Rice crop classification and yield estimation using multi-temporal sentinel-2 data: a case study of terai districts of Nepal

Abstract

Crop monitoring, especially in developing countries, can improve food production, address food security issues, and support sustainable development goals. Crop type mapping and yield estimation are the two major aspects of crop monitoring that remain challenging due to the problem of timely and adequate data availability. Existing approaches rely on ground-surveys and traditional means which are time-consuming and costly. In this context, we introduce the use of freely available Sentinel-2 (S2) imagery with high spatial, spectral and temporal resolution to classify crop and estimate its yield through a deep learning approach. In particular, this study uses patch-based 2D and 3D Convolutional Neural Network (CNN) algorithms to map rice crop and predict its yield in the Terai districts of Nepal. Firstly, the study reviews the existing state-of-art technologies in this field and selects suitable CNN architectures. Secondly, the selected architectures are implemented and trained using S2 imagery, groundtruth and auxiliary data in addition for yield estimation. We also introduce a variation in the chosen 3D CNN architecture to enhance its performance in estimating rice yield. The performance of the models is validated and then evaluated using performance metrics namely overall accuracy and F1-score for classification and Root Mean Squared Error (RMSE) for yield estimation. In consistency with the existing works, the results demonstrate recommendable performance of the models with remarkable accuracy, indicating the suitability of S2 data for crop mapping and yield estimation in developing countries.