EGU21-1698
https://doi.org/10.5194/egusphere-egu21-1698
EGU General Assembly 2021
© Author(s) 2021. This work is distributed under
the Creative Commons Attribution 4.0 License.

Automatic tree species classification by using field data, image analysis and deep learning techniques in riparian forests

Sarah Kentsch1, Maximo Larry Lopez Caceres1, and Yago Diez2
Sarah Kentsch et al.
  • 1Yamagata, Faculty of Agriculture, Tsuruoka, Japan (sarahkentsch@gmail.com)
  • 2Faculty of Science, Yamagata university, Yamagata city, Japan

Mixed forests are still little understood ecosystems. Their structure and composition are not well known and not clearly classified. In times of climate change, monitoring of forests is becoming increasingly important. Forest stands were usually researched by field work, which requires high costs and man-power. Field surveys are further only conducted in small patches of the forests, which does often not represent the whole forest. For mixed forests, usually only a dominate species is mentioned but the forests are not classified further. The greater need of better methods with high accuracies to detect and classify tree species in the forest encouraged this study.

UAVs have been proven to be an efficient tool to conduct automatic field surveys in forestry applications. These easy-to-use and cheap tools are able to gather images with a high resolution. Image processing with image analysis and deep learning techniques is an emerging part in forestry investigations. Therefore, we combined manual field surveys, image analysis and automatic classifications in our study.

The forests, we were investigating, are riparian forests in Shonai area, Japan, which are classified as mixed forests. 7 sites were chosen and field surveys were conducted. Most of the sites are located in flat areas, but 3 sites are located on slopes, where the access is difficult and field work barely possible. We imaged all sites in different seasons with UAVs and performed image analysis with computer vision and ArcGIS methods. Trees were detected and classified manually and automatically. A comparison of all applied methods was drawn, evaluated and will be provided.

Our first results are promising to characterize forests in a new dimension. We will provide detailed information about tree species composition, tree locations and forest structures. Mixed forests can be deeper analysed by maps of dominate and subdominant tree species. Area calculations for tree canopies will be highlighted for the main tree species. We will provide winter images for tree counting in heavy snowfall regions and classification accuracies of deep learning techniques.  

How to cite: Kentsch, S., Lopez Caceres, M. L., and Diez, Y.: Automatic tree species classification by using field data, image analysis and deep learning techniques in riparian forests, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-1698, https://doi.org/10.5194/egusphere-egu21-1698, 2021.

Displays

Display file