Transfer-learning-based approach for leaf chlorophyll content estimation of winter wheat from hyperspectral data

Abstract Leaf chlorophyll, as a key factor for carbon circulation in the ecosystem, is significant for the photosynthetic productivity estimation and crop growth monitoring in agricultural management. Hyperspectral remote sensing (RS) provides feasible solutions for obtaining crop leaf chlorophyll content (LCC) by the advantages of its repeated and high throughput observations. However, the data redundancy and the poor robustness of the inversion models are still major obstacles that prevent the widespread application of hyperspectral RS for crop LCC evaluation. For winter wheat LCC inversion from hyperspectral observations, this study described a novel hybrid method, which is based on the combination of amplitude- and shape- enhanced 2D correlation spectrum (2DCOS) and transfer learning. The innovative feature selection method, amplitude- and shape- enhanced 2DCOS, which originated from 2DCOS, additionally considered the relationships between external perturbations and hyperspectral amplitude and shape characteristics to enhance the dynamic spectrum response. To extract the representative LCC featured wavelengths, the amplitude- and shape- enhanced 2DCOS was conducted on the leaf optical PROperties SPECTra (PROSPECT) + Scattering from Arbitrarily Inclined Leaves (SAIL) (PROSAIL) simulated dataset, which covered most possible winter wheat canopy spectra. Nine wavelengths (i.e., 455, 545, 571, 615, 641, 662, 706, 728, and 756 nm) were then extracted as the sensitive wavelengths of LCC with the amplitude- and shape- enhanced 2DCOS. These wavelengths had specificity to LCC and showed good correlation with LCC from the aspect of photosynthesis mechanism, molecular structure, and optical properties. The transfer learning techniques based on the deep neural network was then introduced to transfer the knowledge learned from the PROSAIL simulated dataset to the inversion tasks of field measured LCC. Parts of the labeled samples in field observations were used to finetune the model pre-trained by the simulated dataset to improve the inversion accuracy of the winter wheat LCC in different field scenes, aiming to reduce the need for the field measured and labeled sample size. To further ascertain the universality, transferability and predictive ability of the proposed hybrid method, field samples collected from different locations at different phenological phases, including the jointing and heading stages in 2013, 2014, and 2018, were utilized as target tasks to validate the proposed hybrid method. Moreover, the LCC of winter wheat estimated with the proposed method was evaluated with the ground-based platform and the UAV-based platform to verify the model versatility for different monitoring platforms. Various validations demonstrated that the hybrid inversion method combining the amplitude- and shape- enhanced 2DCOS and the fine-tuned transfer learning model could effectively estimate winter wheat LCC with good accuracy and robustness, and can be extended to the detection and inversion of other key variables of crops.