Gradient boosting and deep neural network-based forests quality assessment method using spatial-frequency domain spectral indices

High-quality forests enhance air purification, mitigate urban heat accumulation, and provide suitable natural recreational spaces for the public. In this study, we applied the Enhanced Spatial and Temporal Adaptive Reflectance Fusion Model (ESTARFM) to fuse Landsat-8 and MODIS data, generating vegetation indices with high spatial and temporal resolution. The Relative Vegetation Index (RVI) and its Fourier-transform harmonic components were selected as key predictors, and a remote sensing-based index system for Forests Quality (FLQ) was constructed by incorporating climatic and topographic factors. Regression models for FLQ estimation were developed using Deep Neural Networks (DNN), Random Forest (RF), and CatBoost, with DNN showing the highest predictive accuracy. On the training set, the DNN model achieved an R2 of 0.94, an RMSE of 3.65, and an RPD2 of 2.53. On the validation set, it attained an R2 of 0.77, an RMSE of 3.17, and an RPD2 of 3.40, significantly outperforming both RF and CatBoost in robustness and generalization capability. Using the optimal DNN model to map FLQ in Suixi County, the validation R2 was 0.75, RMSE was 3.51, and RPD2 was 3.39. Compared with an evaluation system based solely on a single-date vegetation index, our approach – which incorporates spatial-frequency domain spectral indicators – improved accuracy, reducing RMSE by 0.84 and increasing RPD2 by 0.6, thereby demonstrating the reliability of the proposed method.