Abstract
Total soil nitrogen (TN) is a key indicator for plant growth, and its accurate prediction is crucial for agricultural production and environmental management. This study aims to enhance prediction accuracy of TN by integrating multi-temporal synthetic imagery with environmental data, combined with machine learning methods. Seven single-temporal Sentinel-2 images from 2019 to 2020 were acquired, and four types of multi-temporal synthetic imagery were generated using maximum, minimum, mean, and median synthesis methods. Spectral bands and vegetation indices were separately extracted from these images. Extreme gradient boosting algorithm was then employed to predict TN content using these derived data from remote sensing, topographic and climatic data at four spatial resolutions (10 m, 20 m, 30 m and 60 m). Results indicate significant model performance variations across land use types: single-temporal imagery achieved highest prediction accuracy in orchards (R² = 0.73), while multi-temporal mean synthetic imagery performed best in dry land (R² = 0.63) and paddy fields (R² = 0.61). Spectral bands (particularly B8 and B11) were identified as the most critical predictive variables. This study demonstrates the application value of multi-temporal remote sensing synthesis technology in soil total nitrogen monitoring, providing a technical pathway for advancing precision agriculture and sustainable land management.