Forest under watch: Remote-sensing data processing to monitor forest structures

Remote-sensing technologies are widely used to study forests but integrated toolkit for data processing is still lacking. FuW aims at proposing comprehensive workflow to estimate forest parameters from the most recent remote-sensing data.

Ausgangslage

Forest ecosystems are important carbon sinks worldwide, playing an essential role in climate mitigation. However, the future of forests is currently uncertain due to climate change, threatening forest ecosystem services (FES, e.g., wood production protection against rockfalls, carbon sequestration) and leading to a rethinking of forest management. Yet, the lack of data on forest conditions for modeling and for carbon balance (CB) estimations requires rapid improvements in forest monitoring. The emergence of remote sensing-technologies and data processing workflows assisted by Artificial Intelligence (AI) displays promising perspectives for sustainable forest resource management policies. Most advanced models have already proved successful but only a few studies provided detailed information, limiting practical application. The aim of the FuW project is therefore to propose an integrated tool based on remote sensing and inventory data to calculate FES and CB estimates, and in extenso, forest health. Through case studies, the following objectives are pursued: 1. to test different models and statistical approaches and assess their efficiency and accuracy; 2. to integrate the most robust models into a data processing process for practical use, and 3. to extend the developed approach to larger study areas.

Vorgehen

The main target of this project is to predict the standing volume at plot and tree scales. To do so, we used two approaches based on processing airbone 3D point cloud of forest stands located in Swiss National Forestry Inventory (NFI) surveys. We firstly derived aggregated data from 3D point cloud and satellite imagery. We used field-knowledge to detect the most important forest or tree features contributing to the standing volume prediction and to do feature engineering. Dendrometric variables from NFI were then used to complete the dataset and to get the reference values. We built baseline, i.e. linear and Random Forests, and ensemble learning models using boosting gradient algorithms. Their score performances and their ability to cope overfitting were then compared to select the best model. n a second step, we tested Deep-Learning models from ForestSens tools and no-clustering algorithms directly on the 3D point cloud to segment individual trees.

Ergebnisse

When using aggregated data as inputs to predict standing volume at plot level, the best score without overfitting was reached with Machine-Learning models (R2= 60%). The mean height of the point cloud contributed the most to the predictions. We found that the addition of the satellite data did not improve significantly model performances. Regarding the tree segmentation from 3D point cloud, the Deep-Learning models of ForestSens tools were able to cope the less complex 3D point cloud of Swiss forests, while DBSCAN algorithms showed their limits if not used in a more developed pipeline. Time to parametrize the tools should not be underestimated to assure successful tree segmentation. Our results suggested that the variability of Swiss forest structure, tree architecture and typology induced additional set of information that are not currently well handle in the published workflows. Better performances would need more time and effort, especially in data preparation and model calibration.

Kontakt

Dr. Estelle Noyer

• Wissenschaftliche Mitarbeiterin
• +41 31 848 62 14
• E-Mail anzeigen

• Zum Profil

Dieses Projekt leistet einen Beitrag zu den folgenden SDGs

• 
•