Habitat mapping of coastal dunes with deep learning - Scripts & Data

<p><strong>Authors</strong>: Eva M. Lansu, Val&eacute;rie C. Reijers, Freek Dani&euml;ls, Rebecca James, Marjolijn J. A. Christianen, Tjisse van der Heide</p> <p>&nbsp;</p> <p><strong>Abstract</strong></p> <p><span lang="EN-GB">About one-third of the world's shoreline is defined by sandy coasts with developed dune ecosystems</span><span lang="EN-GB">. </span><span lang="EN-GB">These ecosystems </span><span lang="EN-GB">drastically degraded them due to anthropogenic pressures. </span><span lang="EN-GB">To develop strategic management that counteracts this degradation, it is essential to closely monitor ongoing habitat changes.</span><span lang="EN-GB"> Traditionally, coastal dune monitoring is based on field observations, which are labour intensive and costly. While automated analyses of aerial imagery could reduce monitoring efforts and enhance spatial coverage, t</span><span lang="EN-GB">o date, its application </span><span><span lang="EN-GB">has </span></span><span><span lang="EN-GB">remained limited to a single small-scale trial (&lt;2 km</span></span><span>^{<span lang="EN-GB">2</span>}</span><span><span lang="EN-GB">). </span></span><span><span lang="EN-GB">Here, we trained a Convolutional Neural Network to map the Dutch coastal dunes </span></span><span><span lang="EN-GB">(562 km²) </span></span><span><span lang="EN-GB">at 25 cm resolution using six habitat classes: bare sand, shrubs, fresh water, grass, broadleaf trees, and needleleaf trees. </span></span><span lang="EN-GB">Training the network on only RGB imagery resulted in predictions with 92% accuracy, 80% average recall and 70% precision. Model performance increased when the network was trained on all available data - RGB imagery, near-infrared, distance to sea, digital surface model, and canopy height - resulting in 95% accuracy, 88% averaged recall and 80% precision. Finally, we compared the predictions with 499 in-field observations across the Dutch coastal dunes and found 88% accuracy, 74% averaged recall and 62% precision. We used this model to create a map of the entire Dutch coastal dunes, which enables </span><span lang="EN-GB">rapid and precise </span><span lang="EN-GB">assessments of habitat diversity and extent</span><span lang="EN-GB">. As habitat and species diversity are intrinsically linked, our results showcase how automated image analysis can enable biodiversity monitoring on a national scale. </span></p> <p>==============================================</p> <p><strong>Methods</strong></p> <p>The analyses rely on the following datasets:</p> <ul> <li>Orthophoto mosaics including a near-infrared band (from <u><a href="http://geotiles.nl/">http://geotiles.nl/</a></u>)</li> <li>Digital surface model and a digital terrain model (from <a href="https://www.ahn.nl/">https://www.ahn.nl/</a>)</li> <li>A land-use map (from <u><a href="https://lgn.nl/basiskaart">https://lgn.nl/basiskaart)</a></u></li> </ul>