Deadwood and decaying wood are the most important components for the biodiversity of boreal forests, and it has been estimated that around a quarter of all flora and fauna in Finnish forests depend on that. However, there is a severe lack of stand-level deadwood data in Finland, as the operational inventories either focus on the large-scale estimates or omit deadwood altogether. Unmanned Aerial Vehicles (UAVs) are the only method for remotely mapping small objects, such as fallen deadwood, as even the most spatially accurate commercial satellites provide 30cm ground sampling distance, compared to less than 5 cm that is easily achievable with UAVs.
In this work, we utilized Mask R-CNN to detect individual standing and fallen deadwood instances from RGB UAV imagery. We manually annotated over 14 000 deadwood instances from two separate study sites to use as the training and validation data, and also compared these data to field-measured deadwood data. Our models achieved test set Average Precision (AP) of 0.284 for the same geographical area the models were trained on, and AP of 0.220 for geographically distinct area used only for testing.
In addition to instance-level deadwood maps, we also estimated stand-level characteristics for the numbers of deadwood. In addition, we estimated the approximate total volume of fallen deadwood based on the annotated polygons. These stand-level features clearly show the borders of the conserved forest, and the volume estimations are able to distinguish between naturally formed deadwood hotspot and areas with logging remnants. The proposed method enables deadwood mapping for larger areas, complementing the traditional field work.
Much of the work relies heavily on https://github.com/jaeeolma/drone_detector, and instructions for its installation work here also.
Examples are using UAV RGB Orthomosaics from either Hiidenportti, Kuhmo, Eastern-Finland or Sudenpesänkangas, Evo, Southern-Finland. Hiidenportti dataset has a spatial resolution of around 4cm, and Sudenpesänkangas dataset has a spatial resolution of 4.85cm. Hiidenportti data contains 9 different UAV mosaics, and Sudenpesänkangas data is one single orthomosaic. From these data, we created virtual plots to use as a training and validation data for the models. From Hiidenportti, we constructed 33 virtual plots of varying sizes in such way that all 9m circular field plots present in the area were covered, and each field plot center had at least 45 meter distance to the edge of the virtual plot. For Sudenpesänkangas, due to the area and orthomosaic being larger, we extracted 100x100m plots in such way that each virtual plot contains only one circular field plot. In total, Hiidenportti data contained 33 virtual plots that cover 71 field plots, and Sudenpesänkangas data contaied 71 virtual plots.
Deadwood data that was used for training the models was manually annotated using QGIS software. We annotated all visible fallen deadwood trunks and standing deadwood canopies present in the virtual plots, and saved the results as geojson files. These data were then tiled and converted to COCO-format using functions from drone_detector. Sudenpesäkangas dataset consists of 5334 annotated deadwood instances, and Hiidenportti contains 8479 annotations.
All steps of the workflow are described either in scripts or Jupyter notebooks.
Presented in notebook 1_dataset_description_and_generation
Presented in notebook 2_comparing_annotations_and_field_data
Example of model training is presented in 3_mask_rcnn_model_training. All models were trained as a batch job instead of running the notebook.
The results were interpreted in two levels: patch level and virtual plot level. Patch level (512x512 pixel images) are show in 4_patch_level_results, and virtual plot level results with and without post-processing are show in 5_virtual_plot_level_results.
In addition to object detection metrics, the results were also compared based on forest charasteristics, such as estimated total volume of fallen deadwood, the distributions for length and DBH estimations and such. These are shown in 6_result_comparison_with_field_data.
Finally, we used our models to derive stand-level characteristics for deadwood: number of deadwood per hectare, number of standing deadwood per hectare, number of fallen deadwood per hectare and estimated volume of fallen deadwood per hectare, as show in 7_deriving_standwise_metrics_for_evo.
Model configs and weights are available here. Note that configs for WEIGHTS and OUTPUT_DIR need to be changed according to your needs. Example app running R101 backbone without TTA for image patches can be found here. More information is shown in 8_configs.
- Janne Mäyrä, Finnish Environment Institute SYKE
- Topi Tanhuanpää, University of Eastern Finland
- Anton Kuzmin, University of Eastern Finland
- Einari Heinaro, University of Helsinki
- Timo Kumpula, University of Eastern Finland
- Petteri Vihervaara, Finnish Environment Institute SYKE
This work is as a part of IBC-CARBON WP4.