Detection and Segmentation for Optical Remote Sensing: From Satellites to UAVs

Abstract

Optical remote sensing, including hyperspectral, multispectral, and RGB imagery from satellites and unmanned aerial vehicles (UAVs), plays a pivotal role in environmental monitoring and ecological conservation. However, the challenges associated with optical remote sensing vary depending on data source. Hyperspectral imagery (HSI) is affected by high dimensionality and spectral noise, Sentinel-2 satellite imagery is limited by coarse spatial resolution, while UAV-derived imagery presents difficulties due to landscape-induced noise and the large spatial extent of orthomosaics. These factors collectively complicate accurate object detection and segmentation in remotely sensed data. This thesis addresses these challenges through a series of methodological advancements, organized in three key areas: unsupervised hyperspectral clustering, semi-supervised change detection, and UAV-based efficient object analysis.

First, we introduce Superpixel-based Spatially Regularized Diffusion Learning (S²DL), an unsupervised clustering framework designed to extract meaningful structures from hyperspectral images without manual annotations. S²DL integrates superpixel segmentation with diffusion geometry to capture spatial-spectral relationships, thereby mitigating challenges associated with spectral variability and noise. By utilizing entropy rate superpixel partitioning, the method generates adaptive spatial units that preserve object boundaries, while a density-purity graph construction ensures robust clustering. This framework effectively reduces computational complexity while maintaining clustering accuracy, making it particularly suitable for large-scale HSI applications, especially in environments exhibiting spatial homogeneity. Applied to mangrove species mapping in Hong Kong’s Mai Po Nature Reserve, S²DL successfully differentiates vegetation types, demonstrating its efficacy for label-free ecological analysis in real-world complex environments.

Second, we applied two complementary approaches for change detection in remote sensing: a supervised Extended ReCNN (E-ReCNN) framework for accurate detection and segmentation, and a semi-supervised Support Vector Machine with Smoothed Total Variation (SVM-STV) method tailored for scenarios with computational constraints and limited labeled data. E-ReCNN is employed to monitor artisanal gold mining activities in the Peruvian Amazon using Sentinel-2 satellite imagery, integrating histogram matching and color space transformations to compensate for atmospheric variations across bi-temporal imagery. This ensures high pixel-wise detection accuracy in identifying land cover changes. In contrast, SVM-STV provides a lightweight alternative that utilizes spatiotemporal information and limited labeled samples to track small water body changes in rainforest ecosystems. This semi-supervised method provides a computationally efficient solution for large-scale environmental monitoring, demonstrating the trade-off between model performance and computational feasibility in remote sensing applications.

Third, we introduce Processing, Inference, Segmentation, and Mapping (PRISM), a deep learning-based pipeline for efficient detection and segmentation of canopy palms in UAV orthomosaics, addressing the computational and scalability challenges of high-resolution aerial imagery. PRISM processes large-scale UAV-derived imagery efficiently while preserving fine-grained details essential for ecological analysis. To enhance spatial pattern characterization, we introduce a bimodal point pattern analysis technique that decouples local clustering effects from global distribution patterns, which provides deeper insights into palm population structures. These UAV-driven solutions bridge high-resolution imaging with scalable ecological monitoring and enable more accurate assessments of biodiversity and forest dynamics.

Beyond environmental monitoring, the detection algorithms have important applications to medical imaging domains. In fetal ultrasound video analysis, deep learning models can be applied to the precise localization of the fetal heart and classification of congenital heart defects (CHDs), which aid early diagnosis and clinical decision-making.

By integrating mathematical and statistical models with machine and deep learning techniques, this thesis presents adaptive frameworks for optical remote sensing data. Future research directions include extending these methodologies to broader ecological applications and developing lightweight edge computing architectures for enhanced processing capabilities.

Date of Award	18 Aug 2025
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Jean-Michel Henri Olivier MOREL (Supervisor) & Hon Fu, Raymond CHAN (External Co-Supervisor)