Robust Deep Learning Methodologies for Weakly Supervised Remote Sensing Image Classification

The proliferation of Earth observation missions demands advanced techniques for the automatic analysis of data acquired by satellites. These missions provide increasingly frequent snapshots of the Earth, generating abundant image time-series crucial for applications like climate change studies and environmental monitoring. Deep learning (DL) has proven very effective in addressing the analytical challenges posed by this data, excelling in image analysis and sequential data processing. However, in remote sensing (RS), DL is often hindered by scarce and imperfect labeled data. This limits the effectiveness of DL in RS, requiring the development of DL techniques that exploit different kinds of imperfect and weak annotations. This thesis addresses this critical challenge by developing a suite of novel methodologies for weakly supervised learning (WSL) in RS with a focus on classification tasks such as land-cover mapping and scene classification. WSL strategies are commonly subdivided into three different categories: i) inaccurate supervision, which deals with label noise; ii) inexact supervision, which deals with coarse-grained supervision (e.g., spatial ambiguity), and iii) incomplete supervision, where unlabeled data are available for training. The thesis is structured around five primary contributions addressing the above-mentioned data imperfections. First, we tackle inaccurate supervision from heterogeneous sources by designing a robust framework that handles class-dependent label noise. It effectively fuses data from multiple sources with varying reliability, significantly enhancing model resilience and classification accuracy for tasks like land-cover mapping. Second, we address inexact supervision and spatial ambiguity inherent in RS datasets. We introduce the combination of deep multiple instance learning with positive-unlabeled learning to address the problem of training high-resolution land cover classifiers using low-resolution land cover thematic products. Third, we address incomplete multi-label annotation of RS imagery. We introduce Adaptive Gradient Calibration (AdaGC), a novel and effective method for single-positive multi-label learning. This approach systematically reconstructs comprehensive multi-label information from minimal, single-positive annotations, enabling training of complex models for detailed mapping. Fourth, as a foundational step toward leveraging intrinsic semantic hierarchies, we introduce the Semantics-Aware Hierarchical Consensus (SAHC) approach. This hierarchical self-supervised approach explicitly utilizes the structure of land cover classes to improve feature representation and classification consistency across different class granularities, paving the way for more generalizable and robust models. Fifth, to overcome temporal inconsistencies and discrepancies in multi-year land-cover maps, we developed an effective and robust hybrid Bayesian-DL model. This model integrates any possible DL encoder with Hidden Markov Models to enforce logically consistent land-cover transitions over time. The approach has also shown robustness in a weak supervision scenario where labels are sparsely distributed in time. The proposed methodologies have been rigorously tested on different benchmark RS datasets. They consistently demonstrated significant improvements in classification accuracy, robustness, and generalizability. Collectively, this thesis provides a systematic approach to WSL, fundamentally reducing dependency on high-quality annotations and advancing the state-of-the-art in automated RS image analysis.