Forest inventories are essential for sustainable forest management and health monitoring. Image-based surveys using Unmanned Aerial Vehicles (UAVs) are increasingly adopted for this purpose. DeepForest, a deep learning model pre-trained on large annotated datasets, enables scalable and cost-efficient tree detection in the resulting imagery. While it is known that fine-tuning DeepForest to specific target sites improves its performance, optimal fine-tuning strategies remain unclear.
In this study, we investigate the fine-tuning of DeepForest on a small dataset from a temperate mixed forest. We examine the influence of four parameters on model performance: (1) the method used to label training data (manual labeling, manual correction of DeepForest predictions, and automatic labeling using 3D point cloud segmentation), (2) input image resolution, (3) the number of fine-tuning epochs, and (4) the size of the training dataset.
Our results show that the out-of-the-box DeepForest model performs poorly on the target dataset but improves significantly with fine-tuning, even when using limited data and few training epochs. Manual and automatic labeling achieve similarly high performance, both clearly outperforming the manual correction approach. While the model initially performs better on lower-resolution images, this difference diminishes after fine-tuning on higher-resolution inputs. We observe no clear benefit from increasing the training set size, possibly due to the low diversity of our test dataset.
Overall, our findings confirm that DeepForest is well-suited as a base model for tree detection in high-resolution UAV imagery and can be adapted to new environments with low effort.