Learning with limited labeled data for Remote Sensing Image Analysis

• Autores: Itzá Alejandra Hernández Sequeira
• Directores de la Tesis: Filiberto Pla Bañón (dir. tes.) , Rubén Fernández Beltrán (codir. tes.)
• Lectura: En la Universitat Jaume I ( España ) en 2026
• Idioma: inglés
• Número de páginas: 109
• Tribunal Calificador de la Tesis: Luis Gómez Chova (presid.) , José Salvador Sánchez Garreta (secret.) , Rosa María Valdovinos Rosas (voc.)
• Enlaces
  • Tesis en acceso abierto en: TDX
• Resumen

  • The increasing availability of satellite and aerial imagery from Earth observation missions has created unprecedented opportunities for land cover analysis. However, applying deep learning to remote sensing faces a fundamental challenge: the scarcity of labeled training data. Unlike standard computer vision with large annotated datasets, Remote Sensing (RS) requires specialized expertise for annotation, creating bottlenecks for operational applications. This dissertation addresses these challenges by systematically investigating data-efficient deep learning methods that leverage both limited labeled samples and abundant unlabeled remote sensing imagery. Through four peer-reviewed publications, we explore how transfer learning, semi-supervised learning, and metric learning can reduce annotation requirements for scene classification and semantic segmentation tasks. First, we demonstrate that domain-specific pre-training on RS datasets significantly outperforms general-purpose initialization, challenging the assumption that transfer learning from natural images is always optimal. Second, we establish that state-of-the-art semi-supervised methods with only 4 labeled samples per class exceed supervised approaches using 50\% of training data, providing comprehensive computational cost analysis to guide practical deployment. Third, we introduce a novel framework that integrates triplet metric learning with semi-supervised strategies, achieving improved class separation and faster convergence for confusable categories. Fourth, we pioneer the adaptation of semi-supervised segmentation to multimodal remote sensing, successfully combining RGB, near-infrared, and elevation data to achieve significant performance gains over single-source approaches.