抄録
In this paper, we propose four kinds of landcover classification models using neural networks which are driven by co-occurrence matrices. The first and the second models classify each band image using a neural network at the first stage, then perform a final decision based on the first stage results. At the second stage, an arithmetic decision algorithm is used in the first model, and a second neural network is used in the second model. The third model has individual input and hidden layers for each band. However, the output layer of the third model collects all hidden layer outputs. The fourth model is a single stage classifier, and co-occurrence matrices of all bands are inputted into an input layer at the same time. In order to evaluate the performance of these proposed models, landcover classifications using these models and three kinds of conventional classifiers were conducted for Landsat TM and SPOT HRV data. The conventional methods used are a maximum likelihood classifier and two kinds of neural network classifier which use a single pixel and 3X3 pixels as input data. As a result, the best performance was shown by the fourth model, which showed 3% to 21% and 9% to 20% higher accuracy than the maximum likelihood classifier and the conventional neural network classifiers, respectively.
