Mapping Italian Alpine Peatlands Using Multisource Satellite Imagery and Machine Learning Approach

While we are aware that the Italian Alps host thousands of small peatlands, the precise estimate remains uncertain due to the absence of a comprehensive map. These ecosystems are extremely valuable because, in addition to storing large amounts of organic carbon, they provide many other ecosystem services. They regulate water flow, retaining it during wet seasons and releasing it during dry periods. Furthermore, they purify water by retaining nutrients such as nitrogen and phosphorus and provide water to wildlife even during droughts. Moreover, they are characterized by high biodiversity, serving as habitats for several endangered species.

Conventional approaches to mapping peatlands typically involve surveys characterized by long update cycles and considerable costs. Some remote sensing approaches, such as UAV and aerial photography, have the disadvantages of being weather dependent, and have high costs and limited coverage. In contrast, satellite remote sensing imagery presents several advantages, including broad coverage, cost-effectiveness, and frequent temporal resolution. Hence, our research emphasizes the mapping of Alpine peatlands by integrating multiple remote sensing datasets and employing machine learning algorithms. The spatial distribution of Alpine peatlands shows a correlation with topographic and hydrological conditions. These peatlands, averaging around 1 hectare in size, exhibit distinctive vegetation, topographic, and hydrological characteristics compared to non-peatland regions. Therefore, the differentiation in these features extracted from remote sensing imagery stands as a critical factor for identifying peatlands.

We present the results of integrating Sentinel-2 optical data, Sentinel-1 radar imagery, and the CLO-30 from the Copernicus digital elevation model (DEM) through the Google Earth Engine (GEE) platform. This integration aims to map Alpine peatlands employing a pixel-based Random Forest algorithm. We focus on a section of the Adige River basin, located within the Trentino Alto-Adige Region in Italy. Within this area, we collected and updated an inventory of 157 peatland sites, divided into two groups. One subset was used to calibrate the algorithm, while the other served to validate the results. Several sets of features were extracted from the multi-source remote sensing dataset. The findings suggest that both the DEM itself and the topographic features derived from it contributed most significantly to the classification results. Hydrological connectivity was also found to be a significant feature, probably due to the crucial role that water flow and retention play in the establishment and sustainability of peatlands. A key finding is the impact of these features surpassed that of optical and radar data in enhancing the accuracy of the classification. Since our peatland mapping methodology is implemented on the GEE platform and uses freely available datasets, it can be applied across the entire Alpine region and in other mountainous areas worldwide.