Abstract
Desertification has hampered eco-environment sustainable development in arid, semi-arid, and dry sub-humid areas. However, the effect of the desertification process on vegetation carbon sources and sinks remains unclear in Central Asia. Based on Landsat images and cloud computing, this study applied and evaluated five machine learning methods (i.e., classification and regression tree, random forest, support vector machine, gradient tree boost (GTB), and naive bayes) and desertification difference index method to improve desertification estimation by integrating vegetation, soil, terrain, and climate conditions. According to the optimal method and net ecosystem production (NEP) model, we quantitatively explored vegetation carbon sources and sinks in Central Asia from 1990 to 2020, and then the effect of the desertification process on them was quantified under different aridity stress. The results showed that GTB method performs best on the test set and spatial pattern, which has higher overall accuracy (82.1 %) and Kappa coefficient (0.78) than other five methods. The desertification area has decreased by 8.58 % (341,643 km2) from 1990 to 2020. Among them, the severe and slight desertification areas decreased by 62.42 % and 32.11 %, respectively, while the moderate and high desertification areas increased by 24.6 % and 13.11 %, respectively. In particular, land restoration areas where the desertification restored one or above levels, accounted for 33.91 % of the total area. NEP in Central Asia showed an increasing trend at a rate of 0.54 g C m−2 yr−1 during 1990–2020, and the area passed the t-test (p < 0.05) was mainly located in Kazakh Steppe, Kazakh Uplands, and the edge of Tianshan Mountains. In general, restoring the land of degraded ecosystems has stored up 61.08 × 103 t carbon, accounting for 59.61 % of the total net change of NEP, but the fragile ecological environments in the existing desertification areas have been further aggravated.
