Enhancing soil total nitrogen prediction in rice fields using advanced Geo-AI integration of remote sensing data and environmental covariates

• The lack of high-resolution STN data for Indonesian rice fields results in significant fertilizer mismanagement. • Integrating RS indices and environmental covariates enhances predictive modelling of STN in rice fields. • STN content variability is closely linked to landform conditions. • The integration of RS indices and environmental covariates, including landform, with the RF algorithm achieves an R² of 0.94 and RMSE of 0.05. • Geo-AI techniques show great potential for predicting STN content. The absence of high-resolution soil total nitrogen (STN) is an obstacle to the sustainable management and degradation assessment of Indonesian rice fields. Recently, advanced Geospatial-Artificial Intelligence (Geo-AI) techniques such as the random forest (RF) algorithm have been developed to increase the accuracy and spatial representativeness of STN prediction. However, critical challenges remain in using datasets from multiple locations. The Geo-AI techniques in this study utilizes soil and vegetation indices data from Sentinel-2A MSI imagery, integrated with various environmental covariates such as topographic factors (aspect, curvature, elevation, slope, and RDLS), soil properties (fractions of sand, silt, clay, bulk density, and pH), climate attributes (precipitation and temperature), and field-measured STN content to build predictive model. STN content data were obtained for the top 20 cm soil from a total of 318 sampling points across all landforms—alluvial, karst, and volcanic—of the Malang Regency in East Java, Indonesia. The results point to the differing importance of the predictors, with most remote sensing indices contributing significantly to the models. The performance of the machine learning-based prediction was also excellent, with an R² of 0.94 and a root mean square error (RMSE) of 0.05. The results demonstrate the promising potential of Geo-AI techniques for predicting STN content, offering a sophisticated solution to improve rice production through more precise fertilization. Our analysis provides a solid foundation for the future development of large-scale STN prediction and highlights the potential benefits of incorporating diverse types of predictors to enhance model performance.