Weed segmentation in sugarcane crops using Mask R-CNN through aerial images

In this paper we describe an approach to detect weed regions in sugarcane crops using Mask Regions with Convolutional Neural Networks (Mask R-CNN) on aerial images. Experiments were held with twelve combinations using ResNet-50 and ResNet-101 as backbones. We used both "trained from scratch" and transfer-learning from pre-trained datasets combined with image augmentation techniques for training. The ResNet-101 model with pre-trained COCO achieved an Average Precision (AP
50
) of 65.5% and obtained values of 0.803, 0.707 and 0.752 for Precision, Recall and F
1
score. These results suggest a potential use of Mask R-CNN for weed mapping using aerial images.