Impervious surfaces (ISs) are significant indicators of the overall environmental health of watersheds. Monitoring the IS spatial patterns of watersheds is helpful for improving water management and regional pollution assessment. A random forest regression (RFR) approach was developed to estimate the subpixel IS percentage (ISP) from Chinese GF-5 hyperspectral imagery in 30 m pixels in the Nansi Lake Basin. Initially, object-based image analysis and overlay analysis were adopted to generate a very high-resolution ISP reference dataset using GF-1 multispectral imagery. Subsequently, an RFR-based ISP prediction model was established and evaluated using training and validation samples, respectively. The experimental results demonstrated that the proposed method shows a good performance with a root mean square error of 0.17, a mean absolute error of 0.15, and an R2 value of 0.89. Meanwhile, both band 10 and band 53 of GF-5 made the highest contributions to ISP modeling. Further comparison with other machine learning algorithms revealed that the RFR-based ISP model outperforms support vector machine regression and partial least squares regression, demonstrating the potential usability of the proposed method in ISP estimation from hyperspectral imagery.
Read full abstract