ABSTRACT
This chapter introduces a simple approach to produce land cover maps from remote sensing images using deep nets. It shows that the patch-based approach has some drawbacks. In particular, the receptive field of one model has an impact on the spatial resolution of the output map generated with the trained model. Since the image content in the receptive field activates the single-valued model output, one can expect that the localization of the predicted class is not reliable if predicted at the same spatial resolution as the input image. In practice, when a land cover map is generated at the same resolution as the input image, rather than the resolution of the last feature map of the network. The chapter examines some new architectures that aim to preserve the spatial resolution of the final features of the model output.
