiom does not exactly relate to international efforts to increase crops yields for small-scale farmers in developing countries, but it is pretty close.

Without historical records on farm productivity levels, most efforts to measure harvests in the developing world require farmers to either engage in often-inaccurate self-reporting exercises or undertake costly and time-consuming “crop-cuts” in which samples are taken from specific points in a field and then estimated out for the entire harvest. Due in part to the expense and time required for these methods, the outcomes of many agricultural interventions are simply never measured.

This dynamic is what makes leveraging the explosion in high-resolution satellite imagery to remotely sense yields in developing countries so exciting. Marshall Burke and David Lobell, professors at Stanford University and faculty affiliates of the Center for Effective Global Action (CEGA) at University of California, Berkeley, have recently demonstrated that images of maize farms across Western Kenya, captured by Planet’s Terra Bella satellite constellation, can be used to make yield estimates approximately as accurate as traditional field survey-based measures. The implications of this study are pretty intoxicating; in the foreseeable future, estimates of yield gains that previously took months of time on the part of large field teams could be as simple as a few clicks of the mouse.

Unfortunately, before this dream is realized, there is a pretty significant catch. What makes remote sensing for yield measures possible in the developing world are dramatic technological advancements occurring in the private sector. Planet and other cutting-edge “micro” or “nano” satellite companies now offer one-meter resolution images capable of capturing the tiny farms that characterize agricultural systems in sub-Saharan Africa or South Asia. However, pictures are not enough. Remote sensing for agriculture in the developing world is fundamentally constrained by the lack of ground-based data on yields, plot boundaries and other geo-referenced information. Investment in the curation and collection of new and existing sources of ground data is critical for this method to realize its potential.

Lobell's and Burke’s work in Western Kenya was enabled in part by a long-term study they undertook in the region as part of the Agricultural Technology Adoption Initiative (ATAI), generously supported by the the Department for International Development and the Bill & Melinda Gates Foundation. This multi-year project captured sufficient ground-based data, which was then used to train models using satellite imagery. A similar effort to use existing ATAI data on social networks in Malawi to develop a remote sensing model is underway on the part of CEGA researchers Jeremy Magruder and Solomon Hsiang.

ATAI is actively pursuing the opportunity to curate and leverage more data from the 50 randomized evaluations it has carried out throughout sub-Saharan Africa and South Asia but needs partners and investment to help support the coordination required to make this happen. In other cases, new data collection efforts will be required. ATAI hopes to join forces on this as well, helping to fill gaps in the existing sources of ground data.  

The Center for Effective Global Action recently held a large-scale event on geospatial applications for international development with a focus on agricultural development. Please stay tuned for slides for the event (to be posted at Geo4Dev) and listen to a panel on remote sensing for agricultural development, including myself, Marshall Burke, University of California, San Diego professor; environmental scientist Jen Burney; and Daniel Platt, solutions engineer with Harris Geospatial Solutions. Spoiler alert: during the panel, Marshall predicts it may only take a year to provide a decent estimate of maize productivity in East Africa given existing imagery and models he and others have been able to validate for this context. Very cool.