SMFS-RF: a knowledge-guided machine-learning method for crop phenology extraction from fine-resolution vegetation index data

Phenological stages are critical indicators for assessing crop productivity and understanding ecosystem responses to climate change. However, existing remote-sensing approaches show reduced accuracy and poor transfer when applied to heterogeneous cropping regions. Here, we propose SMFS-RF, a knowledge-guided machine learning framework that integrates Shape Model Fitting by Separate Phenological Stage (SMF-S) with Random Forest (RF) regression, to estimate eight key rice phenological stages from fine-resolution Normalized Difference Vegetation Index (NDVI) time series. In this framework, SMF-S-derived phenological dates serve as training samples, enabling RF to capture the nonlinear relationships between vegetation dynamics and developmental timing. Validation against phenological records from 226 agricultural meteorological stations across China demonstrates that SMFS-RF effectively captured spatial variability and achieved moderate temporal errors. The method yielded mean spatial RMSEs of 8.69–10.41 days and mean temporal RMSEs of 5.52–6.36 days across transplanting, regreening, tillering, stem elongation, booting, heading, milk ripening, and maturity stages. The SMFS-RF method overcomes major constraints in remote sensing–based phenology monitoring, including the limited range of extractable phenological stages and low accuracy across heterogeneous cropping regions. Overall, these findings highlight the potential of SMFS-RF to advance large-scale, fine-resolution phenology monitoring, offering a robust tool for precision agriculture and climate adaptation research.