Abstract
With the ever-growing urgency of food insecurity and the threat of climate change, there is increasing interest in using artificial intelligence for Earth observations (AI-EO) for agriculture, particularly in Sub-Saharan Africa (SSA). This paper provides an overview of the primary research areas within AI-EO for agriculture in SSA. We discuss examples and limitations of current research as well as opportunities for future work. In addition, we identify ten key considerations for future efforts involving AI-EO for agriculture in SSA.
Introduction
While "precision agriculture" in the Global North often involves high-cost infrastructure like autonomous machinery, these solutions remain largely inaccessible in Sub-Saharan Africa (SSA). However, the region faces an urgent need to mitigate climate change impacts and address chronic food insecurity. AI-EO—the application of machine learning to satellite data—offers a scalable and affordable alternative.
The primary challenge lies in the mismatch between existing models and the SSA landscape. Most AI models are trained on data from the Global North, failing to account for the heterogeneous nature of SSA agriculture, which is dominated by smallholder farms (often <1 hectare) and complex intercropping practices that confuse standard algorithms.
Technical Applications in AI-EO
Research in AI-EO for SSA agriculture generally falls into five primary categories, moving from simple classification to complex predictive modeling:
- Cropland Mapping: Identifying where crops are growing versus other land covers.
- Crop Type Mapping: Distinguishing between specific crops (e.g., maize vs. beans).
- Field Boundary Delineation: Using instance segmentation to identify individual farm plots.
- Yield Estimation: Predicting harvest volume per unit area.
- Pest and Disease Detection: Identifying regional anomalies and threats.
Technically, the field is transitioning from traditional "shallow" learners like Random Forests and XGBoost toward deep learning architectures. U-Nets are increasingly used for segmentation tasks, while Recurrent Neural Networks (RNNs) are employed to capture the temporal signatures of crop growth cycles.
The Data Bottleneck and Resolution Gap
A critical barrier to progress is the lack of high-quality "ground-truth" labels. While satellite imagery from sources like Sentinel-2 (10m) and Landsat (30m) is available, the resolution is often too coarse for the small, irregular fields typical of SSA. A single 10m pixel may contain multiple crops and trees, leading to significant classification errors.
Ten Key Considerations for Future Research
To ensure AI-EO solutions are ethical and practically useful, the authors propose ten essential considerations, including:
- Decolonizing Research: Avoiding "parachute science" by establishing meaningful partnerships with local institutions and ensuring data collection respects local consent.
- Resource Context: Designing models that do not require prohibitive data costs or high-end hardware for end-users.
- Interdisciplinary Collaboration: Integrating agronomists, climate scientists, and farmers into the development process from the outset.
- Addressing the Resolution Gap: Prioritizing multi-fidelity fusion to combine free, low-resolution data with expensive, high-resolution commercial imagery.
Results and Impact
The performance gap between global and localized models is stark; global models have shown user accuracy as low as 17% in parts of Eastern Africa. Conversely, localized "Rapid Response" maps, such as those deployed in Togo, demonstrate that AI-EO can effectively support policy and aid distribution. Furthermore, yield forecasting can now provide critical insights 2 to 2.5 months before harvest, offering a vital window for food security interventions.
Future Directions
Future research must tackle "grand challenges" such as intercropping, where new methods are needed to disaggregate signals from mixed-crop fields. Additionally, the field should move toward few-shot learning and meta-learning to reduce reliance on expensive labeled datasets. Ultimately, the goal is the institutionalization of these tools within national ministries of agriculture across SSA to ensure long-term sustainability.
Summary updated