Construction and Scaling of a Combined Spectral Index-Based Maturity Estimation Model for Cold-Region Japonica Rice

Precise timing of rice harvesting is critical for ensuring grain yield and quality. Traditional manual evaluation methods are highly subjective and time-consuming, highlighting the critical demand for rapid, non-destructive approaches to estimate rice maturity. This study focused on cold-region japonica rice grown in Heilongjiang Province, aiming to develop and validate dual-scale (pot and field) maturity estimation models. For model development, canopy spectral data were collected using two complementary acquisition tools: a ground-based active sensor (CGMD402) and UAV-borne multispectral imagery. Four modeling algorithms—Linear Regression (LR), Decision Tree (DT), Random Forest (RF), and Support Vector Machine (SVM)—were utilized, with input variables comprising single spectral indices (Normalized Difference Vegetation Index, NDVI; Ratio Vegetation Index, RVI) and composite spectral indices (Normalized Difference Maturity Ratio Vegetation Index, NDMRVI; Normalized Difference Pigment Ratio Vegetation Index, NDPRVI). At the pot scale, composite spectral indices showed stronger correlations with rice maturity than single indices. Among the four algorithms, the DT model with combined NDVI + RVI input yielded the optimal comprehensive performance, with a coefficient of determination (R2) of 0.957, a root mean square error (RMSE) of 0.064, and a relative error (RE) of 4.8% in the test set. At the field scale, NDVI and RVI both exhibited strong negative correlations with maturity (Spearman’s correlation coefficients of −0.76 and −0.79, respectively). While the RF model performed best in the training set (R2 = 0.752), it was prone to overfitting; in contrast, Multiple Linear Regression (MLR, Ridge Regression) with NDVI + RVI combination demonstrated greater stability in the test set (R2 = 0.515, RMSE = 0.116). Notably, composite spectral indices consistently outperformed single indices across all modeling algorithms, but their accuracy was comparable to the optimal single index combination model. To tackle the challenge of scaling models from pot to field conditions, this research developed a “modeling–validation–evaluation–scaling” framework and a four-indicator combined judgment criterion (ΔR2–ΔRMSE–ΔRE–SF). Quantitative analysis showed that the optimal pot-scale model suffered significant accuracy loss during cross-scale transfer: ΔR2 = 0.447, ΔRMSE = 0.120, ΔRE = 22.84%, and Scale Transfer Factor (SF) = 2.875. A “regional calibration + residual correction” scheme was proposed, which is expected to reduce the transferred RMSE to below 0.12 and SF to 1.8–2.0. Overall, this research offers a reliable technical method for large-scale, non-destructive monitoring of rice maturity, which can facilitate data-driven precision harvesting decisions.