Abstract
The objective of this study is to improve the accuracy of the land-use-capability-classification model which is called the “LF (Latency Factor) model” developed by ourselves. In this LF model, we use not only geographical information but also satellite multispectral data. As for the two kernel functions in the LF model, “LFn model” and “LFq model” applying the Quantification method type II and the Neural Network respectively have been newly defined. As the similar area for the characteristic of the training data was discriminated, the land use capability classification map made by using LFn model and LFq model respectively must be ideally corresponded. Based on this basic concept, the purification procedure for the training data was proposed, which is extracted the corresponded pixels of discriminated results with both LFn model and LFq model. Through this procedure, the improvement of 15% or so on the land use capability classification accuracy could be achieved. Furthermore, it was confirmed that the differences of the land use capability classification maps with both of LFn model and LFq model was remarkably decreased.
