Integration of ground-based and remote sensing data with deep learning algorithms for mapping habitats in Natura 2000 protected oak forests

Abstract

Landscape changes caused by climate change require new methods for forest research, analysis, mapping, and monitoring. This study aims to combine ground-based and remote sensing data utilising deep learning techniques to map protected forest habitats and communities within the Natura 2000 network. The study also seeks to evaluate the accuracy of this approach, specifically in oak-dominated forests, as well as identify the optimal time period within a year for effective habitat identification.

Using the specialised software NaturaSat, automated segmentations were performed based on the coordinates of phytosociological relevés and forest strands defined in database. Oak-dominated forest habitats were differentiated solely through multispectral data obtained from Sentinel-2 satellites. A dataset was selected for the training of a deep learning algorithm called the Natural Numerical Network on the basis of the analysis results. This algorithm aims to create a prediction map of habitats dominated by Quercus cerris, which is also known as the relevancy map.

Through the utilisation of the Natural Numerical Network, a training accuracy of 95.24% was achieved. Field validation, which was conducted at randomly generated locations within the relevancy map, yielded an accuracy of 98.33%. The most distinguishing differences in band characteristics between the two oak-dominated habitats were observed during the autumn months.

This study presents a framework that integrates terrestrial and remote sensing data. This method can serve as a basis for mapping forest habitats and observing changes related to climate change. Moreover, it contributes to the documentation of nature conservation and the mapping of landscapes.