Regional-scale rice-yield estimation using stacked auto-encoder with climatic and MODIS data: a case study of South Korea

In South Korea, rice is the most important grain crop that it is crucial to develop yield estimation model for supporting sustainable agriculture and national food security. The main objectives of this paper are (1) to propose a Deep Learning (DL) algorithm, the Stacked Sparse Auto-encoder (SSAE), for rice-yield estimation using climatic and Moderate Resolution Imaging Spectroradiometer (MODIS) data; (2) to choose scenarios showing the best combined performance in terms of length of crop season (before and after harvest) and aggregation periods (7, 10, 15, and 30 days) of climatic data; and (3) to analyse the results in both the temporal and spatial perspectives. In this procedure, the SSAE model was built around the study objectives and compared with the artificial neural network (ANN) model for evaluation of its performance. According to the results, the combined 15-day-aggregated climatic data (between June and August) and MODIS data were selected as the optimal feature set for rice-yield estimation in the study area, Jeolla-do; the SSAE model outperformed ANN, showing root mean square error (RMSE) and RMSE% of 33.09 kg(10a) ⁻¹ (5.21 kg(10a) ⁻¹ lower than ANN) and 6.89% (1.14% lower than ANN). Throughout the experiments, the rice-yield-estimation potentiality of a DL algorithm, namely the SSAE, was substantiated.