New deep learning based approaches for land cover classificationin satellite images

Abstract

This dissertation provides an in-depth examination of strategies for improving agricultural monitoring through remote sensing. It focuses on three main contributions: the development of FAUNet for delineating parcel boundaries, a novel technique that combines the Segment Anything Model (SAM) with principal component analysis (PCA) to refine segmentation, and the use of thermal time modeling to enhance crop classification across different climates. Parcel boundary delineation serves as a crucial step in agricultural monitoring, ensuring precise segmentation of land parcels for applications such as yield forecasting and land-use planning. To address limitations in existing models, this study presents FAUNet—an innovative dual-headed U-Net specifically tailored for boundary detection in agricultural imagery. FAUNet employs a high-frequency attention module (based on high-pass filters) and a dual-path design that predicts both edge and extent masks. When compared to leading models like U-Net, ResUNet-a, SEANet, and BsiNet, FAUNet delivers the highest object-level F1 score (0.7734) and exhibits notably low over-segmentation (0.0341) and under-segmentation (0.1390) rates. Building on these insights, the dissertation introduces a new method to enhance segmentation by coupling SAM—a foundational segmentation model originally trained on diverse natural images—with PCA. Since SAM's training data do not include specialized remote sensing imagery, its performance in this domain can be limited. To address this issue, SAM's high-dimensional embeddings are first extracted, then reduced with PCA, followed by guided filtering to refine the inputs. This iterative feedback loop helps SAM generate more precise boundary delineations, ultimately improving segmentation results in challenging remote sensing scenarios. The thesis then turns its attention to the challenge of generalizing crop classification models in regions with varying climates. Standard machine learning models (e.g., Random Forest, Gradient Boosting, XGBoost, SVM, and MLP) often encounter difficulties when facing the temporal shifts driven by different local growing conditions. To mitigate this, thermal time modeling based on Growing Degree Days (GDD) is introduced. By aligning crop phenological stages and smoothing out timing discrepancies, GDD helps these models adapt to spatial variability more effectively. Experiments on datasets from Turkish regions with diverse climates show that incorporating GDD boosts classification performance, allowing models to generalize more reliably across geographically distinct environments. Overall, this dissertation tackles significant obstacles in agricultural remote sensing, ranging from accurate parcel boundary detection to robust crop classification under complex environmental conditions. The proposed FAUNet framework streamlines boundary delineation, the SAM modification allows it to perform better in boundary delineation, and thermal time modeling underscores how classification can be adapted for real world agricultural scenarios.