Automatic Tree Counting from Satellite Imagery Using YOLO V5, SSD and UNET Models: A case study of a campus in Islamabad, Pakistan

Abstract

Since the last 3 years, Pakistan has been focusing considerably on the increase in tree plantation in several areas throughout the country. With this increase in plantation, the need for up-to-date record keeping for upkeep of these trees across the country arises. The extensive research in object detection and image segmentation models have led to a much faster method of satellite image based tree counting to replace conventional counting methods. This paper focuses on tree detection and counting using satellite images, spanning a total of 8 years, of a university campus located in the capital of Pakistan. It effectively makes use of data augmentation techniques to improve the accuracy of the implemented models which include YOLOV5, UNET and SSD. The satellite images taken over the years are used to generate a new data set and then the produced dataset is augmented using the techniques of rotating, flipping, and patching. The augmented data set is fed into the object detection and image segmentation models for training. The models are then compared on the basis of loss and accuracy to see which model was better suited to carry future work. The concluding results gave accuracy of 32%, 81%, and 24% for the YOLO, UNET and SSD models respectively. Future improvements include the use of high-resolution images and a larger data set to enhance the accuracy of the resulting models.