The lack of timely, high-resolution data on agricultural production is a major challenge in developing countries where such information can guide the allocation of scarce resources for food security, agricultural investment, and other objectives. While much research has suggested that remote sensing can potentially help to address these gaps, few studies have indicated the immediate potential for large-scale estimations over both time and space. In this study we described a machine learning approach to estimate smallholder maize yield in Ethiopia, using well-measured and broadly distributed ground truth data and freely available spatiotemporal covariates from remote sensing. A neural networks model outperforms other algorithms in our study. Importantly, our work indicates that a model developed and calibrated on a previous year's data can be used to reasonably estimate maize yield in the subsequent year. Our study suggests the feasibility of developing national programs for the routine generation of broad-scale, high-resolution estimates of smallholder maize yield, including seasonal forecasts, on the basis of machine learning algorithms and well-measured ground control data and currently existing time series satellite data.

https://doi.org/10.1016/j.crope.2023.07.002