EffResUNet: Encoder Decoder Architecture for Cloud-Type Segmentation

Abstract

Clouds play a vital role in Earth’s water cycle and the energy balance of the climate system; understanding them and their composition is crucial in comprehending the Earth–atmosphere system. The dataset “Understanding Clouds from Satellite Images” contains cloud pattern images downloaded from NASA Worldview, captured by the satellites divided into four classes, labeled Fish, Flower, Gravel, and Sugar. Semantic segmentation, also known as semantic labeling, is a fundamental yet complex problem in remote sensing image interpretation of assigning pixel-by-pixel semantic class labels to a given picture. In this study, we propose a novel approach for the semantic segmentation of cloud patterns. We began our study with a simple convolutional neural network-based model. We worked our way up to a complex model consisting of a U-shaped encoder-decoder network, residual blocks, and an attention mechanism for efficient and accurate semantic segmentation. Being an architecture of the first of its kind, the model achieved an IoU score of 0.4239 and a Dice coefficient of 0.5557, both of which are improvements over the previous research conducted in this field.