Precision Agriculture General Agricultural and Biological Sciences
Estimation of leaf nitrogen content of typical fruit trees in arid regions based on fusion of UAV multispectral information and texture features
Abstract
Accurate estimation of leaf nitrogen content (LNC) is essential for precision fertilization in arid orchards. However, the "same spectrum, different objects" phenomenon in single spectral data often limits inversion accuracy across diverse tree species. This study aimed to develop an LNC estimation framework for walnut, jujube, and apricot trees by integrating UAV-based spectral and texture features using an interpretability-driven approach. UAV multispectral imagery was utilized to extract 8 optimized spectral indices (SI) and 40 texture features. The SHAP (SHapley Additive exPlanations) algorithm was introduced to quantify and select optimal feature combinations. Three machine learning algorithms—K-Nearest Neighbors (KNN), Ridge Regression, and Support Vector Machine (SVM)—were compared to evaluate model performance. (1) SHAP analysis revealed distinct response patterns: walnut LNC was dominated by texture features (NIR _ Mea, SHAP = 0.282), while jujube favored background-resistant indices (SAVI, SHAP = 0.562). (2) For LNC estimation based on single spectral indices, algorithm performance showed species-dependent differences. KNN performed relatively well for walnut and apricot, whereas RR and SVM achieved higher validation accuracy for jujube. (3) Compared with single-spectral-index models, the SHAP-selected spectral–texture fusion model achieved improved LNC estimation for all three fruit tree species, with the most pronounced improvement observed for walnut, where validation R², RMSE, and RPD improved by approximately 22.2%, 13.2%, and 15.4%, respectively. (4) The mixed-species model achieved validation R², RMSE, and RPD values of 0.64±0.06, 2.15±0.18 g/kg, and 1.71±0.15, respectively, indicating its potential for approximate LNC estimation in mixed orchards without prior tree-species classification. Overall, the proposed SHAP-KNN spectral-texture fusion framework may provide a useful technical reference for monitoring fruit tree leaf nitrogen concentration in arid regions and for supporting precision orchard management.