Apricot Tree Detection from UAV-Images Using Mask R-CNN and U-Net

Abstract

Monitoring trees is necessary to manage and take inventory of forests, monitor plants in urban areas, distribute vegetation, monitor change, and establish sensitive and renewable agricultural systems. This study aims to automatically detect, count, and map apricot trees in an orthophoto, covering an area of approximately 48 ha on the ground surface using two different algorithms based on deep learning. Here, Mask region-based convolutional neural network (Mask R-CNN) and U-Net models were run together with a dilation operator to detect apricot trees in UAV images, and the performances of the models were compared. Results show that Mask R-CNN operated in this way performs better in tree detection, counting, and mapping tasks compared to U-Net. Mask R-CNN with the dilation operator achieved a precision of 98.7%, recall of 99.7%, F1 score of 99.1%, and intersection over union (IoU) of 74.8% for the test orthophoto. U-Net, on the other hand, has achieved a recall of 93.3%, precision of 97.2%, F1 score of 95.2%, and IoU of 58.3% when run with the dilation operator. Mask R-CNN was able to produce successful results in challenging areas. U-Net, on the other hand, showed a tendency to overlook existing trees rather than generate false alarms.