Transferability of Economy Estimation Based on DMSP/OLS Night-Time Light

Despite the fact that economic data are of great significance in the assessment of human socioeconomic development, the application of this data has been hindered partly due to the unreliable and inefficient economic censuses conducted in developing countries. The night-time light (NTL) imagery from the Defense Meteorological Satellite Program’s Operational Linescan System (DMSP/OLS) provides one of the most important ways to evaluate an economy with low cost and high efficiency. However, little research has addressed the transferability of the estimation across years. Based on the entire DN series from 0 to 63 of NTL data and GDP data in 31 provinces of mainland China from 2000 to 2012, this paper aims to study the transferability of economy estimation across years, with four linear and non-linear data mining methods, including the Multiple Linear Regression (MLR), Local Weighted Regression (LWR), Partial Least Squares Regression (PLSR), and Support Vector Machine Regression (SVMR). We firstly built up the GDP estimation model based on the NTL data in each year with each method respectively, then applied each model to the other 12 years for the evaluation of the time series transferability. Results revealed that the performances of models differ greatly across years and methods: PLSR (mean of R2cv = 0.84) and SVMR (mean of R2cv = 0.86) are superior to MLR (mean of R2cv = 0.72) and LWR (mean of R2cv = 0.75) for model calibration; only PLSR (mean of R2 p = 0.88, mean of RPD = 1.8) holds a strong transferability among different years; the frequency of three DN sections of (0–1), (4–16), and (57–63) are especially important for economy estimation. Such results are expected to provide a more comprehensive understanding of the NTL, which can be used for economy estimation across years.