Comparison of object detection methods for crop damage assessment using deep learning
Severe weather events can cause large financial losses to farmers. Detailed information on the location and severity of damage will assist farmers, insurance companies, and disaster response agencies in making wise post-damage decisions. The goal of this study was a proof-of-concept to detect damaged crop areas from aerial imagery using computer vision and deep learning techniques. A specific objective was to compare existing object detection algorithms to determine which was best suited for crop damage detection. Two modes of crop damage common in maize (corn) production were simulated: stalk lodging at the lowest ear and stalk lodging at ground level. Simulated damage was used to create a training and analysis data set. An unmanned aerial system (UAS) equipped with a RGB camera was used for image acquisition. Three popular object detectors (Faster R-CNN, YOLOv2, and RetinaNet) were assessed for their ability to detect damaged regions in a field. Average precision was used to compare object detectors. YOLOv2 and RetinaNet were able to detect crop damage across multiple late-season growth stages. Faster R-CNN was not successful as the other two advanced detectors. Detecting crop damage at later growth stages was more difficult for all tested object detectors. Weed pressure in simulated damage plots and increased target density added additional complexity.