Open Access

ARTICLE

Improved Leaf Chlorophyll Content Estimation with Deep Learning and Feature Optimization Using Hyperspectral Measurements

Xianfeng Zhou^1,2,*, Ruiju Sun¹, Zhaojie Zhang¹, Yuanyuan Song¹, Lijiao Jin¹, Lin Yuan³

1 Zhejiang Zhengyuan Geomatics Co., Ltd., Huzhou, 313203, China
2 School of Artificial Intelligence, Hangzhou Dianzi University, Hangzhou, 310018, China
3 School of Computer Science and Technology, Zhejiang University of Water Resources and Electric Power, Hangzhou, 310018, China

* Corresponding Author: Xianfeng Zhou. Email:

Phyton-International Journal of Experimental Botany 2025, 94(2), 503-519. https://doi.org/10.32604/phyton.2025.060827

Received 11 November 2024; Accepted 24 January 2025; Issue published 06 March 2025

Abstract

An accurate and robust estimation of leaf chlorophyll content (LCC) is very important to better know the process of material and energy exchange between plants and the environment. Compared with traditional remote sensing methods, abundant research has made progress in agronomic parameter retrieval using different CNN frameworks. Nevertheless, limited reports have paid attention to the problems, i.e., limited measured data, hyperspectral redundancy, and model convergence issues, when concerning CNN models for parameter estimation. Therefore, the present study tried to analyze the effects of synthetic data size expansion employing a Gaussian process regression (GPR) model for simulation, input feature optimization using different spectral indices with a competitive adaptive reweighted sampling (CARS) algorithm, model convergence issue combining transfer learning (TL) method for accurate and robust estimation of plant LCC with a deep learning framework (i.e., ResNet-18) using the ANGERS data (a public dataset containing foliar biochemical parameters spectral data for various plant types). Results showed that ResNet-18 training using 800 simulated reflectances (400–1000 nm) and partial ANGERS data exhibited better results, with an R² value of 0.89, an RMSE value of 6.98 μg/cm², an RPD value of 3.70, for LCC retrieval using remanent ANGERS data, than models that using simulations with different amounts of data. The estimation accuracies obviously increased when nine spectral indexes, selected from the CARS algorithm, were used as model input for running the ResNet-18 model (R² = 0.96, RMSE = 4.65 μg/cm², RPD = 4.81). In addition, coupling transfer learning with ResNet-18 improved the 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 adding appropriate synthetic data, input features optimization, and transfer learning techniques could be effectively used for improved LCC retrieval with a ResNet-18 model.

---

Keywords

---