| Image classification, one of the most important aspects of remote sensing, is the categorization of all pixels in an image into land cover types. The majority of image classification methods fall into either statistically-based, structurally-based, or Artificial Neural Network (ANN) approaches. Until recently, very few attempts have been made to increase classification accuracies by combining two or more of these approaches. Furthermore, most of these attempts were limited to post classification process.; This research employed a modified version of the Maximum Likelihood (ML) classifier. Typically, a conventional ML classifier computes the statistical probability of each pixel being a member of every land cover class, then assigns the pixel to the class with the highest probability. By doing that, the ML classifier generates a large amount of information but discards most of it after it assigns the pixel to one particular class. The Modified Maximum Likelihood (MML) classifier, on the other hand, outputs layers of classes along with corresponding probability values. These classes are then ordered, according to the probability values from the highest value to the lowest. Later, the original image pixel values, the output from the MML, and the spatial information are introduced to a trained Back Propagation Feed Forward Neural Network (BPFFNN) to make a final decision as to which class the input pixels are to be assigned.; This combined classification approach was tested on two scanned aerial photographs, one a color image and the other a color IR image. classification results were compared to those from conventional supervised and unsupervised ML classifiers. The combined classification approach improved the overall classification accuracy by an average of 17% over the conventional ML classifier. Not only did this method improve classification accuracies, but it also reduced the human interaction time required. Finally, this research employed statistical experimental designs to evaluate the factors involved in the image classification process and to determine the significance of all the potential input features relative to classification accuracies. |
