Combination of feature selection methods and lightweight Transformer model for estimating the canopy water content of alpine shrub using spectral data

Abstract

Monitoring canopy water content in alpine scrub is crucial for hydrological and ecological studies in the Qilian Mountains. However, traditional measurement methods are time-consuming and inconvenient to implement due to the harsh climate and lack of transportation in alpine regions. In our study, a new method was proposed to estimate the canopy water content of alpine shrub using spectral data. The water content data for samples with alpine scrub canopy reflectance spectra in the range of 350–2500 nm was collected. Then, a lightweight Transformer regression model was built based on three different feature selection methods (the successive projections algorithm, genetic algorithm, and principal component analysis) to predict the canopy water content of shrubs. The proposed model was compared with three traditional machine learning models: random forest, support vector regression, and back-propagation neural network. The results reveal that the lightweight Transformer model demonstrates strong fitting and generalization capabilities in using only full-band spectra as input, which may be attributed to Transformer's own self-attention mechanism. From cross-validation, the RMSE and R2 of the model on the training (validation) set are 0.0201 (0.0233) and 0.968 (0.965), respectively. Furthermore, the performance of all regression models is enhanced after feature selection. On the validation set, PCA-Transformer has the smallest RMSE (0.0147) and largest R2 (0.985), and its running efficiency is improved by 81.69 % relative to the full-band spectra as input. In the additional test set, the predictive ability of the models remain stable. SPA-Transformer has the smallest RMSE (0.0111) and largest R2 (0.986). These suggest the excellent performance of lightweight Transformer on the task of estimating scrub water content and the ability of feature selection methods to optimize models. The results show that PCA-Transformer is an accurative and effictive tool to estimate plant water content, and offer a foundation for monitoring large-scale canopy water content in alpine regions using remote sensing technology.