Mapping forest cover dynamics in the Swiss Alps using 70 years of aerial imagery

The availability of high-resolution remote sensing imagery has enabled precise mapping of the forest cover at large scale. Since forest cover evolves due to land use change, climate and extreme events, understanding its past dynamics becomes crucial in a changing climate context. In this work we analyze historical aerial imagery acquired in Switzerland since 1946 [2] for high-resolution forest mapping. We focus on the 1500–2500m a.s.l. altitude range in the Valais and Vaud Alps, where agricultural land abandonment and climate change have caused forest cover changes.

The times series are composed of single-band, panchromatic images until year 1998, then RGB images up to the year 2020, the last acquisition date over our study area. As a reference for the forest cover in 2020, we use the Topographic Landscape Model SwissTLM3D [1]. For previous years, we plan to manually generate labels to evaluate our results.

We frame forest mapping as a multi-temporal semantic segmentation task: given a time series of images, we predict a map for each image
attributing every pixel to the class "forest" or "non-forest". To solve this task, we develop a deep learning model composed of:

• a segmentation module, trained with the images and labels from the year 2020;
• a temporal module, which takes consecutive features generated by the segmentation module and outputs a multi-temporal segmentation map. This module is trained using a Mean Squared Error (MSE) loss enforcing temporal consistency.

We analyze predictions obtained with three models, each one containing one or two of the modules described above. We observe that using the full spectral information of the input images leads to a better delineation of forest borders for both old and recent images (Table 1, Figure 1). By adding the temporal module, the accuracy on the last image is practically unchanged (Table 1), while temporal consistency along the time series is improved (Figure 2).

 

Table 1: Segmentation scores for the year 2020 on the validation set, for all pixels and for pixels under 10m distance of forest borders

Model	# inputs	Temporal module	Mean F-1 score (all)	Mean F-1 score (forest borders)
Mono-temporal grayscale	1	no	0.86	0.63
Mono-temporal RGB	3	no	0.89	0.72
Multi-temporal RGB	3	yes	0.88	0.72

 

 

 

Our method is currently not suited for abrupt forest loss, and is prone to error spreading from previous predictions. Future work will consist in designing a temporal consistency loss that better reflects known dynamics of the forest cover, in order to obtain a more accurate segmentation for the oldest images and encourage physical consistency across time.

References
[1] Swisstopo. SwissTLM3D. https://www.swisstopo.admin.ch/en/geodata/landscape/tlm3d.html [Online; accessed 06.01.2023].
[2] Swisstopo. Orthoimages. https://www.swisstopo.admin.ch/en/geodata/images/ortho.html [Online; accessed 06.01.2023].