Multi-temporal Nonlinear Regression Method for Landsat Image Simulation

Optical remote sensing is limited in its potential for acquiring time-series images due to the restricted weather conditions in which it may be used. The proposed method simulates a Landsat image at a specific time and applies a multiple nonlinear regression equation that provides a higher degree of correlation with the observed data distribution than the commonly used multiple linear regression equation. In this study, Multivariate Adaptive Regression Splines (MARS) and Gaussian Process Regression (GPR) were considered as methods of multiple nonlinear regression. In addition to weather, environmental parameters such as temperature and humidity were added to analyze the input parameters in the regression process. Here, the GPR method of nonlinear regression results show significant improvement in Landsat image simulation. Furthermore, regardless of the season, simulation results using multiple parameter combinations showed the highest correlation with the reference images when temperature (ground), humidity, precipitation, visibility distance, Normalized Difference Vegetation Index (NDVI), and three types of radiation were applied. It was confirmed that introduction of Moderate Resolution Imaging Spectroradiometer (MODIS) products had little positive effects on the results. Thus, the GPR method defined here provides the best simulation results by employing multiple parameters in the calculation.