Optimal segmentation scale selection and evaluation of cultivated land objects based on high-resolution remote sensing images with spectral and texture features

As the remote sensing technology develops, there are increasingly more kinds of remote sensing images available from different sensors. High-resolution remote sensing images are widely used in the detection of land cover/land change due to their plenty of characteristics of a specific feature in terms of spectrum, shape, and texture. Current studies regarding cultivated land resources that are the material basis for the human beings to survive and develop focus on the method to accurately obtain the quantity of cultivated land in a region and understand the conditions and the trend of change of the cultivated land. Pixel-based method and object-oriented method are the main methods to extract cultivated land in remote sensing field. Pixel-based method ignores high-level image information, while object-oriented method takes the image spot after image segmentation as the basic unit of information extraction, which can make full use of spectral features, spatial features, semantic features, and contextual features. Image segmentation is a key step of object-oriented method; the core problem is how to obtain the optimal segmentation scale. Traditional methods for determining the optimal segmentation scale of features (such as the homogeneity-heterogeneity method, the maximum area method, and the mean variance method), in which only the spectral and geometrical characteristics are considered, while the textural characteristics are neglected. Based on this, the Quickbird and unmanned aerial vehicle (UAV) images obtained in Xiyu Village, Pengzhou City, Sichuan Province, China, were selected as experimental objects, and the texture mean and spectral grayscale mean method (MANC method based on GLCM), which comprehensively considered the spectrum, shape, and texture features, was proposed to calculate the optimal segmentation scale of cultivated land in the study area. The error segment index (ESI) and centroids distance index (CDI) were adopted to evaluate image segmentation quality based on the method of area and position differences. The experimental results show that the MANC method based on GLCM can obtain higher segmentation precision than the traditional methods, and the segmentation results are in good agreement with the cultivated land boundary obtained by visual interpretation.

Automated summary

A new method was proposed in the study, which comprehensively considered spectral, shape, and texture features for calculating the optimal segmentation scale of cultivated land. Experimental results showed that this method achieved higher segmentation precision compared to traditional methods and aligned well with the results obtained by visual interpretation.