Developing a crop yield estimation method from remotely sensed metrics using artificial intelligence; case study: Spain and Western Kenya

Details

Supervisors

Defourny, Pierre ; Houdmont, Pierre

Faculty

Faculté des bioingénieurs

Degree label

Master [120] en sciences agronomiques et industries du vivant, à finalité spécialisée: Geo-Information Science and Earth Observation for Environmental Modelling and Management

Abstract

Crop yield estimation is vital for combating food insecurity by aiding policy making at national and regional levels, and supporting farm-level decision making. However, it is a complex task, especially at the field level, due to limitations in data availability, quality, and standardization, as well as the lack of accurate crop classification maps. Despite these challenges, studies in this area demonstrate that integrating remotely sensed data with artificial intelligence improves yield estimation since ML and deep learning models can detect and analyze complex patterns in large datasets. Thus, this study focuses on field-level yield estimation, aiming at developing a method that can be applied in two diverse contexts: wheat in Spain and maize in Western Kenya. The main datasets used in both cases are LAI derived from Sentinel-2 and meteorological data from the ERA5 Land product. In Western Kenya, additional datasets such as elevation, soil, and long-term regional yield averages are incorporated to enhance estimation. ML algorithms including RFR, HGBR, VR, MLP, and ETR are employed for prediction. In Spain, VR marginally outperforms other models with an R² of 0.59, RMSE of 940.22 kg/ha and a MAE of 714.02 kg/ha at the field level, while HGBR achieves the best results at the provincial level with an R² of 0.78, RMSE of 315.34 kg/ha and a MAPE of 258.04 kg/ha. LAI-related features are the most significant predictors in Spain, whereas elevation is the most important feature in Kenya. The study in Western Kenya is constrained by data quality issues, leading to minimal influence of LAI-related features on the predictions. As a result, the trained RFR and HGBR models exhibit relatively low accuracy.