Data augmentation is standard practice in remote sensing (RS) detection pipelines. However, the performance degradation it induces through sensor physics violations, semantic corruption, and domain shift remains largely uncharacterised across application domains. No prior survey has addressed these mechanisms across the full breadth of RS detection tasks. This paper presents the first PRISMA-guided systematic review of the negative impacts of data augmentation in RS detection, synthesising 113 peer-reviewed publications (2014–2026, inter-rater agreement \(\kappa = \boldsymbol{0.81}\) ) spanning nine application domains: aerial object detection, flood inundation mapping, infrastructure crack detection, building damage assessment, wildfire detection, landslide and change detection, crop and vegetation monitoring, synthetic aperture radar-based detection, and oil spill detection. We identify and mathematically characterise six principal failure classes: oriented bounding box annotation corruption (up to 6.2% annotation error rate, \(-\) 3.3% mean average precision loss), synthetic domain shift (generative adversarial network data fractions above 30% breach confidence calibration thresholds), context-inappropriate copy-paste ( \(+\) 8.2–14.7% false positive inflation), spectral integrity violation, comprising normalised difference vegetation index inversion ( \(-\) 4.7% accuracy) and synthetic aperture radar speckle model corruption ( \(-\) 2.8% mean intersection-over-union), semantic coherence destruction ( \(-\) 11.3% damage-class recall, \(+\) 7.3% false change inflation), and training instability (up to 38 \(\times\) computational overhead). The review further delivers a 24-subtype cross-domain failure taxonomy, a severity heatmap across nine tasks and six sensor modalities, 24 physics-grounded mathematical models, and eight evidence-based open research directions. Together, these contributions establish a principled, sensor-aware framework for safe augmentation design in operational Earth observation systems.
