Identification of Bias in Satellite Measurements Using its Geospatial Properties

Identification of bias in satellite retrievals is a challenging task as the bias, by definition, is the difference between the average of measurements made on the same object and its true value. Given so, the identification of bias requires knowledge of the true value, which is sometimes impossible to obtain except through actual measurements. Two common types of approaches are deployed to avoid this circularity: 1) either measurements are compared with a secondary measurement platform, which often only partially overlaps with primary measurements in space–time, or 2) by examining the internal properties of the measurements as different sorts of inconsistencies under specific circumstances can point to a bias. In this letter, we use the recent advances in space and space–time modeling and show that inconsistencies in interpolated satellite retrievals using spatial-only and spatio-temporal kriging point to possible bias in the measurements due to specific spatio-temporal properties of the field of bias, which often mimics the spatio-temporal properties of the causal phenomena. We suggest a new data quality ranking system based on the absence of this inconsistency. We demonstrate the method using the Global Ozone Monitoring Experiment-2 (GOME-2) satellite retrievals of chlorophyll-induced fluorescence (SIF) and Greenhouse Gases Observing Satellite (GOSAT) retrievals of XCO₂.

I. Introduction

Bias is a quantitative term that describes the difference between the average of measurements made on the same object and its true value. A measurement process is biased if it systematically overstates or understates the true value of the measurement. The hermeneutical aspects of bias are specific to scientific fields. For example, in biomedical research studies at least 12 types of bias are identified [1], which have substantially different causes and properties than are found in biases for other types of research. From the definition of bias, it follows that the basic prerequisite for the correction of bias is the knowledge of a true value. The relationship between bias and the true value is somewhat circular: the removal of bias is required to infer the true value from measurements, but the knowledge of the true value is required to quantify the bias, which then makes the measurement itself obsolete. In practice, this circularity is overcome in two conceptually different ways: 1) comparing external measurement platforms or strategies that act as gold standards (e.g., [2]) or 2) using internal properties of the measured/sampled data (e.g., [3], [4]). The problem associated with the first approach is that alternative data sets (“gold standards”) are, at most, transiently coincident and collocated with the measurement and cannot guarantee unbiasedness outside the spatio-temporal comparison window.