Advancing Land-Cover and Land-Use Classification with Semantically Consistent and Data-Efficient Deep Learning on Multi-Source Earth Observation Data

Land-use and Land-cover (LULC) mapping from Earth Observation (EO) imagery plays a fundamental role in environmental monitoring, urban planning, agricultural management, and climate change assessment. Recent advances in satellite missions, Machine Learning (ML), Deep Learning (DL), and foundation models have significantly improved large-scale mapping capabilities. However, important challenges remain, including semantic inconsistencies between satellite imagery and in-situ observations, the lack of large-scale benchmark datasets with detailed LULC information for training and evaluating advanced DL models, heterogeneous data sources, mixed pixel effects, limited availability of annotated training data, and the need for computationally scalable processing frameworks. Addressing these challenges requires methodologies that integrate multiple data modalities while improving semantic consistency, data efficiency, and spatial representation.

This thesis presents a comprehensive framework for advancing satellite-based LULC mapping through four complementary research directions. First, it investigates the semantic relationship between satellite observations and in-situ field data by leveraging geo-tagged field photographs as an intermediate observational layer. The analysis demonstrates that apparent disagreements between satellite-derived and field-based information often arise not only from limitations of satellite observations but also from ambiguities in field interpretation and inconsistencies in land-cover (LC) taxonomies.

Second, the thesis produces the Sen4Map, a large-scale benchmark dataset, integrating Sentinel-2 imagery with harmonized Land Use/Cover Area Frame Survey (LUCAS) field observations and ancillary information. The benchmark provides semantically rich annotations designed to support the development and evaluation of robust DL models for LULC mapping.

Third, the thesis develops methods for sub-pixel LULC estimation to capture fine-scale landscape heterogeneity beyond the native spatial resolution of satellite imagery. As a proof of concept, the proposed framework estimates built-up fractions from Sentinel-2 imagery using geospatial foundation model embeddings, demonstrating the potential of pretrained spectral-spatial representations for sub-pixel characterization.

And finally, the thesis explores data-efficient learning strategies for EO by investigating few-shot learning with meta-learning on publicly available Remote Sensing (RS) benchmark datasets. These approaches demonstrate improved generalization under limited annotation availability, reducing the dependence on extensive labeled datasets while maintaining competitive mapping performance across diverse environments.

Beyond methodological advances, the thesis also contributes scalable computational solutions through a HDF5-based data architecture optimized for High-Performance Computing (HPC) environments, enabling efficient storage, access, and processing of large multi-source EO datasets. Collectively, the proposed methodologies contribute toward more semantically consistent, data-efficient, and scalable satellite-based LULC mapping systems. As EO transitions toward foundation models, unified embedding representations, and agentic geospatial intelligence, the contributions presented in this thesis establish the datasets, semantic frameworks, and embedding-centric methodologies required to support the next generation of LULC mapping.