Single tree detection using machine learning and remote sensing data

With an open source approach, an automated detection of single tree locations from remote sensing data was developed using a machine learning (ML) approach.



The development of a method for the automatic detection of individual trees using machine learning (ML) and aerial imagery poses several challenges. Here are some of the most important ones:

  • Training ML models: accurate training data is essential for training ML models. Data creation and quality control should be automated as far as possible.
  • Variability of tree species and environments: Trees vary greatly in size, shape and color, depending on the species and location. Different environments (urban, rural, forest, park) present different challenges, such as distinguishing trees from other green areas or structures.
  • Complex backgrounds and overlaps: In densely forested areas, trees can overlap, making them difficult to recognize individually. Shadows, buildings and other structures can also interfere with detection.
  • Scalability and computational requirements: ML models, especially deep neural networks, often require significant computational resources for training and inference, especially when processing large amounts of high-resolution images.
  • Generalizability of the model: A model that works well on a certain set of aerial images might perform worse on other images. The challenge is to develop a robust model that works reliably under different conditions and in different regions.

The company mundialis has fully implemented the project. The implementation took place in the following work steps:

  • Analysis of the customer’s inventory data
  • Creation of a methodology for single tree detection
    • Comparison of two approaches: Machine learning (random forest) vs. neural network
    • Machine learning as the method of choice
    • Automatic generation of training data
    • Post-processing: pixel-based classification output in object-based single trees
  • Derivation of parameters
    • Home position
    • Tree height
    • Distance to the next building/next tree
    • Crown diameter, volume, area
    • Differentiation between deciduous and coniferous trees (with their own classification)
  • Quality assurance based on reference tree register
  • Handover of software and result data as well as documentation to the client
  • Creation of a methodology for single tree detection using machine learning and deep learning
  • Implementation of the machine learning method for automated classification
  • Derivation of tree parameters
  • Provision of the results and the implemented software
Ruhr Regional Association (RVR)


Knowledge of the locations of individual trees and the development of the tree population is an important basis for various issues relating to green infrastructure, spatial observation and climate. Such information was previously only sporadically and incompletely available in the Ruhr metropolis, for example in the form of municipal tree registers. For an area-wide approach, an automated detection of individual tree locations from remote sensing data using a machine learning (ML) approach was developed as part of the project for the Ruhr Regional Association (RVR).