Regional Forest Mapping over Mountainous Areas in Northeast China Using Newly Identified Critical Temporal Features of Sentinel-1 Backscattering

Abstract

Sentinel-1 provides an extraordinary opportunity to explore the temporal behavior of backscattering of C-band synthetic aperture radar (SAR) due to its unique capability of successive observations every 12 days. This study reported new findings on the critical temporal features of Sentinel-1 backscattering over mountainous forested areas in northeast China and their application in regional forest mapping. Two interesting phenomena were discovered through the analysis of 450 scenes of images acquired by Sentinel-1A or Sentinel-1B over an area of 318,898.62 km2. The first phenomenon was that the dates of the largest drops of backscattering coefficients over forest and non-forest plots were different during the transition from autumn to winter. The largest drop of non-forest plots occurred around the date of the minimum temperature decreasing below 0 °C, while that of forest plots occurred around the date of the maximum temperature decreasing below 0 °C. The second phenomenon was that at the dates where these two drops occurred, the magnitude of the drop was negatively correlated with the forest canopy coverage for the first date and positively correlated for the second date. Based on these two phenomena, two methods for the forest mapping, referred to as the direct method and the indirect method, were proposed using only three dates of Sentinel-1 images, i.e., Date1: before the minimum temperature decreased below 0 °C, Date2: after the minimum temperature decreased below 0 °C but before the maximum temperature decreased below 0 °C, and Date3: after the maximum temperature decreased below 0 °C. The results showed that the overall accuracy of the forest map produced by the direct method was 93.60%, while that by the indirect method was 93.80%. Their accuracies were comparable with those of forest maps derived from publicly released land cover maps, which was approximately 94.42% for the best one. This study proposed a new way to do large-scale forest mapping in annually frozen regions using as few Sentinel-1 SAR images as possible.