Abstract
To get the better land-cover classification using the satellite multi-spectral data, the modified Maximum Likelihood classification method applying Genetic algorithms (termed MLG method) is newly proposed. As a major premise of the MLG method, the prior-probability terms of the discriminate function of maximum likelihood process are dealt with the "revision-terms" for classifying each category in the multi-dimensional feature space. A distinctive feature of this method is that the revision-terms of the discriminate function could be estimated though "GA operations".
The required conditions in identifying the discriminate functions are not only to maximize the classification accuracy for the training data itself (division accuracy) and for the reference data, which is used to evaluate the overall accuracy (PCC : Probability of Correct Classification) in the image, but also to minimize the error rate considering both the "omission error" and the "commission error" as well.
To satisfy those conditions simultaneously, then the revision-terms (prior-probability) of the discriminate function are estimated through the "GA operations". Three examination cases associated with modifying the discriminate functions are executed : Case 1) using the equal prior-probabilities, Case 2) using the prior-probabilities estimated from the training data, and Case 3) using the prior-probabilities estimated by using GA operations.
Through those experiments, we conclude :
The GA operations functioned well to increase both the division accuracy and PCC, as well as to decrease the error rate simultaneously.
The average of division accuracy and PCC increase to 80.2% (Case 3) from 78.5% (Case 1) and to 82.1% (Case 3) from 72.2% (Case 1), respectively. Also, the average of error rate decreases to 41.1% (Case 3) from 46.9% (Case 1).
These results indicate that the GA-based classification (MLG method) of Case3 is the most effective for improving the classification accuracy.
