Estimating crop leaf area index and chlorophyll content using a deep learning-based hyperspectral analysis method

Abstract

The crop leaf area index (LAI) and leaf chlorophyll content (LCC) are essential indicators that reflect crop growth status, and their accurate estimation is helpful for agricultural management decision-making. Traditional hyperspectral estimation methods for crop LAI and LCC from canopy spectra face challenges due to intricate soil backgrounds, canopy structural environments, and varying observational conditions. This paper proposes an LAI and LCC estimation method based on hyperspectral remote sensing, a radiative transfer model (RTM), and a leaf area index and leaf chlorophyll content deep learning network (LACNet). The LACNet architecture was developed utilizing deep and shallow feature fusion, blocks, and a hyperspectral-to-image transform (HIT) concept, aiming to improve LAI and LCC estimation. We used a field-based spectrometer to collect a dataset comprising 1,234 spectral measurements across five crop types: wheat, maize, potato, rice, and soybean. We used properties optique spectrales des feuilles and scattering by arbitrarily inclined leaves (PROSAIL) to generate a simulated spectra dataset (n = 145,152) representing complex farmland conditions for the five abovementioned crops, considering the variations in soil type, soil moisture, LAI, LCC, etc. The LACNet deep learning model sequentially uses RTM simulated and field-based spectra datasets for training, achieving higher universality and validation accuracy. We also analyzed the LACNet model’s interpretability for LAI and LCC estimation based on the gradient-weighted class activation mapping theory. From our research, we drew the following conclusions: (1) The shallow network features are sensitive to the LAI and LCC in the entire visible band, consistent with our correlation analysis results, while the deep network sensitive areas are mainly concentrated in the RE + VIS and RE + NIR regions of the HIT images. (2) The LACNet deep learning model (LAI: coefficient of determination (R²) = 0.770, root mean square error (RMSE) = 0.968 m²/m²; LCC: R² = 0.765, RMSE = 4.547 Dualex readings) can provide higher crop LAI and LCC estimation accuracy than widely used spectral feature and statistical regression methods (LCC: R² = 0.491–0.620, RMSE = 5.804–6.691 Dualex readings; LAI: R² = 0.476–0.716, RMSE = 1.089–1.482 m²/m²). The results of this study highlight the potential of the LACNet deep learning model as an effective and robust tool for accurately estimating crop LAI and LCC.