Object-oriented crop classification for GF-6 WFV remote sensing images based on Convolutional Neural Network

Abstract

The GF-6 WFV image is the first remote sensing image of the 8-band multi-spectral satellite with medium and high resolution in China. 4 spectrum bands including two red-edge bands are added to the image based on the conventional red, green, blue and near-infrared band. As a vegetation sensitive band, the red-edge band is one of the methods used for crop classification and identification in remote sensing images. Research on the impact of GF-6 WFV image and its red-edge bands on crop classification is urgently needed. This research uses GF-6 WFV images as the data source. The main work is (1) proposes a convolutional neural network (RE-CNN) remote sensing image crop classification model suitable for GF-6 WFV red-edge bands; (2) conducts crop classification research about GF-6 WFV imagery and its red-edge bands and evaluates effectiveness of red-edge bands due to the lack of relevant research, (3) uses the strategy of combining object-oriented and deep learning for crop classification. The core idea of this research: multi-scale segmentation is used in order to avoid the influence of salt and pepper phenomenon on image classification, and image segmentation is completed by selecting the best segmentation parameters with ESP tools and ROC-LV. Object-oriented classification by CART decision tree can extract vegetation area while eliminating salt and pepper noise, and convert it into input data of convolutional neural network. The network structure of Inception was introduced to extract the multi-scale features of the image and then a convolutional neural network model (RE-CNN) for GF-6 WFV imagery was constructed for crop classification. A control experimental group with or without red-edge bands was set up and the RE-CNN model was used for crop classification and accuracy verification. The effect of the newly added red-edge bands on crop classification is studied, and the effectiveness and sensitivity of the red-edge band are evaluated in crop classification by GF-6 WFV imagery. The experimental results of this study show that: (1) Object-oriented CART decision tree classification effectively eliminates salt and pepper noise in vegetation area extraction, and the classification strategy of combining with deep learning achieve better classification results in remote sensing image crop classification. (2) The RE-CNN model proposed in this paper can be used for GF-6 WFV remote sensing image crop classification. The classification accuracy of the experimental group in the group of red-edge bands is as high as 94.38%, and the Kappa coefficient is 0.92. (3) The newly added red-edge bands in GF-6 WFV images can effectively improve the crop classification accuracy of remote sensing image. Compared with the group without red-edge bands, the classification accuracy is increased by 2.83%, which verifies the effectiveness and sensitivity that the newly-added red-edge bands in GF-6 WFV images improves the classification accuracy. Moreover, it provides a reference for the research of GF-6 WFV image and its red-edge band for crop classification.