Abstract:

The supervised classification algorithm highly depends on the training samples so that it faces big challenge in the monitoring and analysis of natural resources for the inaccurate data caused by remote sensing imaging radiation deviation. Normally, for the same category on land surface, there are always some spectrum differences in multi-temporal images and different spatial locations on the images for the different conditions of atmosphere, radiation, illumination and acquisition geometries. This paper presents a method of extracting photovoltaic (PV) power plants based on multi invariant feature combination using Landsat 8 OLI remote sensing imagery. Based on the analysis of spectral uncertainty of the ground objects, this method combines the transformed spectral characteristics, spectral curves, texture and band ratio to improve the generalization performance of the ground object extraction algorithm by utilizing the complementation of multiple features. Firstly, the RGB band of the remote sensing image is converted into the HLS format, and the Fourier transform texture feature is calculated according to the brightness dimension L, and the spectral angle, band ratio, hue and saturation are added as the spectral signatures, which is not sensitive to the changes in pixel brightness values. Finally, we choose One-class support vector machine (OCSVM) as a classifier. The experimental results show that our method not only can effectively overcome the differences in pixel brightness values of the objects but also has a better performance on extracting the complex PV power plants.

Key words: spectral uncertainty, generalization performance, multi invariant feature, photovoltaic power plants, one-class support vector machine, dataset shift