Search

Extensive use of drainage ditches in European boreal forests and in some parts of North America has resulted in a major change in wetland and soil hydrology and impacted the overall ecosystem functions of these regions. An increasing understanding of the environmental risks associated with forest ditches makes mapping these ditches a priority for sustainable forest and land use management. Here, we present the first rigorous deep learning–based methodology to map forest ditches at regional scale. A deep neural network was trained on airborne laser scanning data (ALS) and 1,607 km of manually digitized ditch channels from 10 regions spread across Sweden. The model correctly mapped 86% of all ditch channels in the test data, with a Matthews correlation coefficient of 0.78. Further, the model proved to be accurate when evaluated on ALS data from other heavily ditched countries in the Baltic Sea Region. This study leads the way in using deep learning and airborne laser scanning for mapping fine-resolution drainage ditches over large areas. This technique requires only one topographical index, which makes it possible to implement on national scales with limited computational resources. It thus provides a significant contribution to the assessment of regional hydrology and ecosystem dynamics in forested landscapes. [ABSTRACT FROM AUTHOR]

Comparisons between field data and available maps show that 64% of wet areas in the boreal landscape are missing on current maps. Primarily forested wetlands and wet soils near streams and lakes are missing, making them difficult to manage. One solution is to model missing wet areas from high-resolution digital elevation models, using indices such as topographical wetness index and depth to water. However, when working across large areas with gradients in topography, soils and climate, it is not possible to find one method or one threshold that works everywhere. By using soil moisture data from the National Forest Inventory of Sweden as a training dataset, we show that it is possible to combine information from several indices and thresholds, using machine learners, thereby improving the mapping of wet soils (kappa = 0.65). The new maps can be used to better plan roads and generate riparian buffer zones near surface waters. [ABSTRACT FROM AUTHOR]