Redesigning U-Net with dense connection and attention module for satellite based cloud detection

Abstrak

In this paper, we present an upgraded U-Net technique for satellite-based cloud detection, with additional features, such as, more relevant spatial information, improvement in gradient propagation, feature reuse and controlling the network parameters using growth rate by adding dense connections. Furthermore, incorporation of attention module helps to learn strong inter-spatial and inter-channel relationships of feature maps by adding a few trainable parameters to the network. The two attention blocks namely position attention module (PAM) and channel attention module (CAM) focus on important parts of the image by neglecting the redundant information. The experimental results prove that the put forward technique with dense and attention modules could detect cloud with an accuracy of 95.69%.