A Novel Approach to Improve Leaf Chlorophyll Content Estimation with Resnet-18 and Feature Optimization Using Hyperspectral Measurements

Abstract

Accurate and robust estimation of leaf chlorophyll content (LCC) is very important to better known the process of material and energy exchange between plants and environment. Compared with traditional remotely sensing methods, convolutional neural network (CNN) excels in representation learning ability, which makes it suitable for hyperspectral data processing. Abundant researches thus have been focused on agronomic parameters retrieval using different CNN frameworks. Nevertheless, few reports have paid attentions to the problems, i.e., limited measured data, hyperspectral redundancy and model generalization issue, when concerning CNN models for parameters estimation. Therefore, the present study tried to propose a method for accurately access plant LCC using a ResNet-18 model with the ANGERS data. Effects of synthetic data size expansion, employing a gaussian process regression (GPR) model for simulation, input feature optimization, combing using different spectral indices with a competitive adaptive reweighted sampling (CARS) algorithm, and model generalization issue utilizing transfer learning (TL) method, on its performance for LCC retrieval were thus comprehensively investigated. Results showed that the established multi-output GPR model, based on the ANGERS data, exhibited good accuracies (a RMSE value of 0.04, a RRMSE value of 27.01%, and a NRMSE value of 12.77%) for 400 - 1000 nm reflectance data simulation. ResNet-18 training with 800 simulated reflectance (400 - 1000 nm) and partial ANGERS data exhibited overall best results, with a R2 value of 0.89, a RMSE value of 6.98 μg/cm2, a RPD value of 3.70, for LCC retrieval using remanent ANGERS data. The estimation accuracies obviously increased when nine spectral indexes, selected from CARS algorithm, were used as model input for running ResNet-18 model (R2 = 0.96, RMSE = 4.65 μg/cm2, RPD = 4.81). In addition, coupling transfer learning with ResNet-18 improved model convergence rate, and TL-ResNet-18 exhibited accurate results for LCC estimation (R2 = 0.94, RMSE = 5.14 μg/cm2, RPD = 4.65). These results suggest that the proposed method with ResNet-18 could be accurately used for LCC retrieval.

Keywords: Convolutional neural network, gaussian process regression, spectral index, competitive adaptive reweighted sampling, transfer learning, leaf chlorophyll content