Accurate and timely identification of plant abnormalities is vital in agriculture, since delays can severely reduce both crop health and yield. Automating this task has the potential to generate significant environmental and economic benefits. In this work, we investigate the detection and quantification of abnormalities in maize fields using high-resolution RGB imagery gathered by Unmanned Aerial Vehicles (UAVs). A limitation of many existing methods is that they depend on relatively small datasets that are also tuned to the characteristics of specific fields. As a result, models trained on one field do not perform well on data from another field due to domain shift. While data augmentation and synthetic image generation have been used to increase dataset size and diversity, these techniques often fail to provide the data variability and fidelity required for robust agricultural applications. An alternative strategy is to merge data from multiple fields; however, this approach is ineffective without first normalizing the data to account for irrelevant variations such as lighting conditions, sensor characteristics, and soil types. To address this challenge, we propose a framework for standardizing multi-field data by mapping them into a shared domain using unsupervised domain adaptation (UDA). To further improve quality, we introduce a selective UDA strategy that filters out poorly adapted images prone to artifacts. During training, images captured under diverse conditions are aggregated in this shared domain, enriching the training set with realistically domain-adapted data rather than relying solely on synthetic augmentations. For detection and quantification, we leverage Vision Transformers (ViTs), while an ensemble of CycleGANs is used for domain adaptation. We validate our framework on a publicly available dataset of UAV-based, high-resolution RGB images of both healthy and abnormal maize plants across multiple growth stages, collected from two fields with diverse environmental conditions.