Implementation of Atrous Convolution for Cloud Segmentation in Grayscale & RGB Images from Landsat 8

Advancements in remote sensing over the past two decades have enabled the production of high-resolution multispectral imagery, necessitating improved analytical methods. Object-Based Image Analysis (OBIA) has become a principal approach with image segmentation being a critical initial step. This study examines various image segmentation methods, particularly focusing on semantic segmentation for cloud images. Accurate cloud detection in satellite imagery is vital for numerous applications, leading to the development of Cloud-Net, an end-to-end cloud detection algorithm utilizing a Fully Convolutional Network (FCN) based on a U-Net architecture. This research implements atrous convolution on Cloud-Net in both contracting and expanding blocks, reinstating vanilla skip connections to reduce complexity. Experimental results using Landsat 8 imagery demonstrate that while the proposed Atrous Cloud-Net improves precision and specific metrics evaluation for grayscale images, the original Cloud-Net performs better for RGB images. The findings highlight the trade-offs between different segmentation methods and the importance of multi-channel imagery for higher accuracy in cloud detection tasks. Future work will focus on further refining segmentation models and exploring additional channels to enhance performance.