SEMANTIC SEGMENTATION OF AGRICULTURAL FIELD ANOMALY PATTERNS IN AERIAL IMAGES USING U-NET

Abstract

Precision agriculture requires efficient monitoring of large agricultural fields with the aim that crop anomalies can be identified and spatially located as early as possible to enable targeted and timely field management. The traditional in-field inspection is time-consuming, tedious and measures only a small number of points. Deep learning technologies are coupled with unmanned aerial vehicles (UAVs) to create the opportunity to automate highresolution agricultural imagery. This work examines a task specific U-Net approach to multi label segmentation of agricultural anomaly patterns, but not a novel network architecture, and rather emphasizes the impact of practical design decisions. The evaluated configuration has four-channel RGB-NIR (NRGB) input, uses sigmoid-based multilabel prediction, and uses a combined Binary Cross-Entropy (BCE) and Dice loss. The experiments are carried out on the Agriculture-Vision 2021 benchmark dataset with over 21,000 labeled aerial images of eight agriculture anomaly classes. To isolate the contributions of NIR channel and Dice loss component, a controlled ablation study is performed, and additional experiments compare the resulting configuration to an identical experiment protocol against the standard U-Net, FCN, SegNet and a modern segmentation baseline. The performance is measured by Intersection over Union (IoU) and mean Intersection over Union (mIoU) on each class and the overall mean, respectively. Results show the effectiveness of NIR information and Dice-based loss for anomaly segmentation in agriculture and the performance of a relatively simple U-Net-based configuration on the benchmark dataset. The results validate the potential of using task-specific U-Net structure for automated UAV monitoring in agriculture and demonstrate individual input and loss-function contribution to the task.