DOGO Warn Levels: An Upgrade to Quality Flags for the OCO-2 Data Products

All large observational datasets contain varying quality data due to instrument limitations, retrieval software imperfections, and confounding (poorly modeled) effects in the system being observed. Traditionally, we image large percentages of data that are of limited or no scientific use, so called "bad data." The previous solution was for an instrument team to create a binary filter (quality flag) to remove unwanted observations before any analyses are performed. However, quality flags assume fully separable "good"/ "bad" data and produce a fixed percentage of data that may still be too high or low for a particular user's needs. Additionally, atmospheric soundings are confounded by clouds and aerosols that continuously vary in optical depth and size scale without a sharp, simple boundary between "good" and "bad," breaking the data quality concept. OCO2 has for the first time at JPL/NASA developed an alternative to quality flags: the award-winning DOGO (Data Ordering Genetic Optimization) system. DOGO uses machine-learning to derive a coarse ordering of the data from most to least trusted without making arbitrary good/bad decisions. By seeking to reduce metrics like stdev of southern hemisphere XCO2, difference from TCCON, and stdev of global small areas, DOGO searches the space of all possible filters and isolates a handful of predictive metadata features that, through simple formula, can be used to predict an optimal ordering. This ordering seeks to keep the above metrics reduced while gradually admitting more data for analysis. The ordering is crystallized into the Warn Level (WL) product, so named because higher Warn Levels for an individual sounding indicate higher chances that the sounding will not prove useful. Users use WL's by accepting the soundings with lowest values first, then proceeding higher until the data volume need, coverage need, or problem tolerance is reached. DOGO's WL's are the official quality estimation tool provided with OCO-2 data products starting with version 7.1. This discussion will focus on the precise metadata features and formula derived by DOGO for the v7.1 dataset, demonstrations of its positive impact on science results, and the critical role DOGO's Warn Levels play in determining Bias Correction for OCO-2 XCO2.