GIS & Geospatial
Automatic tree recognition from orthophotos
We built a system that automates forest inventory. Using geoinformation technologies and machine learning, it locates trees accurately and quickly over large areas.
- Industry
- GIS & Geospatial
- Location
- Germany
- Team
- GisPoint covered: the project manager planned, coordinated, and controlled the project. GIS specialists collected the data, developed the algorithms, and tested and implemented the machine learning models in ArcGIS Pro. The quality controller assessed the accuracy of the results against the quality standards.
- Stack
- The project runs on ArcGIS Pro with deep learning tools, which is what gets tree locations found automatically at the accuracy required. The output is a detailed map of forest areas for forest management.
About the client
German Federal Forestry Administration
The challenge, and what closed it
Technical challenge
Finding tree locations automatically with ArcGIS Pro is not simple. What we ran into: Noise and artifacts: shadows, clouds, atmospheric phenomena, and sensor noise interfere with segmentation algorithms. Spectral characteristics: tree species differ, so does their condition (healthy, diseased, young, old), and so does the weather, which moves the spectral signature enough to make classification hard. Computing resources: the data volumes and the machine learning models needed hardware to match. Training sample preparation: building a training set with accurate annotations was labour-intensive. These come with projects of this kind, and each one had to be worked through.
What we did
Automatic tree recognition from orthophotos. Using machine learning algorithms for data analysis. Integrating algorithms into GIS applications.
Why us
The project ran on ArcGIS Pro, which handled both the data and the compute. The stages: Development of the classification algorithm: ArcGIS Pro and its machine learning tools were used to build an algorithm that classifies image pixels and locates each tree. Model testing and validation: the model’s accuracy was tested on an independent sample against manual counting and postal surveys. Map generation: the classification results became digital maps of each tree location.
Results
The project resulted in a high accuracy of tree coordinates (90%), which allows us to obtain detailed information about their location in the forests. A significant amount of data (600 hectares) was processed and detailed tree distribution maps were created.
What we delivered
Automatic tree recognition from orthophotos.
Using machine learning algorithms for data analysis.
Integrating algorithms into GIS applications.
Tech stack
- Survey & CAD
- The project runs on ArcGIS Pro with deep learning tools, which is what gets tree locations found automatically at the accuracy required. The output is a detailed map of forest areas for forest management.
