Research On Multi-scale Texture Features And Classification For Main Fruit Trees Species In Southern Xinjiang Basin

Xinjiang Tarim Basin has become the dominant region of apples, walnuts, dates, pears and other fruit characteristics industry with its vast geographical space and unique climatic conditions. So far, the characteristic fruit area has more than 2000 acres. Characteristic fruit industry has become a new growth pole to promote regional economic development and an important way to increase farmers’ income in Xinjiang. However, in the process of fruit industry scale and industrialization, information construction of fruit industry lags seriously. The traditional fruit resources investigation more depend on ground-based survey, not only time-consuming, labor-intensive, and has long life cycle and low efficiency. Therefore,accelerating the Xinjiang fruit industry information construction, quickly and accurately grasping the fruit resources layout, scale and other basic information has become an urgent demand to the rapid and healthy development in Xinjiang fruit industry. In this process, the remote sensing identification of fruit tree species is the core content of this work, has important actuality significance in the sustainable management of forest fruit industry.We draw the following main conclusions through research conducted to texture extraction at different resolutions,dimension reduction of texture features and high spatial resolution remote sensing identification of main fruit tree species in southern Xinjiang Basin.Textures differences in four kinds of fruit trees had the first-increase and follow- decrease trend as the window size changes under different resolutions(2m, 8m, 16m).Distinguishing four kinds of fruit trees suitable window are: 19×19 to 27×27,17×17 to 23 ×23,11×11 to 17×17 windows under 2m, 8m, 16 m resolution.The textures distinguishing four kinds of fruit trees are:variance, homogeneity, dissimilarity usefully.Four kinds of fruit trees species has the same separability trend under different resolutions(2m, 8m,16m). Dates has a steady trend in separability with the change of window, and the other three kinds of fruit trees had the first-increase and follow- decrease trend as the window size changes.Distinguishing four kinds of fruit trees suitable window are: 23×23,21×21,13 ×13 windows under 2m, 8m, 16 m resolution. Best texture combining as follow: variance, dissimilarity, correlation; variance, homogeneity, contrast;variance, contrast, dissimilarity under 23×23,21×21,13×13 windows.The classification accuracy based spectrum and texture features improved greatly than that based spectral characteristics and improved 15.45%、5.00%、20.12% from 58.32%、62.46%、39.14% to 73.77%、67.46%、59.26% at different resolutions(2m, 8m, 16m) separately.View classification results on a single species, classification accuracy of the dates is much higher than the other three kinds of fruit trees. apples,pears,walnuts had lower classification accuracy. According to classification results under different resolution images,support vector machines had a better classification results, the classification accuracy is higher than the mahalanobis distance, maximum likelihood, neural network classification methods.Classification accuracy of four kinds of forest tree species increase with the increasing of resolution under different resolutions(2m, 8m, 16m). The classification accuracy were as follow : 76.91%,67.46%, 61.23% under 2m, 8m, 16 m resolution. So 2m resolution remote sensing image is the best remote sensing data sources to identify the fruit tree species.