An automated framework to develop field level crop yield prediction models using remote sensing [A case study on the Wonji-Shoa region in the Awash Basin, Ethiopia]

Attaining national food security and adopting sustainable agricultural practices has gained unprecedented global attention in recent years. Accurate and reliable crop yield estimates provide quantitative information to decision-makers to formulate policies on crop management, risk mitigation, import-export strategy, maintaining national food reserves, and contributing significantly to the nation's food security. Such forecasts also help farmers understand associated risks, decide input costs, and improve field management, thereby making agriculture more sustainable. Remote Sensing (RS) based crop monitoring has provided immense opportunities to researchers to acquire timely and reliable crop growth status information at diverse scales, and at lower costs. Although crop yield prediction at the field level gained importance to formulate efficient water and farm management practices, still there is no easyto- use tool that can be utilized to predict crop yields at the farm level, even after the availability of high-resolution satellite data sets. Inter-field heterogeneity arising from varied field management and agricultural practices further necessitates this forecasting. Therefore, this study aimed to develop RS based crop yield prediction models per field using multi-source high-resolution satellite data and estimate the relative predictive power of different RS input variables. Further, different regression algorithms were compared to test their suitability in building such a data-driven model. Finally, it was also attempted to determine the time before harvest at which fairly accurate yield estimates can be made available. To carry out this research, Wonji-Shoa sugarcane farms located in the southwestern region of the Awash River basin in Ethiopia were chosen as the study area. WaPOR derived Net Primary Production (NPP) data was used as an alternate variable to yield; since spatially distributed and temporally continuous information on crop yield is usually not available. In addition to deriving model input variables from high-resolution imagery data from Sentinel-2 and Sentinel-1, WaPOR derived actual evapotranspiration (ETa) data was used as a third category of control variables, as it is representative of water stress experienced by the crop. Field-level phenology, another significant input variable, was derived using a semi-automated methodology based on finding the correlation between NDVI time-series data of all the fields, followed by visually inspecting a few. Finally, five regression algorithms viz. Multivariate Linear Regression (MLR), Stepwise Multivariate Linear Regression (SMLR), Boosted Regression Tree (BRT), Support Vector Regression (SVR), and Random Forest Regression (RFR) were tested to develop field-wise crop prediction models in the statistical environment of R software and compared to determine the most suitable algorithm. The research results revealed that for sugarcane NPP predictions in the study area, linear regression models (MLR and SMLR) yielded more accurate results than the non-linear machine learning models (BRT, SVR and RFR). The highest accuracy was achieved for MLR models for which reliable estimates could be made around 4 months prior to the harvest (R2=0.89, RMSE=159 gC/m2, MAE= 123gC/m2, and MAPE=10.8%). The non-linear machine learning models could not provide reliable estimates of sugarcane NPP. As compared to Vegetation Indices (VIs), cumulative VIs were found to have higher prediction powers in forecasting future sugarcane NPP. Cumulative EVI was determined to be the variable with the highest predictive power, followed by radar derived VH and WaPOR deduced ETa. By carrying out a field-level NPP analysis, it was concluded that information from such a model based prediction tool can contribute significantly to the uplifting of socio-economic status of the farmers in the region. This study contributes an automated framework to develop empirical crop yield prediction models per field iteratively, using high-resolution RS data that predict sugarcane NPP for every field separately. Nevertheless, the study has its limitations, including necessity for extensive testing before implementation in other geographical regions, dependency on long-term yield data, issues with phenology deduced using NDVI time-series data, statistically biased results, amongst others