Attribution of climate and human activities to vegetation change in China using machine learning techniques

• Published in: Agricultural and forest meteorology 2020-11, Vol.294, p.108146, Article 108146
• Main Authors: Shi, Yu, Jin, Ning, Ma, Xuanlong, Wu, Bingyan, He, Qinsi, Yue, Chao, Yu, Qiang
• Format: Article
• Language: English
• Subjects: Attribution analysis; Climate change; Climatic zone; Human activity; Vegetation greening
• ISSN: 0168-1923; 1873-2240
• DOI: 10.1016/j.agrformet.2020.108146

•Machine learning has great potential for assessing vegetation restoration and greenness.•RFR model using meteorological factors explained 80% of NDVI variation•Climate and human activities affect greening in China's ecological engineering zones•Human activity was the most important factor associated with NDVI increase A series of policies and laws have been implemented to address climate change impacts in China since the 1980s. One of the most notable policies is ecological restoration engineering. However, there are many environmental factors that affect vegetation in the ecological restoration engineering zones. The relationships among different factors cannot be explained well by traditional statistical methods due to the existence of hidden non-linear features. Moreover, it is difficult to adopt threshold methods to accurately define vegetation areas fully, or to quantitatively analyze and assess the effects of climate factors and human activities on vegetation changes. The objective of this study was to determine vegetation area and distribution using Landsat TM/ETM/OLI images combined with a support vector machine (SVM) classification model. We analyzed the dynamic characteristics of vegetation area and greenness (NDVI, Normalized Difference Vegetation Index) in China's ecological restoration engineering zones from 1990 to 2015. Based on random forest regression (RFR) with a residual analysis method, the contributions of meteorological factors and human activities to vegetation greenness changes were quantitatively evaluated. Vegetation area and NDVI changed significantly in the study areas, increasing by more than 50% and 40%, respectively, from 1990 to 2015. Temperature, sunshine hours, and precipitation impacted vegetation greenness, which caused NDVI fluctuations in specific years. However, the NDVI increase was difficult to explain fully with meteorological factors. Using cross-validation, we predicted about 80% of the observed NDVI variation occurring from 1984 to 1994. Nine meteorological factors were related to vegetation growth, of which the average temperature, minimum temperature, maximum temperature, and average relative humidity were most critical. The combined effect of the nine climatic factors contributed less to NDVI increase than human activities. Human activity was the most important factor associated with NDVI increase, with contributions of more than 100% in most study areas. Human activities derived from national or local polic