Extraction Of Greenhouse Using Texture Features Within High Spatial Resolution Images

Greenhouse plays a critical role in coping with difficulties of winter shortage of vegetables in northwestern China. It is of importance to extract the spatial extant of greenhouse and to understand their distribution and future trends. In this article, greenhouses in ALOS image were extracted by texture features from image objects and moving windows. The effects of texture variables, window size, direction and displacement and other parameters of greenhouse information extraction were discussed. Five texture variables selected by information entropy involved in the classification, to find strongest correlation between object and texture features by factor analysis method, multi-scale texture were constructed based on the correlation. The following conclusions are obtained in this study:1) Texture features from image objects and moving windows have different effects on the overall accuracy(OA) of greenhouse extraction. The OA of greenhouses extraction does not increased significantly by Object-based texture, sometimes reduced the OA, the enhanced range is-0.36%~1.39%, however window-based texture can significantly improve the extraction accuracy, the enhanced range is 4.66%~7.88%. For the different object, the contribution of two types of textures on the accuracy of classification is different, window-based texture feature can improve the classification accuracy of most objects, but Object-based texture can only improve the classification accuracy covered greenhouse, fallow land and orchard,and window-based texture increased the classification accuracy larger than Object-based texture. In short, window-based texture is better than Object-based texture in the overall accuracy and the classification accuracy of object.2) The information entropy method is an effective way to select the texture variables. The 5 texture variables selected by the information entropy method, can improve the classification accuracy compared with all the texture variables. Finally, the mean, homogeneity, contrast, entropy and correlation were selected to extract the information of greenhouses.3) The correlation of texture variables, the KMO and the Bartlett sphere test proved that it is necessary and feasible to use the factor analysis method to optimize the texture variables. The best texture parameters selected by the comprehensive score is difference selected by classification accuracy. The why is that: The best texture parameters selected by the classification accuracy were based on ALOS Image Texture feature and spectral feature, which ALOS Image spectral information certain influenced the optimal texture parameters. The texture parameters selected by factor analysis method is based on the texture variables, which have different weighting coefficients, but by classification accuracy did not consider weighting coefficients of texture variable, all the weighted coefficient of texture variable is 1.4) The best window size was affected by distance value, and different distance value corresponds to different best window size. For distance value is 1 pixel, the best window is the maximum texture window size(15×15), the best window of the remaining distance value is the minimum texture window size, and the overall accuracy is 15×15(135°directions,distance value is 1 pixel). This is because, For distance value is 1 pixel, with the increase of the window size, internal details of the object were reduced, homogeneity was growing,and the fuzzy degree of the object boundary was acceptable, the separability of different objects was increasing. For distance value is not 1 pixel, with the increase of the window, texture boundaries were more blurred, deviated from the boundary the objects were more distantly, the phenomenon that objects have the different texture unit was more serious, the separability of different objects was reducing.5) Window size and direction affected the optimal distance value, the best distance value concentrated in 1 and 3, the best distance value of window size of 7×7 and 9×9 were 3 pixels(except for 135°directions, the best distance value is 1 pixel), the best distance value of window size of 11×11, 13×13 and 15×15 was 1 pixel. This is because, for distance value were 1 and 3, the majority pixels were likely to belong to the same object, and the GLCM texture features can reflect the spatial characteristics of the object. for distance value were 5 and 7, corresponding to the actual distance were 12.5m and 17.5m, larger than part of the object(such as roads, fallow land), the compared pixels were not usually located within the same object, the GLCM texture features can’t objectively reflect space feature.6) The texture direction effected classification accuracy limitedly, the best texture direction of most objects concentrated in two direction, and the direction of concentrated is not always as same as the highest classification accuracy’s. By statistical the main direction of the object, it was found that the main direction was not fixed, and the range of the main direction is relatively large. In a word, the main direction of object were not fixed, which is the main reason for effecting classification accuracy limitedly.7) The overall accuracy of multiscale texture is higher than object-based and window-based texture, which is improved by 3.98% and 10.56% respectively. Multiscale texture also can improve the classification accuracy of uncovered greenhouse, greenhouse covered by plastic, Residential land, ponds, roads, fallow land and orchard.