Optimum sampling window size and vegetation index selection for low-altitude multispectral estimation of root soil moisture content for Xuxiang Kiwifruit

Early detection of water stress is essential for orchard management; however, existing methods are unable to accurately monitor individual plant water status over large areas, and the shaded nature of kiwifruit orchards further complicates the monitoring of root soil moisture content (RSMC). In this study, we used multilayer perceptron (MLP) and canopy vegetation indices, estimated by unmanned aerial vehicle remote sensing, to predict RSMC at a depth of 40 cm during the fruit expansion stage in Kiwi orchards (August 2021 and 2022). Using artificial intelligence algorithms, we assessed the effect of image sampling size and model input combi-nations on estimation accuracy. In the inversion model building process, 247 MLP models were built based on a combination of eight vegetation indices and trained with 18 datasets according to different sampling widths to compare model evaluation parameters. To reduce the amount of input parameters, we selected parameters based on the Pearson correlation between the input (individual vegetation indices) and output (RSMC) and finally compared the coefficients of determination of the models for different combinations of vegetation indices. We found that the coefficient of determination and explained variance score increased and the root mean square error decreased as the model inputs increased. The coefficient of determination and root mean square error had a strong positive correlation with sampling width (r = 0.7082 and 0.7273, respectively). When training a model with green index-green normalized difference vegetation index-optimized soil-adjusted vegetation index (OSAVI) or -modified SAVI (MSAVI) as inputs, the accuracy of the model remained approximately 0.7447, which did not vary significantly from models with eight vegetation indices as inputs but presented a simplified network structure. This study provides a model-building framework for the analysis of soil moisture conditions in shaded orchards.