Day 4

Detailed paper information

Back to list

Paper title Interpretable mapping of high-altitude forest in the Swiss Alps
  1. Thiên-Anh Nguyen Ecole Polytechnique Federale de Lausanne - EPFL Speaker
  2. Benjamin Kellenberger Ecole Polytechnique Federale de Lausanne - EPFL
  3. Devis Tuia EPFL
Form of presentation Poster
  • C1. AI and Data Analytics
    • C1.07 ML4Earth: Machine Learning for Earth Sciences
Abstract text In high mountain regions such as the Swiss Alps, the expansion of forest towards high altitudes is limited by extreme climatic conditions, particularly related to low temperatures, thunderstorms or snow deposition and melting [1]. All these factors, together with human land use planning, shape the upper forest limit, which we refer to as the alpine treeline. The complex topography of such regions and the interplay of a large number of drivers makes this boundary highly fragmented. Remote sensing-based land cover products tend to oversimplify these patterns due to insufficient resolution or to a need for excessively labor-intensive labeling. When higher resolution imagery is available, the accuracy of automated forest mapping methods tends to drop close to the treeline due to fuzzy forest boundaries and lower image quality caused by complex topography [3]. High- resolution maps of forest that are specifically tailored for the treeline ecotone are thus needed to accurately account for this complexity.

Mapping forest implies formulating a clear definition of forest. A large number of such definitions exist, most of them based on tree height and tree canopy density thresholds, but also spatial criteria (area, width/length), as well as structural form (e.g. shrubs) and land use. The position of the treeline can vary greatly depending on the chosen definition. While traditional machine learning methods are able to reach high accuracy with respect to the training labels, they do not provide additional information about underlying relevant variables and how they relate to the final map. For this reason, they are often referred to as ‘black boxes’. The results of such models are implicitly linked to a forest definition through the training labels, if those are accurate enough and based on a fixed definition, but spatially-explicit and disentangled concepts are missing to explain the model’s decisions in terms of forest definition.

To tackle the high-altitude forest mapping task, we propose a deep learning-based semantic segmentation method which uses optical aerial imagery at 25 cm resolution over the 1500-2500 m a.s.l. altitude range of the Swiss Alps and forest masks from the SwissTLM3D landscape model, which provides a spatially explicit, detailed characterization of different types of forest [2]. After proper training, the model yields a fine-grained binary forest/non-forest map, and is also able to classify the forest into three types (open forest, closed forest, shrub forest), despite noisy labels and heavy class imbalance. We obtain an overall f-1 score above 90% with respect to the SwissTLM3D labels both for the binary task and when including the forest type classification into the task.

From this baseline model, we then developed an interpretable model which estimates intermediate forest definition variables for each pixel, explicitly applies a target forest definition and highlights systematic discrepancies between the target forest definition and the noisy training labels. These pixel-level explanations complement the resulting forest map, making the model’s decision process more transparent and closely related to relevant and widely-used variables characterizing Swiss forests.


[1] George P. Malanson, Lynn M. Resler, Maaike Y. Bader, Friedrich-Karl Holtmeier, David R. Butler, Daniel J. Weiss, Lori D. Daniels, and Daniel B. Fagre. Mountain Treelines: A Roadmap for Research Orientation. Arctic, Antarctic, and Alpine Research, 43(2):167–177, 5 2011.
[2] Swisstopo. SwissTLM3D. tlm3d.html, 2021. [Online; accessed 04.11.2021].
[3] Lars Waser, Christoph Fischer, Zuyuan Wang, and Christian Ginzler. Wall-to-Wall Forest Mapping Based on Digital Surface Models from Image-Based Point Clouds and a NFI Forest Definition. Forests, 6(12):4510–4528, 12 2015.