PMTFIM: Integrating machine learning with nutrient balance theory to estimate multi-stage paddy fertilization information at field scale over large regions

Housheng Wang, Xiang Gao, Wei Jiang, Xuerong Lang, Xian Hu, Meihua Qiu, Qiankun Guo, Yonghong Liang, Xuelei Wang, Yue Mu, Rui Ren, Ganghua Li, Hengbiao Zheng, Yanfeng Ding, Xiaosan Jiang,
PMTFIM: Integrating machine learning with nutrient balance theory to estimate multi-stage paddy fertilization information at field scale over large regions,
ISPRS Journal of Photogrammetry and Remote Sensing,
Volume 230,
2025,
Pages 693-715,
ISSN 0924-2716,
https://doi.org/10.1016/j.isprsjprs.2025.10.006.
(https://www.sciencedirect.com/science/article/pii/S0924271625003946)
Abstract: Accurate estimation of multi-stage field-level paddy fertilization information over large regions provides crucial support for optimizing fertilization management and evaluating greenhouse gases emission and agricultural non-point source pollution risks. Most existing models have utilized survey data from the government statistical agencies or international databases such as FAOSTAT, to estimate annual fertilization information, ignoring the different phenological growth stages of fertilization during the growing season. Although some studies have used LAI to estimate multi-stage fertilization information based on remote sensing, they utilized statistical models, leading to significant uncertainties. In this study, we present a precise tracing multi-stage paddy fertilization information model (PMTFIM), a novel remote sensing-driven framework that integrates Gaussian Process Regression-based LAI time series, phenological modeling, and machine learning coupled with nutrient balance theory to estimate multi-stage paddy fertilization information at the field scale across fragmented agricultural regions. Firstly, we obtained a daily 10-m LAI dataset by using the Gaussian Process Regression method and then integrated a double logistic function to generate training phenological data. Therefore, we estimated the fertilization dates based on rice phenology by coupling the optimal machine learning models. The fertilization amounts models were developed by using optimal machine learning, which accounts for interactions between fertilization, meteorological conditions, soil properties, and LAI dynamics based on nutrient balance. We determined that the random forest model is the optimal model among multiple machine learning models. The results demonstrated that PMTFIM captured the heterogeneity in fertilization information in fragmented paddy fields. The overall prediction accuracy based on training and test datasets (dates: R2 > 0.95, RMSE < 5 days; amounts: R2 > 0.80, RMSE < 14 kg/ha) has improved compared to existing statistical models. The PMTFIM achieved reliable accuracy at multiple growth stages with R2 ranging 0.53–0.80 for fertilization dates and amounts at field scale and in independent sub-regions evaluated on test dataset, while maintaining high overall accuracy across the entire growing season, with R2 of 0.80–0.96. Our proposed method has great potential for estimating multi-stage field-level crops fertilization information over large regions, especially in areas with fragmented fields.
Keywords: Remote sensing-based fertilization modeling; High resolution; Machine learning; Spatially-explicit nutrient management; Paddy fields