Semantic segmentation architecture based on Atrous Spatial Pyramid Pooling and convolutional based attention module for mapping of burned areas in satellite images
Loading...
Authors
Meeting name
Sponsors
Date
Journal Title
Format
Thesis
Subject
Abstract
Computer vision is a rapidly advancing field within artificial intelligence and deep learning. It has found applications in various areas, including climate. Remote sensing data is now widely accessible; medium-resolution satellite images can be obtained for free. A custom dataset was created to train deep learning models to map areas affected by wildfires. Developing such models can significantly aid in post-wildfire assessments and recovery efforts. Mapping burned areas can be framed as a semantic segmentation problem. Existing state-of-the-art convolutional neural network (CNN) architectures, such as U-Net and DeepLab have been trained to semantically segment burned areas. Initially successful in natural language processing tasks, transformer networks have recently been adapted for computer vision applications. These networks utilize self-attention layers, which are crucial to their design. SegFormer is a transformer-based model for semantic segmentation that has shown considerable promise. The attention mechanism is not exclusive to transformer networks; it can also be integrated with convolutional operations to enhance performance. For instance, SegNext is an architecture that effectively employs this combination. Seg- Former and SegNext models have also been trained using the same custom dataset and showed better performance than CNN architecture. However, these architectures have a large number of parameters and require a large amount of data to train. Seg- ConFormer, an architecture that combines the strengths of each of these architectures, has been trained from scratch using the custom dataset only, and it outperformed all other architectures using significantly less number of parameters.
Table of Contents
PubMed ID
Degree
Ph. D