Interpine has developed a complete solution for accurate tree counts and stocking assessments, using a machine learning (ML) model to detect and count trees in aerial imagery. This solution requires little effort to train, and provides highly accurate results.

Interpine can undertake the drone survey, or the client can collect the aerial imagery with their own drone, following our recommended specs, then send the data to us. Provided the data meets a minimum quality, the model is then trained on the imagery, processing is carried out, and a tree count is produced

Figure 1. Image from pre-thin stocking assessment.

Figure 2. Image from nursery seedling count. 

For other commercial tools, building a good machine learning model requires manually labeling a large dataset (usually 1000-2500 trees), which can take up to 8 hours. With Interpines ML model, we only need to label 100-200 trees to achieve a similar result.

In a recent trial (Figure 3), the ML model outperformed a LiDAR data based algorithmic approach in a complex plot with overhanging tree canopies. LiDAR data has traditionally given better results than aerial imagery alone, as the height value of the 3D point cloud helps with the tree detection, so it is encouraging to see the how well the ML model performs in these situations.  

Figure 3. Comparason of LiDAR data based tree detection (pink cross) and Interpines ML model (red dots). 

Results of Interpines ML tree detection can be visualised in a map showing stems per hectare (SPH), providing a useful tool for forest managers to evaluate stocking density, or the quality of thinning operations at a stand level. The image below shows the results of a ML post-thin tree detection assessment, with SPH thresholds represented through the following colours: 

Yellow: < 425 SPH

Orange: 425 – 525 SPH

Green: 525-600 SPH  (target for thinning)

Red: > 600 SPH (rework)

Figure 4. Results of ML tree detection for a post thinning assessment, showing stems per hectare.