Evaluation of surface reflectance bandpass adjustment techniques

Abstract

There is an increasing number of on-orbit earth observation sensors in the solar domain (visible to middle infrared). While most the sensors are designed with apparent analogous bandpasses, they often exhibit differences in terms of Relative Spectral Response (RSR). Although the differences are well characterized, there is a need for techniques to adjust for the effect of these differences on surface reflectance after accounting for calibration, directional, and atmospheric effects. This research evaluated eight bandpass adjustment techniques using synthetic and actual remote sensing observations on seven widely used sensors. The scope of the study is global.The eight evaluated techniques were: a) spectral band adjustment factor technique (SBAF, i.e., one-coefficient regression), b) linear fit technique (i.e., two-coefficient regression), c) three Adaptive SBAF techniques (ASBAF), which are defined according to three different cost functions and apply a target-dependent SBAF, d) two artificial neural network approaches using (i) a single band (ANN-1B) and (ii) all bands (ANN-NB) as inputs, and e) a Classification and Regression Tree method (CART) using all bands as inputs. All these techniques require a calibration phase, which was carried out separately using three hyperspectral data sets: (i) USGS spectral measurements, (ii) simulated reflectance by the ProSail radiative transfer model, and (iii) real-world Hyperion observations. The full data set is publically available.The evaluation using the three hyperspectral data sets alone showed that the techniques with a single band as input (SBAF, LIN, and ANN-1B) generated higher uncertainty than the techniques with all bands as input (ASBAF, ANN-N, CART). The ASBAF techniques performed well with a few differences among the three ASBAF variants. The ANN-N was the technique that performed the best. The evaluation based on actual remote sensing data sets (Landsat-8 and Sentinel-2 data were used) suggested that all techniques reduced the deviation between the two data sets. Nonetheless, the deviation between the two data sets caused by factors other than RSR differences did not permit a clearcut comparison of the techniques. Some recommendations regarding the techniques to select for a particular application are provided.Finally, this research introduces the concept of RSR difference index (RDI) to quantify the spectral differences between two spectral bands.