Comparing U-Net convolutional network with mask R-CNN in the performances of pomegranate tree canopy segmentation

Abstract

In the last decade, technologies of unmanned aerial vehicles (UAVs) and small imaging sensors have achieved a significant improvement in terms of equipment cost, operation cost and image quality. These low-cost platforms provide flexible access to high resolution visible and multispectral images. As a result, many studies have been conducted regarding the applications in precision agriculture, such as water stress detection, nutrient status detection, yield prediction, etc. Different from traditional satellite low-resolution images, high-resolution UAVbased images allow much more freedom in image post-processing. For example, the very first procedure in post-processing is pixel classification, or image segmentation for extracting region of interest(ROI). With the very high resolution, it becomes possible to classify pixels from a UAV-based image, yet it is still a challenge to conduct pixel classification using traditional remote sensing features such as vegetation indices (VIs), especially considering various changes during the growing season such as light intensity, crop size, crop color etc. Thanks to the development of deep learning technologies, it provides a general framework to solve this problem. In this study, we proposed to use deep learning methods to conduct image segmentation. We created our data set of pomegranate trees by flying an off-shelf commercial camera at 30 meters above the ground around noon, during the whole growing season from the beginning of April to the middle of October 2017. We then trained and tested two convolutional network based methods U-Net and Mask R-CNN using this data set. Finally, we compared their performances with our dataset aerial images of pomegranate trees. [Tiebiao- add a sentence to summarize the findings and their implications to precision agriculture]