Advanced remote sensing for environmental monitoring. Plastic pollution detection and coral reef habitat mapping using UAV and satellites based on deep learning

Global environmental issues, notably plastic waste and coral reef degradation, necessitate scalable monitoring methods capable of functioning under varied settings. This thesis establishes a methodological framework that integrates aerial and satellite remote sensing with machine learning and deep learning techniques to tackle these difficulties. The study used UAV-mounted hyperspectral sensors targeting the lower bands (900-1700 nm) of the Short-Wave Infrared (SWIR) spectrum for plastics detection. Plastic polymers possess distinct spectral signatures in this region, facilitating automated differentiation from natural backgrounds. The project advanced methodically through 6 phases for plastic detection, from classical supervised machine learning as post-classification to achieving effective real-time detection appropriate for embedded UAV systems, all the way to unsupervised and deep learning approaches. Cross-validation experiments were conducted across multiple independent flights achieving enhanced robustness relative to traditional baselines while preserving pixellevel segmentation accuracy. We extended the study to satellite-scale monitoring by employing two benchmark datasets (MARIDA and MADOS) comprising Sentinel-2 imagery. Problem reformulation (consolidating non-debris classes into unified background) combined with appropriate loss functions achieved effective classification for satellite-based marine debris identification despite severe class imbalance. Complementing the broader remote sensing framework as part of the NASA MarineVERSE project, this work advances coral reef habitat mapping across multiple spatial scales. We developed a modified NeMO-Net model using Attentive Residual U-Net with an EfficientNet- B3 encoder to map benthic habitats in Tumon Bay, Guam, and track temporal changes between 2022 and 2024. Ground truth annotations were obtained using crowdsourcing on NASA’s NeMO-Net citizen science game, tackling the significant limitation of training data paucity. To assess model transfer across sensors and scales, we validated with extensive training data from the Khaled bin Sultan Living Oceans Foundation Global Reef Expedition, spanning 65,000 km² of diverse reef systems surveyed between 2006 and 2015. The workflow builds upon NASA NeMO-Net’s convolutional neural network (CNN) model trained on MAXAR WorldView-2 imagery at meter-scale resolution (2 meters). The model learns habitat-spectral relationships from this training and is tested on an entirely unseen Guam site. The same trained model weights are subsequently applied via transfer learning to ESA Sentinel-2 at decameter-scale resolution (10 meters) and validated again in Guam. This establishes a three-platform validation chain spanning three orders of magnitude in spatial resolution across WorldView-2, Sentinel-2, and Fluid Lensing (sub-centimeter), supporting the development of scalable multimodal learning pipelines from local airborne campaigns to regional satellite mapping for reef monitoring through time. The thesis substantiates a scalable framework that encompasses multiple platforms. Laboratory spectral characterization offers a controlled foundational understanding. UAV systems allow for high-resolution operational implementation. Satellite observations support large-area monitoring. The systematic progression from classical machine learning to deep learning architectures elucidates when and why heightened model complexity is essential, particularly when robust generalization across varied environmental conditions becomes the constraining factor instead of within-domain classification accuracy. Future efforts will develop integrated multi-sensor data fusion to broaden these verified approaches for global-scale implementation