Identification of Cotton Using Random Forest Based on Wide-Band Imaging Spectrometer Data of Tiangong-2

Crop identification needs the relatively high spatial, temporal and spectral resolution due to the similarity spectral or spatial features of crops. Tiangong-2 (TG-2) Space Laboratory which payloads the Wide-band Imaging Spectrometer (WIS) can provide 14 bands covering from visible to near-infrared (VIS-NIR) wavelength with spatial resolution of 100 m and a repeat circle of 3 days. The high resolution of spatial-spectral-temporal is of promising advantage in precision crops identification. Random forest (RF) is a powerful machine learning classifier and has a good potential in the remote sensing classification. In this paper, TG-2 WIS image was used to extract the cotton in Xinjiang area with field experimental validation. RF along with support vector machine (SVM) and maximum likelihood classification (MLC) were utilized and compared in the cotton classification study. Three types of features of principal component, texture feature and spectral index are extracted for the three classifiers. A total of 29 features are used as input bands for MLC, SVM and RF. The results showed that RF was superior to the other methods. The overall accuracy and Kappa coefficients of RF are 91.2% and 0.871, respectively. This paper proves that the RF classification method has a good effect in the vegetation extraction of TG-2 data and positive significance for expanding the application of TG-2 data.