When We Can't Ask Farmers How They're Doing, Their Fields Tell The Story: Using Satellite Data to Estimate Cropped Area and Crop Production in Mali

USAID’s Bureau for Resilience and Food Security and the US Government’s Feed the Future Initiative implement development activities to improve resilience, food security, nutrition, and livelihoods, with a focus on smallholder farmers and agriculture-led growth. Many of these programs are carried out in remote, hard-to-reach locations; and they are inherently challenging to design, implement, monitor, and evaluate without the use of newer technologies such as satellite remote sensing data and the corresponding methods for their appropriate use. Some development programs are implemented in areas experiencing instability or conflict, yet there is still a need to monitor those areas to track programmatic results and ensure accountability to stakeholders.

The Challenge: Feed the Future ZOI survey coverage in Mali
While programmatic activities in conflict-affected communities may be led by local residents, household survey operations for collection of monitoring data typically are not. Areas affected by instability or conflict can be difficult or dangerous for non-residents to access. Consequently, the survey field teams tasked with collecting representative monitoring data in Feed the Future’s Zone of Influence (ZOI – the sub-national geography where Feed the Future focuses its programming) may not be able to enter the required geographies, rendering the survey data unrepresentative, or they may be placed at risk of violence if they draw attention while attempting to enter or stay overly long in an unstable area.
This was the challenge faced by the teams charged with implementing a survey of Mali’s Feed the Future ZOI (which includes Sikasso, Mopti, and parts of Timbuktu; see Figure 1). Fieldwork for the 2019 Mali ZOI Survey was initiated; however, conflict-related insecurity concerns developed over time, and it was finally determined that the survey fieldwork could only be safely implemented in Sikasso, leaving Mopti and Timbuktu without monitoring data.
These circumstances underscored the need to consider alternative data sources and methods for monitoring program performance in areas experiencing instability or conflict.

The Solution: Use earth observations data and machine learning methods to meet selected monitoring needs
To overcome the significant challenge of access on the ground in conflict-affected geographies, USAID is working with leading researchers from the NASA Harvest consortium to develop capabilities to monitor and evaluate aspects of our programs. This work involves using data from Earth-observing satellites combined with the ground-referenced data collected in the 2019 ZOI Survey in Sikasso and by Lutheran World Relief colleagues in Segou.
The 2019 ZOI Survey in Sikasso collected detailed information from households on agriculture, including collection of georeferenced field boundaries where selected crops (maize, millet, okra) were cultivated (see Figure 2), soil characteristics in those fields using LandPKS, farmer-reported crop production estimates, and farmer use of agricultural technologies, inputs, and other practices. Some of the agriculture-related data collected during the 2019 Mali ZOI Survey fieldwork, including those that allow for estimation of key Feed the Future indicators like yield, represent biophysical attributes of specific geographies that could be observable by satellite remote sensing methods.
Key Questions:
Given the need to provide monitoring data in the Feed the Future Zone of Influence for geographies in Mali and elsewhere that are too insecure to implement standard ZOI Survey operations, NASA Harvest is working to answer several key research questions:
- To what extent can the detailed, georeferenced agriculture data collected in the 2019 Mali ZOI Survey, as well as other similar agricultural data, be used in conjunction with high-resolution earth observation imagery to train machine learning algorithms to allow for accurate estimation of planted area, crop classification, production, and yield across the entirety of the Feed the Future Zone of Influence for maize, millet, and okra?
- To what extent can the methods and tools developed be scaled and utilized in other inaccessible areas (for example, other Feed the Future target countries facing similar constraints to household survey operations)?
- Are the agricultural data collected by Feed the Future’s ZOI Surveys sufficient for training satellite models for crop identification and yield indicators? Are there ways in which the surveys could be enhanced in this regard?
Driving Progress:
The NASA Harvest team at the University of Maryland (comprised of researchers Brian Barker, Inbal Becker-Reshef, Mehdi Hosseini, Michael Humber, Hannah Kerner, Catherine Nakalembe and Ritvik Sahajpal) is driving the effort forward by first identifying available high-resolution satellite imagery for past seasons that correspond to the available ground-referenced data, as well as for future seasons. They are also cleaning, assessing, and analyzing the plot and crop data from the 2019 Mali ZOI Survey in conjunction with high-resolution satellite data that they have acquired and processed, including HLS v1.5 (Harmonized Landsat and Sentinel-2 at 30m, ~2-3 day revisit), Sentinel-2 (10m, ~5 day revisit), PlanetScope (3m, ~daily), and Sentinel-1 SAR (20m).
Below are images of the same scene showing several of the ground-referenced plot polygons collected during the 2019 Mali ZOI Survey, as seen by the Landsat, Sentinel-1, Sentinel-2, and PlanetScope satellites (Figures 3-6):




The team’s next steps will be to use satellite imagery to identify cropped areas, utilizing the 2019 Mali ZOI Survey data from Sikasso and other available ground-referenced data including from Lutheran World Relief and ICRISAT. They will also gather new ground data (crop type and yield) for training and validation purposes, leveraging remote sensing field experts and Lutheran World Relief field agents.
These data will then be used to train machine-learning algorithms that will help produce cropland maps, dominant crop type maps, indicators of crop condition, quantitative yield estimates, and field boundary delineation. These products will help USAID, Feed the Future, and the Government of Mali understand the agricultural situation in the country and the results of our programmatic efforts at a time when we can’t visit and learn directly from farmers themselves––We rely on their fields, and the satellites that observe them, to tell the story.