Cotton leaf water potential prediction based on UAV visible light images and multi-source data

Regular monitoring of crop moisture conditions has the potential to enhance both crop production efficiency and water utilization efficiency. To generate a moisture status map using Unmanned Aerial Vehicle (UAV) visible light imagery, the focus of the present study was on developing a predictive model for cotton leaf water potential utilizing UAV visible light imagery and multi-source data. By processing the R, G, and B channel values of UAV visible light images, the grayscale histogram mean was extracted as an indicator of vegetation growth status. The Extreme Learning Machine (ELM) was employed in conjunction with the Sparrow Search Algorithm (SSA) optimization to create the SSA-ELM model. This model integrates the grayscale histogram method with soil data and meteorological data, and trains with leaf water potential as the output feature, establishing a cotton leaf water potential prediction model. The SSA-optimized model exhibited significantly improved performance on two test sets in Alaer and Tumushuke. After SSA optimization, the R² values increased by 0.05 and 0.04 respectively, reaching 0.80 and 0.85, while the RMSE values were 0.211 MPa and 0.215 MPa, indicating the accuracy and stability of the model predictions. By utilizing data partitioned into three distinct cotton growth stages, the model was remodeled, followed by assessment using test sets from the same stages in the cities of Alaer and Tumushuke in 2023. On the Flowering test set, the R² values were 0.70 and 0.71, with corresponding RMSE values of 0.272 MPa and 0.269 MPa, respectively. For the Full boll test set, the R² values were 0.78 and 0.76, with corresponding RMSE values of 0.247 MPa and 0.245 MPa. Lastly, within the Open boll stage test group, the R² values reached 0.75 and 0.80, with corresponding RMSE values of 0.240 MPa and 0.245 MPa. These outcomes further affirm the robustness, adaptability, and generalizability of the method for predicting leaf water potential based on RGB imagery combined with multi-source data across different geographical settings.