Skip to main content

Using AI and Earth Observations in Sub-Saharan Agriculture Requires Special Considerations to Ensure Success

header image


Earth observation (EO) data collected from satellites orbiting the Earth is helping scientists, policymakers, and decision-makers learn more about our planet and make better-informed decisions. With recent advancements in artificial intelligence (AI), researchers can process massive amounts of data and perform analyses that were impossible to perform until recently. One area that has significantly benefited from the synergy of AI and EO is agriculture. For example, AI-EO for agriculture has allowed for the rapid assessment of crop damage after natural disasters and the ability to evaluate pest impacts.

However, while AI-EO applications for agriculture are immense, special considerations must be accounted for when adapted in different parts of the globe. In Sub-Saharan Africa, for instance, the frequency of small-holder agriculture—small-scale fields composed of heterogeneous crops—is in contrast to the larger mono-crop fields common in higher agricultural production countries.

These smaller fields require large amounts of high resolution EO data to be detected, resulting in higher computational processing demands for any AI analysis. Additionally, many existing AI techniques were developed for the larger, homogeneous fields and are not generalizable across the heterogeneous crop landscapes typical in Sub-Saharan Africa.

NASA Harvest’s Catherine Nakalembe (Harvest Africa Lead) and Hannah Kerner (Harvest AI and Data Lead) recently wrote an open letter detailing the difficulties that come with utilizing AI and EO for agriculture in Sub-Saharan Africa. In the letter, they discuss key AI-EO applications and the current implementation limitations in Sub-Saharan Africa. They also list ten considerations that scientists and stakeholders should incorporate into future work to ensure that AI-EO research in the region has a positive impact.


Popular AI-EO Applications

Cropland and Crop Type Mapping

Cropland maps show where crops are growing on a landscape, delineating between cropped vs. uncropped land, whereas crop type maps indicate specific crops at a specific location (e.g. wheat). In addition to informing users of where agricultural land is within a region; both products also are important for yield and crop condition models as they help to restrict analysis to cropped land. These maps vary in scale dependent on the spatial resolution of the EO data used to create them. Within Sub-Saharan Africa, many fields are quite small and thus require very high-resolution data to map, which is not freely available. Further, the creation of cropland and crop type maps require labeled African agriculture datasets to train the classification models.


Yield Estimation Models

These models estimate the amount of crop harvested per unit area, which, when combined with cropland and crop type maps, allow for the estimation of agricultural production. There have been several yield estimate models created for high-exporting countries, however, few studies have focused on developing yield estimates for Africa. As many of the crop production increases seen across Africa are due to the expansion of cropland rather than yield intensification, developing and improving yield estimate models for the continent can provide the information required to support yield increase strategies.


Field Boundary Delineation

Understanding the boundaries of individual crop fields helps guide field data collection and area estimation. Knowing the limits of a particular field also allow researchers to assess better the impacts of different inputs like irrigation, tillage management, and fertilizer application. Similar to cropland and crop-type maps, the main factor limiting progress in developing field boundary products is the lack of publicly available very high-resolution EO data.


Pest, Disease, and Anomaly (PDA) Detection

Quickly detecting pest, disease, and other anomalies within a field is vital to limiting potential damage and salvaging as much of the specific crop as possible. AI-EO for agriculture provides a much less labor-intensive method for this detection and allows for a much more pinpoint selection of problem areas. As with cropland and crop type maps and field boundary delineation, PDA Detection is hampered in Sub-Saharan Africa by the lack of publicly available very-high-resolution EO data.


Special Considerations for AI-EO for Agriculture in Sub-Saharan Africa

Nakalembe and Kerner also listed ten considerations that scientists and stakeholders should account for developing future research plans and products.These considerations are paramount in ensuring the utmost impact while simultaneously mitigating any adverse consequences.. A full description of each consideration can be read in the original letter.

  1. Interdisciplinary teams are a requirement.
  2. Consider the resource context
  3. Prioritize methods for limited labeled data
  4. Methods should be transparent and reproducible.
  5. Work with stakeholders from the beginning.
  6. Decolonize research methods and practices.
  7. Form meaningful partnerships with local institutions.
  8. Institutionalized investments.
  9. Open access to high-resolution imagery.
  10. Limitations of AI-EO solutions should be assessed and communicated.

The full letter, Considerations for AI-EO for agriculture in Sub-Saharan Africa, is available open access in Environmental Research Letters.

News Date
Jun 20, 2023
Keelin Haynes, Catherine Nakalembe, Hannah Kerner