on the Use of Class Activation Map for Land Cover Mapping

In a supervised classification of a remote sensing image, a ground survey is needed to collect sufficient training data to train a classifier. The ground survey is expensive and time consuming. Thus, the idea of the class activation map (CAM) is proposed for land cover mapping in this paper. Here, instead of training with ground data where the underlying class label of each training pixels must be identified, our goal is to classify every pixel in an image with a training set, where a training sample is the association between an image and their class composition. The identification of class composition with an image is an easier task to perform, since human perception can accurately detect the composition of objects. The CAM is a collection of “heatmap” images associated with different combinations of land cover classes, where the high value indicates a high probability that a pixel belongs to that class. Next, we label a pixel to the class where the corresponding CAM is maximum.