Frozen ground degradation profoundly impacts the hydrology, ecology and human society on the Tibetan Plateau (TP) and the downstream regions. The spatial distribution and potential changes of permafrost and maximum thickness of seasonally frozen ground (MTSFG) on the TP is of great importance and needs more in-depth studies. This study maps the permafrost and MTSFG distribution in the baseline period (2003-2010) and in the future (2040s and 2090s) with 1-km resolution. Logistic regression (LR), support vector machine (SVM) and random forest (RF) are validated using 106 borehole observations and proved to be applicable in estimating permafrost distribution. According to the majority voting results of the three algorithms, 45.9% area of the TP is underlain by permafrost in the baseline period, and respectively 25.9% and 43.9% of the current permafrost will disappear by the 2040s and the 2090s projected by mean of the projections from the five General Circulation Models under the Representative Concentration Pathway 4.5 scenario. SVM performs better in spatial generalization than RF based on the results of nested cross validation. According to the MTSFG results derived from SVM, the most dramatic decrease in MTSFG will occur in the southwestern TP, which is projected to exceed 50 cm in the 2090s compared with the baseline period. This study introduces the statistics and machine learning algorithms to frozen ground estimation on the TP, and the high resolution permafrost and MTSFG maps produced by this study can provide useful information for future studies on the third pole region.
Keywords: Climate change; Machine learning; Maximum thickness of seasonally frozen ground; Permafrost degradation; Tibetan Plateau.
Copyright © 2018 Elsevier B.V. All rights reserved.