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

<p><span dir="ltr" role="presentation">The availability of high-resolution remote sensing imagery has enabled precise </span><span dir="ltr" role="presentation">mapping of the forest cover at large scale. Since forest cover evolves due </span><span dir="ltr" role="presentation">to land use change, climate and extreme events, understanding its past dynamics </span><span dir="ltr" role="presentation">becomes crucial in a changing </span><span dir="ltr" role="presentation">climate context. </span><span dir="ltr" role="presentation">In this work we analyze historical aerial imagery acquired</span><span dir="ltr" role="presentation"> in </span><span dir="ltr" role="presentation">Switzerland since 1946 [2] for high-resolution forest mapping. We </span><span dir="ltr" role="presentation">focus on the 1500&#8211;2500m a.s.l. altitude range in the Valais and Vaud Alps, where </span><span dir="ltr" role="presentation">agricultural land abandonment and climate change have caused </span><span dir="ltr" role="presentation">forest cover changes.</span></p> <p><span dir="ltr" role="presentation">The times series are composed of single-band, panchromatic images until </span><span dir="ltr" role="presentation">year 1998, then RGB images up to the year 2020, the last acquisition date over our </span><span dir="ltr" role="presentation">study area. As a reference for the forest cover in 2020, we use the Topographic </span><span dir="ltr" role="presentation">Landscape Model SwissTLM3D [1]. For previous years, w</span><span dir="ltr" role="presentation">e plan to manually generate labels to evaluate our </span><span dir="ltr" role="presentation">results.</span></p> <p><span dir="ltr" role="presentation">We frame forest mapping as a multi-temporal semantic segmen</span><span dir="ltr" role="presentation">tation task:</span> <span dir="ltr" role="presentation">given a time series of images, we predict a map for each image</span><br role="presentation" /><span dir="ltr" role="presentation">attributing every </span><span dir="ltr" role="presentation">pixel to the class "forest" or "non-forest". To solve </span><span dir="ltr" role="presentation">this task, we develop a deep learning model composed of:</span></p> <ul> <li><span dir="ltr" role="presentation">a segmentation module, trained with the images and labels from the year 2020;</span></li> <li><span dir="ltr" role="presentation">a temporal module, which takes consecutive features generated by the segmentation module and outputs a </span><span dir="ltr" role="presentation">multi-temporal segmentation map. This module is trained using a Mean </span><span dir="ltr" role="presentation">Squared Error (MSE) loss enforcing temporal consistency.</span></li> </ul> <p><span dir="ltr" role="presentation">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).</span></p> <p>&#160;</p> <table class="co_table co_table_bordered w-100" border="1"><caption>Table 1: <span dir="ltr" role="presentation">Segmentation scores for the year 2020 on the validation set, for all </span><span dir="ltr" role="presentation">pixels and for pixels under 10m distance of forest borders</span></caption> <tbody> <tr> <td>Model</td> <td># inputs</td> <td>Temporal module</td> <td>Mean F-1 score (all)</td> <td>Mean F-1 score (forest borders)</td> </tr> <tr> <td>Mono-temporal grayscale</td> <td>1</td> <td>no</td> <td>0.86</td> <td>0.63</td> </tr> <tr> <td>Mono-temporal RGB</td> <td>3</td> <td>no</td> <td>0.89</td> <td>0.72</td> </tr> <tr> <td>Multi-temporal RGB</td> <td>3</td> <td>yes</td> <td>0.88</td> <td>0.72</td> </tr> </tbody> </table> <p>&#160;</p> <p><img src="" alt="" width="514" height="508" /></p> <p>&#160;</p> <p><img src="" alt="" width="587" height="334" /></p> <p>&#160;</p> <p><span dir="ltr" role="presentation">Our method is currently not suited for abrupt forest loss, and is prone to </span><span dir="ltr" role="presentation">error spreading from previous predictions.</span> <span dir="ltr" role="presentation">Future work will consist in </span><span dir="ltr" role="presentation">designing a temporal consistency loss that better reflects known dynamics of </span><span dir="ltr" role="presentation">the forest cover, </span><span dir="ltr" role="presentation">in order to obtain a more accurate segmentation for the oldest images and </span><span dir="ltr" role="presentation">encourage physical consistency across time.</span></p> <p><span dir="ltr" role="presentation"><strong>References</strong><br role="presentation" />[1] Swisstopo. SwissTLM3D. https://www.swisstopo.admin.ch/en/geodata/landscape/tlm3d.html [Online; accessed 06.01.2023].<br role="presentation" />[2] Swisstopo. Orthoimages. https://www.swisstopo.admin.ch/en/geodata/images/ortho.html [Online; accessed 06.01.2023].</span></p>