Understanding ecosystem services of detailed forest and wetland types using remote sensing and deep learning...

Spanning both temperate and sub-frigid zones, Northeast China boasts typical boreal forests and abundant wetland resources. Because of these attributes, the region is critically significant for global climate regulation, carbon sequestration, and biodiversity preservation. While existing research explores the ecosystem service (ESs) functions of different land cover types, a thoroughly in-depth investigation into the ESs of detailed forest and wetland types is essential. This study addresses this deficiency by combining remote sensing and deep learning techniques, employing a lightweight convolutional neural network (CNN) model and a decision tree for the large-scale classification of forests and wetlands. The ESs of various forest and wetland types-encompassing habitat quality, carbon stock, and soil retention-were assessed during two periods (2008 and 2018) in Heilongjiang Province. Key factors determinants of ESs were identified using the Geodetector tool. The results indicated an overall accuracy of 0.77 in 2008 and 0.78 in 2018 for forest type classification, and 0.88 in 2008 and 0.87 in 2018 for wetland type classification. In particular, the transition from mixed broadleaf forests to mixed coniferous-broadleaf forests dominated changes from 2008 to 2018, probably due to natural succession. Among forest types, Mongolian oak forests exhibited the highest carbon stock and soil retention capacity owing to their rapid growth and deep root systems. Mixed broadleaf forests exhibited superior habitat quality, suggesting minimal disturbance. Habitat quality, carbon stock, and soil retention were found to be significantly influenced by human activity, atmospheric quality, and topographic factors, respectively. By leveraging remote sensing and deep learning methodologies, this study offers a comprehensive analysis of forests and wetlands, elucidating the nuanced ecosystem roles of specific forest and wetland types.