• Solving heterogeneity issue by Model-level reconstruction instead of at data-level. • Introducing Rice Growth Changing process into Model Sequences reconstruction. • Unsupervised Long-Term Rice Mapping via Rice Cycle-Constrained Model Sequence. Large-scale rice mapping is crucial for agricultural management and climate change studies but is hampered by challenges such as the spatiotemporal heterogeneity of high-quality data, sample scarcity, and limited model transferability. To overcome these barriers, this study presents a pioneering framework that shifts the paradigm from traditional reconstruction of spectral data to the reconstruction of time-continuous classification model sequences guided by phenological knowledge constraints. Unlike traditional methods that focus on reconstructing raw spectral data, our approach models the continuous evolution of parameters of classification models throughout the growing season. By training discrete Logistic Regression (LR) models regularized by their temporal predecessors, we ensure biophysical realism and link model dynamics directly to rice canopy development. This sparse sequence is then interpolated into a continuous trajectory also under the constraints of rice canopy development. The model-space interpolation enables optimal handling of observations from any arbitrary date and facilitates robust, sample-free model transfer. Long-term validation (1984–2024) and cross-regional rice mapping applications across China, Japan, Italy, and the USA demonstrate that our method achieves high classification accuracy (OA > 0.95) and robust generalization (relative errors < 10%). This study demonstrates that reconstructing model parameter sequences—rather than the raw data itself—is a feasible and powerful strategy for large-scale rice mapping across diverse climatic zones. Beyond rice mapping, the proposed strategy of reconstructing model parameter sequences may offer a transformative perspective for the broader Earth observation community. It demonstrates that when direct high-quality observations are intermittent, modeling the continuous evolution of the classification logic itself under phenological or geographic knowledge is a powerful alternative to data imputation. This conceptual shift provides a scalable solution for monitoring various highly dynamic Earth surface processes, such as forest phenology, and land degradation, where temporal consistency and physical interpretability are paramount.
Zhao et al. (Sat,) studied this question.