Detection of Lodging Rice Areas in Rice Fields by Semantic Segmentation via a Fisheye-lens Camera

Abstract

To detect lodging rice areas with a fisheye-lens camera, two semantic segmentation models based on deep neural networks were developed and tested with fisheye images and converted images. The two models could detect the lodging rice areas and unharvested rice areas in the front of the combine harvester. The two models were trained to detect the following seven classes, 1) unharvested rice areas, 2) harvested rice areas, 3) lodging rice areas, 4) ridges, 5) humans, 6) other combine harvesters, and 7) background. Unharvested and lodging rice areas are the main focus of this study. Intersection over union (IoU) of these two classes was 0.8446 and 0.8524 respectively of the best model FCNResNet50, trained with both fisheye images and converted images.