Design and Development of a Framework for Semantic Segmentation of Satellite Images

Abstract

Semantic segmentation of satellite images is important for a variety of land cover and land use mapping applications. The challenge of creating a framework for the semantic segmentation of high-resolution, multi-spectral (HRMS) optical satellite images has been the focus of the research work that this thesis embodies. This thesis presents the following novel contributions. newlineA novel district-scale land cover dataset was created for the region of Ludhiana district in Punjab, India. The dataset covers an area of 4978 km2. HRMS Sentinel-2 Satellite Image Time Series (SITS) for 2020 is included in the dataset along with the ground truth mask. newlineA novel framework was proposed for the semantic segmentation of HRMS satellite imagery that allows the download of satellite imagery, addition of different spectral indices, removal of overly cloudy scenes, temporal interpolation of satellite scenes to fill out gaps, and feeding the pre-processed data into a machine/deep learning architecture for training and/or inference. newlineA deep learning-based encoder-decoder architecture that combines a lightweight channel attention mechanism with a U-Net++ for semantic segmentation based on pre-trained models as encoders was developed to investigate the generalizability of the features learned by a pre-trained deep learning model on common items to the satellite imagery domain. The segmentation problem was found to benefit from transfer learning, and it was found that pre-trained models can also be employed as encoders for the satellite imagery domain after required architectural modifications. newlineAnother architecture was proposed for segmentation of agricultural fields from satellite images consisting of U-Net++ with pre-trained ResNet encoder, attention and feature fusion bars. newlineThe multi-temporal element of satellite imagery was leveraged by interpreting each satellite image collected over a region as a 3D temporal sequence and by proposing a 3D encoder-decoder semantic segmentation architecture with a lightweight 3D Coordinate attention. newline