Enhancing Change Detection Accuracy in Remote Sensing Images Through Feature Optimization and Game Theory Classifier

Satellite-based change detection involves comparing multi-temporal images to identify modifications in land cover features. This work investigates the application of a game theory classifier to enhance accuracy in medium-resolution multispectral remote sensing images. The proposed post-classification approach includes segmentation, feature extraction, classification, and image differencing to detect changes in multi-temporal images. To optimize multispectral images, land cover types are segmented using a proximal splitting algorithm. Boundary and texture features are then extracted using the Difference of Offset Gaussian Filter and Gray Level Co-occurrence Matrix. Principal Component Analysis is subsequently applied to reduce the dimensionality of the extracted features. Finally, the reduced features are classified using a game theory classifier, which effectively handles the uncertainty and variability inherent in non-smooth multispectral data. Experiments were conducted using Landsat datasets from the Hanoi and Balcoc regions, evaluating parameters such as misclassification rate, mean square error, color peak signal-to-noise ratio, and validity index. Quantitative analysis showed that the proposed approach achieved misclassification rates of 0.10 and 0.11 for dataset 1 and 2, respectively. Qualitatively, the results underscore the effectiveness of the extracted features in aiding the game theory classifier to discern subtle differences among feature classes.