Curriculum Learning in Earth Observation

Curriculum Learning in Earth Observation: This thesis explores the application of Curriculum Learning to the classification of satellite images, a key challenge in the field of Earth Observation. Satellite imagery provides an invaluable source of information for analyzing Earth’s surface, yet its complexity, volume, and diverse data formats make it difficult to process efficiently with conventional machine learning techniques. Deep Learning models, particularly Residual Networks (ResNet), have shown promise in image classification tasks, but often require extensive labeled datasets and high computational resources. This research analyzes how Curriculum Learning, an approach that tries to follow the human learning process by introducing training samples from simple to complex, can improve the efficiency and effectiveness of training models on satellite data. The study begins by reviewing the fundamentals of supervised and unsupervised learning in image classification for Earth Observation and then highlights the difficulties in satellite imagery, including data volume, variable quality, class imbalance, and domain specific labeling needs. The thesis then focuses on how Curriculum Learning can address these challenges by structuring the training process to start with clearer, easier-to-label images and pro- gressively introducing more complex data. This structured approach not only improves generalization but also enhances labeling efficiency and model robustness. The research culminates in the proposal of a structured Curriculum Learning frame- work, integrating modern deep learning architectures such as ResNet and Vision Trans- formers, tailored for satellite image classification. The thesis also outlines potential future directions, including self-supervised learning, dynamic difficulty scheduling, and multi- modal learning.