Scale dependence of forest fragmentation and its climate sensitivity in a semi-arid mountain: Comparing Landsat, Sentinel and Google Earth data

Abstract

Landscape fragmentation is generally viewed as an indicator of environmental stresses or risks, but the fragmentation intensity assessment also depends on the scale of data and the definition of spatial unit. This study aimed to explore the scale-dependence of forest fragmentation intensity along a moisture gradient in Yinshan Mountain of North China, and to estimate environmental sensitivity of forest fragmentation in this semi-arid landscape. We developed an automatic classification algorithm using simple linear iterative clustering (SLIC) and Gaussian mixture model (GMM), and extracted tree canopy patches from Google Earth images (GEI), with an accuracy of 89.2% in the study area. Then we convert the tree canopy patches to forest category according to definition of forest that tree density greater than 10%, and compared it with forest categories from global land use datasets, FROM-GLC10 and GlobeLand30, with spatial resolutions of 10 m and 30 m, respectively. We found that the FROM-GLC10 and GlobeLand30 datasets underestimated the forest area in Yinshan Mountain by 16.88% and 21.06%, respectively; and the ratio of open forest (OF, 10% < tree coverage < 40%) to closed forest (CF, tree coverage > 40%) areas in the underestimated part was 2:1. The underestimations concentrated in warmer and drier areas occupied mostly by large coverage of OFs with severely fragmented canopies. Fragmentation intensity of canopies positively correlated with spring temperature while negatively correlated with summer precipitation and terrain slope. When summer precipitation was less than 300 mm or spring temperature higher than 4 °C, canopy fragmentation intensity rose drastically, while the forest area percentage kept stable. Our study suggested that the spatial configuration, e.g., sparseness, is more sensitive to drought stress than area percentage. This highlights the importance of data resolution and proper fragmentation measurements for forest patterns and environmental interpretation, which is the base of reliable ecosystem predictions with regard to the future climate scenarios.