Inversion study on the nitrogen content of jujube canopy based on multi-source data fusion combined with integrated learning algorithm

• Integrated learning algorithm optimizes multi-source data processing to improve nitrogen inversion accuracy. • NDT technology promotes precise nitrogen management and reduces agricultural production costs. • Combination of growth stage optimization and integrated learning to provide personalized management strategies. Red jujube (Ziziphus jujuba Mill.) is a characteristic and economically important crop in Xinjiang, playing a vital role in regional agricultural development and cultural preservation. Nitrogen, as a crucial macronutrient, significantly influences plant growth, photosynthetic efficiency, and fruit quality. Accurate estimation of canopy nitrogen content is essential for optimizing fertilization strategies and improving both yield and product quality. This study establishes a nitrogen inversion method for red jujube canopies based on multi-source data, aiming to provide a scientific foundation for precision agriculture in Aral City, Xinjiang. The approach integrates three data types: Sentinel-2 multispectral satellite imagery, ground-based hyperspectral reflectance data (325–1075 nm), and standardized leaf image data acquired using dark-box cameras. Following preprocessing procedures such as noise filtering and outlier removal, 13 color channels and 751 hyperspectral bands were extracted. Correlation analysis identified four image color channels (B, a, b, S) and 34 hyperspectral bands as sensitive features associated with leaf nitrogen content. Multiple machine learning models were constructed to evaluate inversion performance, including backpropagation (BP) neural networks, random forest (RF), and ridge regression, as well as ensemble combinations of these algorithms (e.g., BP+RF, RF+ridge, BP+RF+ridge). Results demonstrated that models using fused multi-source data significantly outperformed those based on single data sources. The optimal model, combining BP, RF, and ridge regression, achieved an R² of 0.88044, with a mean absolute error (MAE) of 2.3576 and a root mean square error (RMSE) of 2.0571. Compared to baseline models, this method improved R² by up to 27.58 % and reduced errors by >50 %. Among individual algorithms, RF showed superior performance. Ensemble learning models exhibited enhanced accuracy and robustness, particularly in managing complex heterogeneous data. Across different growth stages, the best performance was observed during flowering stage 1 and fruit maturity, while inversion accuracy during flowering stage 2 was lower. This may be due to limited sample diversity or weak representation of traits, transition from asexual to reproductive growth resulting in changes in nitrogen metabolic pathways and spectral response, and other related reasons. Overall, the findings underscore the advantages of multi-source data fusion and ensemble modeling in nitrogen content inversion, offering valuable insights for precision management in red jujube cultivation.