A random forest method for constructing long-term time series of nighttime light in Central Asia

Abstract

The nighttime light (NTL) data from Defense Meteorological Satellite Program’s Operational Linescan System (DMSP-OLS) and Suomi National Polar-orbiting Partnership Satellite’s Visible Infrared Imaging Radiometer Suite (NPP-VIIRS) have been extensively used to characterize spatiotemporal variations of human activities on the ground. However, inconsistency of these two datasets hampers their applications for long-term analysis. Hence, the current paper managed to introduce the Random Forest model (RF) for constructing a long-term time series of NTL with considering topographical and climatic influences. Performance of the RF was compared with that of several common curve estimation methods, namely linear regression, quadratic regression, cubic regression, logistic regression and Biphasic Dose Response (BDR) model based on two datasets in overlapping years of 2012 and 2013. Relative importances of natural factors were quantified with the initial RF model with all variables inputted into, and four typical RF models with relatively better performance were screened on this basis. The results showed that VIIRS data, latitude and longitude, elevation and precipitation were inputted into RF models in turn and their simulation accuracies gradually increased. According to accuracy indicators, RF models with topographical and climatic factors as input variables were far superior to the BDR which performed better than other curve estimation methods. In particular, the RF model with VIIRS data, latitude and longitude as input variables had higher simulation accuracy and less computation amounts was recommended as the best-fit RF model for the inter-calibration of NTL data in Central Asia. Afterwards, two continuous time series of NTL from 1992 to 2018 generated respectively with the best-fit RF and BDR were compared and results indicated that calibrated data with the RF were more consistent with DMSP-OLS data and Landsat 8 satellite images. Besides, the scope and brightness of NTL with RF had a significant positive correlation with log-transformed socioeconomic factors, confirming the availability of RF. The current paper offers a new thought to settle inconsistency of NTL data from two sensors with considering influences of natural factors, which is expected to help understand and explore complex interactions between natural environment and human activities in the arid Central Asia.