Interpretable and Pattern Aware Cloud Segmentation Using Feature Semantic Learning

Abstract

The Landsat archive is essential for Earth surface monitoring and land cover changes. To utilise this wealth of data effectively, cloud and cloud shadow interference must be filtered, particularly for automated systems that process hundreds of photographs without human input. Cloud and cloud shadow interference often obscure clear ground views. Clouds block optical satellites from tracking the globe. These clouds must be precisely located for remote sensing using satellite picture archives. Despite improvements, cloud identification remains challenging, particularly in bright environments and thin cloud layers. Recent breakthroughs in deep learning-based cloud masking have improved cloud identification accuracy. These approaches are often evaluated differently, and studies compare them to threshold-based systems inconsistently. We propose Ed-CNN, a deep convolutional neural network (DCNN) with an enhanced encoder-decoder architecture, for cloud and snow segmentation. This method enhances segmentation accuracy by augmenting the encoder with the Atrous Spatial Pyramid Pooling module and the decoder by integrating information from multiple encoder stages. We recommend standardising datasets and protocols for cloud detection model benchmarking, as deep learning models trained on similar sensor data can benefit new satellites and heavily rely on training and testing datasets.