SPOT 5 Remote Sensing Image Classification Based On The Methods Of Neural Networks And Object-oriented Technique

With the development of remote sensing technology, the resolution of remote sensing image has been greatly improved, which leaded to widely application in forestry. Compared to the medium- and low-resolution image data, the spatial information, structural features and texture information of the objects and other detail information of the high-resolution image data are abundant. Thus, high-resolution remote sensing images are widely used in the extraction of forest information. In the earliest time, forest information was extracted from the image with the technology of visual interpretation classification, followed by a digital computer classification of remote sensing images. But even the neural network classification, a traditional technology based on pixels, is a relatively mature approach; it could only be better applied for the information extraction of medium- and low-resolution remote sensing image in the computer classification techniques. Neural network classification methods cannot effectively make use of the abundant information of the high-resolution remote sensing images, and is likely to result in a "salt and pepper" phenomenon, leading to a poor quality of classification accuracy. However the object-oriented image classification method contains spatial information and treat texture feature of homogenous objects as the basic processing unit.The data used was SPOT5 images of Nanbu County, which contained a multispectral images with a spatial resolution of 10 m and a panchromatic image with spatial resolution of 2.5 m. We processed neural networks and object-oriented image classification respectively based on the images of Nanbu County, and we compared these two results and assessed the accuracy. In this paper, we conducted neural network classification with ENVI and completed object-oriented image classification using eCognition. After visually compared the neural network and object-oriented classification, the results showed that:using object-oriented classification to extract information from high-resolution remote sensing image can effectively avoid the "salt and pepper" phenomenon. Comparing these two technologies with confusion matrix, we found:(1) In these study area, optimal segmentation parameters differentiate vegetation and non-vegetation, farmland and woodland, forest land and open forest land, paddy fields and waters was 20, the shape factor was 0.1, and the optimal color factor was 0.9. When segmentation scale was 30 and the shape factor was 0.3, supplemented with NDVI index; and when segmentation scale was 60, the shape factor was 0.8; we can get a better segmentation result.(2) After fully considering the segmentation scales and spectral information of each level, we discovered that we can effectively distinguish woodland with NDVI index greater than 0.52 and the red band less than 290. Farmland can be highly segmented at the brightness values between 110 and 235, and the NDVI index less than 0.52. Building land can be best accurately separated at the brightness values between 188 and 216, near red band between 260 and 300, NDVI index less than 0.46.(3) The accuracy of results evaluation showed that the overall accuracy and kappa coefficients of neural network classification were 74.32% and 0.7631 respectively. The overall accuracy and kappa coefficients of object-oriented classification and were 92.89% and 0.9148 respectively, which was significantly higher than the neural network.