| Sumario: | Abstract1. CONTEXTAddressing the limitations of scaling agronomic recommendations, which are usually confined to small areas, requires a spatial framework for characterizing production environments in a timely and cost-effective manner.2.OBJECTIVEThis study aimed to introduce a data-driven framework to characterize rainfed wheat crop production environments in Ethiopia. The framework entails mapping of the annual rainfed wheat area and the delineation of crop-specific and dynamic agro-ecological spatial units (ASUs).3. METHODSAn ensemble machine learning approach built upon time-series satellite images and environmental data was used for crop type mapping while pixel- and object-based clustering algorithms were used to delineate dynamic ASUs from two temporal perspectives: annual ASUs for the 2021 and 2022 growing seasons to assess short-term dynamism, and ASUs from aggregated data (2016 – 2022) to capture long-term variations in the production environment.4. RESULTS AND CONCLUSIONSModel evaluation showed that the ensemble of random forest, gradient boosting, and classification and regression trees predicted wheat cropland in the 2021 and 2022 growing seasons with 88-90% accuracy. A concordance in defining ASUs between pixel- and object-based approaches was observed with consistency and dynamism in ASUs from 2021 to 2022 and between single-year and aggregated ASUs across approaches. This consistency and dynamism in ASUs highlight the spatial scalability and temporal flexibility of the framework, which allows for characterizing production environments across scales and analyzing trends and fluctuations, providing valuable insights for addressing food security and environmental challenges.5.SIGNIFICANCEThe developed spatial framework could facilitate future yield gap analysis and agronomic assessments for rainfed wheat in Ethiopia and be transfered to other crops and production environments.
|
|---|