Assessment on the declining degree of farmland shelter forest in a desert oasis based on LiDAR and hyperspectrum imagery

Abstract

We examined the growth decline and health status of farmland protective forest belt (Populus alba var. pyramidalis and Populus simonii shelterbelts) in Ulanbuh Desert Oasis by using airborne hyperspectral and ground-based LiDAR to collect the hyperspectral images and point cloud data of the whole forest belt respectively. Through correlation analysis and stepwise regression analysis, we constructed the evaluation model of the decline degree of farmland protection forest with the spectral differential value, vegetation index, and forest structure parameters as independent variables and the tree canopy dead branch index of the field survey as dependent variables. We further tested the accuracy of the model. The results showed that the evaluation accuracy of the decline degree of P. alba var. pyramidalis and P. simonii by LiDAR method was better than that by hyperspectral method, and that the evaluation accuracy of the combined LiDAR and hyperspectral method was the highest. Using the LiDAR method, hyperspectral method, the combined method, the optimal model of P. alba var. pyramidalis was all light gradient boosting machine model, with the overall classification accuracy being 0.75, 0.68, 0.80, and Kappa coefficient being 0.58, 0.43, 0.66, respectively. The optimal model of P. simonii was random forest model, random forest model, and multilayer perceptron model, with the overall classification accuracy being 0.76, 0.62, 0.81, and Kappa coefficient being 0.60, 0.34, 0.71, respectively. This research method could accurately check and monitor the decline of plantations.

Keywords: LiDAR; degree of recession; hyperspectrum; random forest model.