Remote Sensing: Improving Land Cover Classification using Satellite Signal and Deep Learning

Abstract

Remote sensing and land cover classification are useful tools for comprehending, monitoring, managing natural resources as well as environmental changes around the world. One key use of remote sensing is to accurately identify and classify different types of land, such as water bodies, flood-prone areas, farms, buildings, rice fields, roads, and construction sites. This plays a big role in helping with disaster response, monitoring agriculture, and using resources in a sustainable way. A key task is accurately segmenting satellite imagery and extracting features to identify regions of interest, using both optical and Synthetic Aperture Radar (SAR) data. However, remote sensing image segmentation presents several technical challenges, including limited availability of high-quality labeled datasets, high computational demands, class imbalance, cloud obstruction, spectral similarity among crop types, and complexities due to spatial and temporal variations. Conventional index-based methods Normalized Difference Water Index (NDWI) and Modified Normalized Difference Water Index (MNDWI) and traditional convolutional neural network (CNN) architectures, though widely used, frequently fall short under real-world complexities and specific sensor limitations such as cloud cover, seasonal changes, etc. Previous methods typically relied on conventional CNN architectures and straightforward thresholding methods, encountering major limitations in handling multichannel data fusion and class imbalance effectively. To address these issues, this thesis introduces several novel approaches successively. First, we propose the Fusion Adaptive Patch Network (FAPNET), designed to improve flood-water detection from SAR imagery by integrating adaptive multi-channel data fusion and innovative Neural Adaptive Patch (NAP) augmentation. This method significantly reduces computational demands while accurately segmenting water-affected areas. Building on this foundation, we further developed the Patch Layer Adaptive Network (PLANET), a more advanced segmentation network dynamically adapting its structure to varying input resolutions. Furthermore, PLANET efficiently segments multiple land cover classes, addressing limitations found in previous conventional CNN architectures by optimizing memory usage and enhancing segmentation precision. Recognizing that deep learning model training requires extensive labeled datasets often unavailable in remote sensing scenarios, we introduced Pixelwise Category Transplantation (PCT), an innovative data augmentation technique specifically tailored for water-body extraction tasks from Sentinel-2 imagery. PCT effectively reduces uneven class representation and boosts model learning by artificially increasing the amount of high-quality labeled data. Moreover, our thesis extensively investigates the use of spectral compositions and phenological stage information in agricultural land segmentation tasks, particularly focusing on rice crop identification. By conducting an in-depth assessment of the spectral band combinations from Sentinel-1 and Sentinel-2 imagery, we demonstrate that the appropriate selection of bands and temporal information significantly enhances segmentation precision, especially in complex agricultural landscapes. This research provides significant contributions to the advancement of remote sensing-based land cover classification, introducing practical solutions to critical limitations inherent in current methodologies. Future research will further refine these models, utilize additional spectral and temporal information, extending their use to a wider range of environmental and agricultural remote sensing applications. The main contributions of this research are given below: 1. In order to maximize the effectiveness of the learned weights, we introduce a multi-channel Data Fusion Module (DFM), which was created using Vertical transmit and Vertical receive (VV), and Vertical transmit and Horizontal receive (VH) polarization data. Also, elevation data from the NASA Digital Elevation Model (NASADEM), incorporating feature fusion, normalization, and end-to end masking, was used. 2. Inspired by CNN architectures, we developed a powerful data augmentation method called the NAP augmentation module. It extracts features at multiple scales using unit kernel convolution to create patches, helping the model learn faster and better understand the meaning of the data. 3. We propose FAPNET, which is a lightweight, memory-time-efficient, and high-performance model, as demonstrated in our quantitative and qualitative analysis. 4. The introduction of PCT, a novel form of data augmentation applicable not only to water body detection but also to a variety of generalized image segmentation tasks. 5. Introducing the PLANET model, which is an enhanced version of our previously proposed FAPNET model and the NAP-based data augmentation method, with an additional capability to handle multi-class segmentation. 6. Integrating the ability to dynamically generate segmentation encoder-decoder layers, providing simpler structures for smaller input sizes, and more complex structures for larger input sizes in order to achieve higher accuracy and efficient memory usage. 7. Design and implement robust methodologies to automate the mapping of irrigated rice fields using a median mosaic image composition.