The effect of pre- and post-processing techniques on tree detection in young forest stands from images of snow cover using YOLO neural networks

Abstract

A neural network model for individual tree detection was developed based on the YOLOv4 architecture, which underwent additional preprocessing and postprocessing steps. The preprocessing step involved expanding the dataset by randomly cutting fragments from images, calculating anchor frame sizes using the K-means clustering algorithm, and discarding anchor frames that were too small a priori. The existing post-processing block of the YOLO architecture was modified by giving more weight to false positives in the error function and using the non-maximum suppression algorithm. Baseline neural networks from the YOLOv4 and YOLOv5 architectures, each in two versions (pre-trained and not pre-trained on the MS COCO dataset), were used for comparison without any additional modifications. In the overgrown experimental field, multi-season aerial copter surveys and ground counts were conducted on several sample plots to gather data. Comparison of multi-season aerial photographs with ground-count data showed that the best images in terms of the percentage of visually identifiable trees were those taken during the snowy season and when there was no foliage. Using these images and some additional images, we manually created a dataset on which we trained and tested neural network models. The model we developed showed significantly better results (2 to 10 times better) on the mAP 0.5 metric compared to the alternatives we considered.