Abstract:

Objective: This study aims to evaluate the feasibility of estimating canopy chlorophyll content (CCC) of Populus spp. (poplar) using UAV-based hyperspectral imagery in combination with the PROSAIL radiative transfer model. Specifically, it focuses on determining the resolution of hyperspectral imagery for CCC inversion within this methodological framework, identifying the optimal band set, developing a high-accuracy inversion model, and quantifying the contribution of individual bands to the model performance. Method: The National Poplar Germplasm Base in Shishou City, Hubei Province was selected as the study area. UAV-based hyperspectral and LiDAR data were acquired at four spatial resolutions (0.05, 0.10, 0.20, and 0.50 m). Following the acquisition of remote sensing data, the measured CCC of 120 poplar trees with different genotypes was collected. Individual tree segmentation was performed using point cloud data to delineate tree crown areas, from which the mean reflectance values of hyperspectral images were extracted. Reflectance data within the 400–900 nm spectral range were retained as the original band set for ground-truth validation. Meanwhile, the PROSAIL model was calibrated using ground measurements, and 100000 simulated samples were generated for model training. A hybrid two-stage feature selection framework was proposed by integrating the Shapley additive explanation (SHAP) algorithm with a long short-term memory (LSTM) regression model for feature ranking. The feature space was progressively reduced with a step size of 10, reconstructing and reordering features at each iteration. After comparing the validation accuracy, the feature subset with the highest validation accuracy was selected as the initial set. Subsequently, the sequential backward selection (SBS) was used to further identify the optimal subset and construct CCC estimation models for different spatial scales. Finally, model adaptability across spatial resolutions and the contribution of each wavelength were systematically analyzed. Result: The results indicated that the accuracy of poplar CCC estimation declined with decreasing spatial resolution, with Scale 1 (0.05 m) yielding the highest precision. Comparative experiments indicated that the integration of the SHAP algorithm within the progressive two-stage feature selection framework enhanced the adaptability of regression models, and the re-ranking strategy further improved the robustness of feature prioritization. The model achieved its optimal validation accuracy when the number of spectral bands was reduced to 50, whereas the accuracy of models without re-ranking decreased as the number of bands decreased. The SBS algorithm was used to search for 38 sensitive bands, predominantly distributed in the red and red-edge regions (660–780 nm). The three most influential bands were 895 nm, 891 nm, and 723 nm, contributing 4.44%, 4.03%, and 3.42% to the model, respectively. As spatial resolution decreased, the number of bands in the optimal feature subset increased, with most additional bands located in the blue and near-infrared regions. Overall, the Scale 1 model achieved the best inversion performance, with R², RMSE, and MAE values of 0.807, 7.087 μg·cm^–2, and 5.452 μg·cm^–2, respectively, based on ground validation data. Conclusion: The integration of the PROSAIL model with deep learning enables the retrieval of poplar canopy chlorophyll content (CCC) from UAV-based hyperspectral data, and high spatial resolution has a significant positive effect on model accuracy. In addition, the progressive two-stage feature selection framework proposed in this study incorporates the SHAP algorithm, a re-ranking mechanism, and the integration of embedded and wrapper algorithms, thereby achieving more robust feature ranking, higher retrieval accuracy, and a more efficient feature selection process.

Key words: Populus spp., canopy chlorophyll content, unmanned aerial vehicle hyperspectral imagery, radiative transfer model, feature selection