Cotton yield prediction utilizing unmanned aerial vehicles (UAV) and Bayesian neural networks

In this study, we propose to utilize the canopy features such as canopy cover (CC), canopy height (CH), canopy volume (CV), and excess greenness index (ExG) extracted from UAVs imagery and Bayesian neural network (BNN) to develop a pipeline for predicting cotton crop yield. The pipeline consisted of two components, data imputation which dealt with irregular spatial and temporal data and yielded prediction with uncertainty quantification. The data was collected from producers' fields in 2020, 2021, and 2022. To assess the performance of the proposed BNN model particularly for the generalization across years, three other models including support vector regression (SVR), random forest regression (RFR), and multiple layer perceptron (ML) were used. In cross year test, our pipeline produced better results with root mean squared error (RMSE) of 365.22 kg ha(-1), mean absolute error (MAE) of 294.5 kg ha(-1), and R-2 of 0.67 between actual yield and the yield prediction by the model. In addition, feature importance analysis showed that the combination of CC, CH, CV, and ExG followed by the combinaton of CC and ExG variables outperformed other combinations or single variables.