Integration of classification methods for improvement of land-cover map accuracy

Abstract Classifiers, which are used to recognize patterns in remotely sensing images, have complementary capabilities. This study tested whether integrating the results from individual classifiers improves classification accuracy. Two integrated approaches were undertaken. One approach used a consensus builder (CSB) to adjust classification output in the case of disagreement in classification between maximum likelihood classifier (MLC), expert system classifier (ESC) and neural network classifier (NNC). If the output classes for each individual pixel differed, the producer accuracies for each class were compared and the class with the highest producer accuracy was assigned to the pixel. The consensus builder approach resulted in a classification with a slightly lower accuracy (72%) when compared with the neural network classifier (74%), but it did significantly better than the maximum likelihood (62%) and expert system (59%) classifiers. The second approach integrated a rule-based expert system classifier and a neural network classifier. The output of the expert system classifier was used as one additional new input layer of the neural network classifier. A postprocessing using the producer accuracies and some additional expert rules was applied to improve the output of the integrated classifier. This is a relatively new approach in the field of image processing. This second approach produced the highest overall accuracy (80%). Thus, incorporating correct, complete and relevant expert knowledge in a neural network classifier leads to higher classification accuracy.