• A hybrid AGB inversion framework integrating 1D-CNN spectral features with machine learning was developed. • Band contributions and synergistic effects were quantified via a two-stage SHAP analysis to optimize vegetation indices. • The integration of deep spectral features and optimized vegetation indices enhanced estimation accuracy and spatiotemporal scalability. Aboveground fresh biomass (called AGB in the study) is a fundamental indicator of grassland productivity, utilization intensity, and ecosystem condition. Although satellite remote sensing has been widely used for large-scale AGB monitoring, further improvement is still needed in spectral information extraction and in exploiting synergistic relationships among multispectral bands. This study develops an inversion framework that integrates 1D-CNN–based spectral feature extraction with machine-learning regression, using extensive field measurements of fresh AGB across natural grasslands and Sentinel-2 multispectral imagery. The results indicated that: (1) the 1D-CNN effectively captured Deep Spectral Features from reflectance sequences; when these features were combined with optimized vegetation indices, the resulting Gaussian Process Regression (GPR) model outperformed the baseline model using only reflectance and conventional vegetation indices, increasing R 2 by 0.06 and achieving a 12.82% relative reduction in RMSE; (2) red, red-edge and shortwave-infrared bands exhibited high synergistic contributions, and constructing optimized vegetation indices through high-contribution band substitution yielded average improvements of 0.02 in R 2 across five machine-learning algorithms; (3) among all feature sets and regression models, the H 2 feature set—screened from reflectance, Deep Spectral Features and optimized vegetation indices—produced the best performance in the GPR model ( R 2 = 0.81, RMSE = 58.66 g·m⁻ 2 ); (4) spatial and temporal generalization analyses demonstrated good transferability of the 1D-CNN–enhanced models. Overall, this study provides a scalable and interpretable approach for estimating AGB across large natural grassland regions.
Gu et al. (Wed,) studied this question.