Estimation of photosynthetic parameters from hyperspectral images using optimal deep learning architecture

Abstract

The maximum carboxylation rate (Vcmax) and maximum electron transport rate (Jmax) of leaves are crucial for comprehending carbon cycling in farmland. Nevertheless, estimating these photosynthetic parameters precisely and rapidly faces a considerable challenge. This study designed an optimal deep learning architecture that accurately extracts photosynthetic parameters from hyperspectral images and evaluated its stability across different crop species. Photosynthetic parameter models jointly driven by hyperspectral images and deep learning were compared with models jointly driven by one-dimensional reflectance and simple machine learning methods, as well as chlorophyll driven models. The proposed optimal deep learning architecture incorporated spatial attention and prior knowledge of spectral indices calculations. Our results demonstrated that the hyperspectral images and deep learning jointly driven models outperformed traditional models. Notably, incorporating the module of spatial attention and spectral indices calculation networks achieved the best estimations for Vcmax25 (R2: 0.86, RMSE: 10.18 μmol m−2 s−1) and Jmax25 (R2: 0.83, RMSE: 24.27 μmol m−2 s−1). In contrast, the performance of the one-dimension reflectance driven models deteriorated (R2: 0.43–0.58, RMSE: 19.71–39.99 μmol m−2 s−1). Moreover, the best architecture was interpreted. Weight analysis revealed that the hyperspectral information on the middle part of the leaf contributed most to photosynthetic parameters. Feature map analysis indicated that the spectral indices calculation module utilized the information of the visible light spectrum. Even though the migration of models across various crop species may lead to a slight degradation, the model performance remained satisfactory after fine-tuning. The proposed deep learning models, which used 3D hyperspectral images for estimation of photosynthetic parameters, outperformed the models jointly driven by 1D reflectance data and conventional machine learning algorithms. The results highlighted the significance of spatial information from hyperspectral images and prior knowledge through spectral indices calculations. Moreover, the stability and effectiveness of the proposed architecture remained excellent across different species. This study presents advanced and highly effective deep learning techniques for evaluating the photosynthetic capacity of crop leaves and modeling the carbon cycle.