Developing wheat crop yield estimation method for Spain from remotely sensed metrics using artificial intelligence

Details

Supervisors

Defourny, 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

Agricultural production plays a crucial role in the global environment, society, and economy, especially the strategic crops directly impacting food security. An accurate crop yield estimation is important for farmers and policymakers to make informed decisions about agricultural practices, production, trade, and food security. However, the current methods used by national statistics offices are based on traditional approaches and uncertain data sources, leading to inaccurate yield estimates. To address this issue, this research aims to provide a state of the art wheat crop yield estimation method at the field level, utilizing remotely sensed features derived from satellite earth observation data and leveraging artificial intelligence techniques. The study is conducted on 849 wheat fields located in Castilla y Leon, Castilla-La-Mancha, and Comunidad Valenciana regions of Spain for 2019. Various multivariate, machine learning, and deep learning models, including LASSO, DTR, RFR, SVR, and MLP, are used to predict crop yield at the field level. MLP outperforms the other models with an R2, MAE, and RMSE of 0.64, 552.65 kg/ha, and 733.53 kg/ha, respectively. The RFR model followed closely, achieving an R2, MAE, and RMSE of 0.64, 565.62 kg/ha, and 744.31 kg/ha, respectively. The DTR model achieved an R2, MAE, and RMSE of 0.55, 630.76 kg/ha, and 817.55 kg/ha, the SVR model obtained 0.48, 680.32 kg/ha, and 876.22 kg/ha, and LASSO model achieved 0.48, 689.65 kg/ha, and 900.45 kg/ha, respectively. Additionally, the feature importance score to identify the significant contributors to yield estimation revealed that vegetation features corresponding to LAI are the most significant. Moreover, among features recorded at different growing stages, the senescence stage emerged as the most important factor in predicting yield at the field level. Overall, the results demonstrate the potential of artificial intelligence in accurate crop yield estimation, even at the field level, and can be used as an alternative to traditional methods. Further research can be conducted to explore the potential of other machine and deep learning models and to investigate the impact of different input features on model performance.