关键词: 等效水厚度, 随机森林, PRO4SAIL模型, 水分指数, 机器学习, 路域植被, 湖南醴潭高速

Abstract:

Vegetation water content is an important evaluation index for vegetation health monitoring. The Equivalent Water Thickness (EWT) of vegetation is of great significance for the monitoring and evaluation of the ecological conditions in the road area, it could provide a guideline in road area environment management. Taking the Litan highway in Hunan Provinces as an example, this research used field data of canopy reflectance and equivalent water thickness of vegetation on the ground, and simulated reflectance and simulated equivalent water thickness established by PRO4SAIL. In total, 12 kinds of water indices were established by using the simulated reflectance of the PRO4SAIL canopy model and the ground measured reflectance. The random forest algorithm (RF) was introduced to analyze the importance of the 12 water indices and equivalent water thickness. We determined the ordination between water indices and equivalent water thickness as well as the optimal number of input water index in the equivalent water thickness estimation model by using the adjusted coefficient of determination. Based on the selected water index, the water index and equivalent water thickness were calculated by the PRO4SAIL simulation reflectance as the training set. Three equivalent water thickness estimation models were constructed: Random Forest Coupled Partial Least Squares (RF-PLS), Random Forest Coupled Support Vector Machine (RF-SVM) model, and Random Forest coupled Genetic Algorithm to optimize the Support Vector Machine (RF-GA-SVM) model. The applicability of 12 water indices in the estmation of equivalent water thickness in road-area of vegetation was also analyzed. The accuracy of the model was validated by measured equivalent water thickness on the ground. The experimental results show: (1) The adjusted determination coefficient of RF-SVM model was the highest, established by Normalized Difference Water Index (NDWI), Normalized Multi-band Drought Index (NMDI), Simple Ratio Water Index (SRWI), Simple Ratio (SR), Normalized Difference Infrared Index (NDII), Water Index (WI), Dattwater Index (DWI), Moisture Stress Index (MSI) and Soil Adjusted Vegetation Index (SAVI), with the determination coefficient of verification set reaching 0.8877. (2) The RF-PLS and RF-GA-SVM models with the four water indices of NDWI, NMDI, SRWI, and SR had the highest adjusted determination coefficient, with the validation set's determination coefficients reaching 0.8053 and 0.8952, respectively. (3) Among them, the RF-GA-SVM model was the best for estimating equivalent water thickness, which met the requirements of vegetation equivalent water thickness monitoring in road area. Our findings provide an effective and accurate method for the estimation of equivalent water thickness, and provide support for road area environment monitoring based on hyper-spectral remote sensing.

Key words: Equivalent water thickness, random forest, PRO4SAILmodel, water index, machine learning, vegetation of road area, Litan highway in Hunan Province