Comparative analysis of deep neural network and artificial neural network for predicting wheat water content using UAV multispectral and thermal imagers

Water is a vital component for the growth of wheat and the quality of its grains. The measurement of plant water content (PWC) serves as a crucial indicator for assessing the water status of crops, thereby guiding effective irrigation management practices. Recent advancements in technology, particularly the use of multispectral and thermal imaging from unmanned aerial vehicles (UAVs), offer the capability to capture a comprehensive view of PWC variability across agricultural fields. This study aimed to create predictive models for PWC utilizing artificial neural networks (ANN), deep neural networks (DNN), and traditional stepwise regression techniques, all based on the analysis of multispectral and thermal imagery. To effectively evaluate the water content in wheat, we combined high-resolution thermal and multispectral imaging techniques with machine learning approaches. This assessment was carried out through three distinct experiments, which included one conducted in a rainout shelter and two performed under rainfed conditions. The findings demonstrated that UAV-derived multispectral imagery, when coupled with machine learning models, can effectively predict wheat plant water content with remarkable precision. Notably by considering all the dataset, DNN model exhibited superior performance (R2=0.96, ENS=0.98, RMSE=1.37%, MAE=0.98%) compared to both the ANN (R2=0.95, ENS=0.95, RMSE=1.88%, MAE=1.46%,) and the stepwise regression model (SRM) (R2=0.67, ENS=0.51, RMSE=10.79%, MAE=9.03%). Across all machine learning approaches, both the DNN and ANN models significantly outperformed the stepwise regression model in predictive accuracy. The statistical outcomes derived from the calibration phase indicate that both the trained Artificial Neural Networks (ANN) and Deep Neural Networks (DNN) serve as effective instruments for accurate prediction of PWC. Notably, the DNN network demonstrates superior accuracy compared to ANN.