Abstract: To address the challenges of complex phenological variations, feature redundancy, and insufficient recognition accuracy in remote sensing identification of fruit tree species under complex orchard environments, a typical orchard area in Aksu, Xinjiang, was selected as the study area, and a fruit tree species identification method integrating phenological information and multi-source remote sensing data was proposed. Multi-source remote sensing features were constructed using Sentinel-2 optical remote sensing data, Sentinel-1 radar data, and terrain factors. Phenological window optimization was used to select remote sensing phases with higher class separability, and the Boruta algorithm was adopted to screen multi-source features and reduce feature redundancy. On this basis, a Bayesian optimization method based on TPE (Tree-structured Parzen Estimator) was introduced to optimize the key hyperparameters of the LightGBM model, and was combined with a Stacking ensemble learning strategy to construct the fruit tree species classification model. Meanwhile, SHAP (SHapley Additive exPlanations) was used to analyze the contribution of different remote sensing features to the classification results. The proposed method achieved an overall accuracy of 90.62% and a Kappa coefficient of 0.889 3 on the test set, and its classification accuracy was higher than that of single-model methods. The combination of multi-source remote sensing information and machine learning methods improved the recognition accuracy of fruit tree species and could provide a methodological reference for regional orchard resource surveys and agricultural management.

Key words: multi-source remote sensing, fruit tree species identification, phenological information, feature selection, machine learning