On Building Detection Using the Class Activation Map: Case Study on a Landsat8 Image

Traditionally, the land cover mapping process needs a ground data to be collected with high precision in both class labeling and spatial locations. To collect enough, high precise ground data require resources. As a result, we proposed an approach for building an image classification based on the class activation map (CAM) where the goal is not to identify the relationship between each pixel and a class label, but to identify whether each sub-images contain the class of interest or not. The output of the class activation map is the filter responds where pixels with high respond are likely to belong to the class of interest. We examined the performance on a LAND-SAT 8 and found. The result of CAM showed that the proposed method achieves high accuracy in identifying whether a sub-image contains the class of interest or not. However, the precision in localizing the class is relatively moderate.