<![CDATA[The paper explores resolution-accuracy trade-offs in UAV-based semantic segmentation for precision agricultural imagery, which is an important challenge in achieving computational efficiency while maintaining satisfactory results for semantic segmentation. It is widely believed that higher resolution imagery leads to more accurate results, but it also raises processing expenses and restricts the ability to deploy in real-time systems with unmanned aerial vehicles (UAV). This study comprehensively analyzes the effect of the spatial resolution (also known as Ground Sampling Distance (GSD)) on segmentation performance on a range of agricultural anomaly types. The agriculture-vision dataset is used in experimental runs of three different GSDs (10 cm, 20 cm and 40 cm per pixel) with UAV-acquired images. The standard semantic segmentation metrics (mean Intersection over Union (mIoU), Dice coefficient, and computational time analysis) are adopted to evaluate several deep learning models such as U-Net, R2U-Net, U-Net3+, Attention U-Net and DeepLabV3+. The results show that there is no clear correlation between achieving higher spatial resolution and achieving better segmentation accuracy. Medium resolution imagery (20 cm/pixel) can produce similar or better results for large scale anomalies, including dryness and nutrient deficiency, and at significantly lower computational costs. On the other hand, fine-grained anomalies require more fine-grained resolution to better display the anomalies, which shows that the best resolution depends on the task. Further, multi-scale feature aggregation models have higher robustness to resolution degradation. The findings offer a practical understanding of designing the resolution-aware model, which could lead to more efficient use of UAV for resolution-aware deployment, flight planning, and the implementation of edge-AI in precision agriculture. The study provides a data-driven tool for optimizing the spatial resolution versus accuracy versus efficiency of real-world monitoring systems in agriculture.]]>
