A new feature extraction algorithm for measuring the spatial arrangement of texture Primitives: Distance coding diversity

Texture is a key spatial feature for object recognition in remote sensing images. Currently, most texture feature extraction methods mainly focus on the repeated patterns of texture primitives (the basic texture units) but rarely consider their spatial arrangement. Although some methods can capture the spatial arrangement of texture primitives to some extent, their principles and algorithms are complex and difficult to implement. In this study, we proposed a new statistical feature extraction method, called distance coding diversity (DCD), which can measure the spatial arrangement of texture primitives with a rotation invariant characteristic. The texture features extracted by DCD were widely tested on a large dataset at three different scales: 50 digital matrix samples at the pixel neighborhood scale, 2508 pairs of small block images of pre- and post-damaged buildings at the image object scale, and five large-scale images covering different main land cover types (cropland, buildings, plantations, natural vegetation and abandoned dumps) at the regional landscape scale, and the test results were compared with Entropy of Grey-level Co-occurrence Matrix (GLCM) and Short Run Emphasis (SRE) of Grey Level Run Length Matrix (GLRLM). DCD can effectively measure the spatial arrangement of texture primitives, and its sensitivity to spatial arrangement is 98%, which is more than 5 times that of SRE (with a sensitivity value of 18%). The identification accuracy for different land cover types based on DCD improved by 12.31% over Entropy, and 14.40% over SRE. Due to the simple algorithm and superior performance, DCD can be widely used for object recognition and land cover/use classification and has wide application prospects.