Exploring the influence of coupling diverse spectral resolutions and feature selection approaches on the estimate of equivalent water thickness in fruit tree leaves

Accurate and effective acquisition of physiological parameter information from fruit trees plays an important role in the fine management of orchards. The extensive use of hyperspectral remote sensing technology in the field of agriculture provides an effective means for the scientific management of orchards. Accurate estimation of leaf equivalent water thickness (EWT) is crucial for applying hyperspectral remote sensing technology to assess the growth status of fruit trees. However, current research lacks a focus on analyzing the spectral resolution within the feature set to further evaluate its robustness and sensitivity to the inversion of the EWT of fruit tree leaves, and to develop an efficient inversion model specifically for fruit tree leaf EWT. This study aims to investigate the impact of coupling different spectral resolutions with feature selection methods on predicting EWT at the leaf level. Specifically, the leaf spectra of walnut, apricot, and jujube trees were measured using an ASD FieldSpec4 spectrometer in the field in Xinjiang, China’s northwest region, while the EWT of the leaves was determined in the laboratory. The Prospect-D model was utilized to generate a simulated spectral dataset, which was subsequently resampled at nine different resolutions, ranging from 1 to 100 nm. The selection of feature sets in this study was based on this simulated dataset. Secondly, various spectral indices, including three-band spectral indices (TBI1, TBI2, and TBI3) and two-band spectral indices (DSI, NDSI, and RSI), were combined with feature band selection methods, namely Competitive Adaptive Reweighted Sampling (CARS) and Successive Projections Algorithm (SPA), as well as the downscaling method t-distributed stochastic neighbor embedding (t-SNE), to couple different spectral resolutions with feature selection strategies. Finally, WOA-RF models were employed to evaluate model performance. The study aimed to identify the optimal model by validating the model for each fruit tree sample set. The findings of this study revealed the following: (1) The EWT of walnut leaf samples was significantly higher compared to that of apricot and jujube leaf samples, with a Coefficient of Variation (CV) less than that of apricot and jujube datasets (CV < 25.7%). (2) The Random Forest (RF) model optimized using the whale optimization algorithm (WOA) demonstrated superior estimation performance compared to the original model. Specifically, it exhibited higher values for R-squared (R²), relative percent deviation (RPD), and the ratio of performance to interquartile distance (RPIQ), along with a lower mean absolute error (MAE). (3) Validation using a single sample set indicated that, for the middle resolution, the optimal combination based on the WOA-RF model was 20 nm-CARS, achieving R² > 0.881, RPD > 2.021, RPIQ > 1.6418, and MAE < 0.00097. Notably, this model requires only lower spectral resolution and fewer band combinations to achieve higher accuracy in estimating EWT. These results can serve as a reference for monitoring and applying hyperspectral remote sensing to assess the water status of various fruit tree leaves at a large scale.