Fast Village Finder

Automatic land cover classification through satellite imagery is an important remote sensing problem. Varying satellite-sensor’s characteristics reflect their own differences in the images and a sensor independent solution to the problem becomes difficult in general. Along with the sensor specific variations, the time at which the image is captured is yet another parameter. We propose an efficient method for village segmentation. The satellite imagery we use has been captured by varying sensors at varying times. To describe village pixels we compute phase-gradient, cornerness and color features at every pixel of the image. We apply feature selection methods to select salient features for discrimination of village pixels from the rest. We select only one cornerness feature at a particular scale having equivalent capability to discriminate the village pixels. The rest of the features need not to be computed and the feature computation module becomes very fast. Using the selected feature, we segment the image into two clusters by applying k-means++ algorithm. The cluster having larger cluster-center (according to the property of selected feature) is labeled as a village. A MATLAB tool is also built for marking the ground truth for large images with a minimal interaction. Experiments further elaborate effectiveness of work in practical applications especially in large images.