ABSTRACT
This chapter proposes to address a problem of image restoration, which is quite different in nature than land cover mapping, to show that deep learning can also be applied to other topics in remote sensing. A well-known issue impacting optical imagery is the presence of clouds. The need for cloud-free images at precise date is required in many operational monitoring applications. Each encoder implements a convolutional neural network with strides. It consists of successive 4×4 convolutions with stride 2 to downsample the input. Each convolution is followed by a leaky rectified linear unit. Each transposed convolution is followed by a rectified linear unit. It is then concatenated with the corresponding feature maps from the encoder.
