Remote sensing information extraction of aquatic vegetation in Lake Taihu based on Random Forest Model

Abstract

Aquatic vegetation is a vital component of the ecosystem of Lake Taihu. Assessment of its distribution and abundance by using remote sensing techniques is important for the protection of Lake Taihu as the information serves as an excellent indicator of aquatic environmental quality. In this research, to extract spatial distribution of the different aquatic vegetation types in Lake Taihu, we took Landsat 8 multi-spectral images as the main data source and applied them with Random Forest Model on the basis of multiple characteristic variables, which were constructed by the method of spectral index and image transformation. The results show: (a) By analyzing and comparing statistics mean, standard deviation and variable coefficient obtained from different training samples, we found NDVI, NDWIF and SR were better characteristic variables for distinguishing open water and floating-leaf vegetation, submerged vegetation and emergent vegetation than others. (b) Under the condition of 1000 classification trees with 4 random variables in split node, the out-of-bag error of Random Forest Model was below 6%. The error of model was mainly affected by SR, MNDVI and NDVI. (c) According to validation analysis, the overall accuracy classification of image based on Random Forest Model was about 88.56% with a high Kappa index of 0.88. The total area of aquatic vegetation was about 306.0 km2 in July of 2014, in which floating-leaf vegetation and emergent vegetation accounted for 84.9% and were mainly distributed in the eastern and southern parts of Lake Taihu.