Overcoming the annotation bottleneck: label-efficient deep learning for precision agriculture and remote sensing

The deployment of deep learning models in unstructured environments, such as precision agriculture and remote sensing, is frequently limited by the annotation bottleneck. Acquiring dense, high-quality labels in these domains is an expensive and time-consuming process, complicated by persistent covariate shifts, sparse targets, and pronounced class imbalances. This dissertation presents a cohesive methodological approach to overcome data scarcity by focusing on label-efficient learning strategies. The research utilizes two primary mechanisms to construct and amplify training signals: surrogate supervision, which derives structured targets from weak cues and structural priors, and compositional synthesis, which reshapes the effective training distribution by recombining object- or region-level instances. The first part of the thesis focuses on precision agriculture, with a specific application to automated harvesting and monitoring in table grape vineyards. It introduces a weakly and semi-supervised pipeline that addresses domain shifts in object detection and instance segmentation through automated pseudo-label generation. To further reduce the reliance on manual field data collection, the research presents a hybrid data generation pipeline that blends real fruit instances into simulated 3D environments. This compositional synthesis approach is subsequently adapted for anomaly detection, utilizing foundational models and classical edge detection to generate realistic samples of rare agricultural defects. Additionally, the research addresses the data limitations associated with measuring internal fruit quality. It demonstrates that a multi-task deep neural network can accurately estimate Soluble Solid Content using low-cost RGB sensors, maintaining robust performance even when ground-truth labels are sparse, thereby providing a practical alternative to specialized instrumentation. The second part transfers these core principles to the macroscopic scale of Earth observation. It adapts instance-level cut-and-paste augmentation to satellite semantic segmentation by extracting discrete objects from composite labels to improve model generalization. Building on this concept, the thesis introduces ChangeMix, a change-aware augmentation strategy for semi-supervised change detection. By maintaining a dynamic bank of high-confidence change regions and injecting them into unlabeled data while balancing temporal order, ChangeMix effectively manages the class imbalance and temporal bias characteristic of remote sensing datasets. Across both domains, the experimental evaluations demonstrate that the proposed methodologies consistently improve model performance and robustness under limited labeling budgets. By maximizing the utility of available data and generating targeted synthetic supervision, this dissertation offers a scalable and accessible approach for deploying reliable perception systems in complex, real-world scenarios.