Active learning regularization increases clear sky retrieval rates for vegetation biophysical variables using Sentinel-2 data

Abstract For typical cloud conditions, a clear sky retrieval rate (CSRR) >67% is required to meet the Global Climate Observing System temporal interval requirement of 10 days when mapping canopy biophysical variables (‘variables’). Physically based algorithms suitable for global mapping of variables using multispectral satellite imagery, e.g. the Simplified Level 2 Prototype Processor (SL2P), typically have a CSRR between 25% and 75%. An Active Learning Regularization (ALR) approach was developed to increase the CSRR rate while satisfying uncertainty requirements. A local calibration database for each variable was produced from representative valid SL2P estimates and associated Sentinel-2 Multispectral Instrument surface reflectance estimates. Predictors for each variable were developed by i) using Least Absolute Shrinkage and Selection Operator regression to select a subset of spectral vegetation indices (VIs) from a provided library, ii) removing outliers from the calibration database by trimming the conditional distribution of each variable given a VI, and iii) calibrating a non-linear regression predictor of the variable given the selected VIs using the trimmed database. ALR was applied to MSI imagery acquired over the Canadian Prairies during the 2016 and 2018 growing seasons and validated with in-situ data collected over 50 fields by the SMAPVEX16-MB campaign. The mean CSRR during the 2018 growing season was ~98% (~70%) for ALR (SL2P) for all canopy variables except FCOVER and ~ 98% for FCOVER using both ALR and SL2P. In comparison to SL2P, ALR had increased agreement rates with in-situ leaf area index (86% versus 79%) and fraction cover (96% versus 79%) but not canopy water content (35% versus 53%). Intercomparison with valid SL2P estimates from different MSI images acquired within ±2 days found that 90% [±5%] of ALR estimates fell within the uncertainty of the valid estimates. These findings support the hypothesis that, over croplands, ALR significantly increases CSRR over SL2P without appreciably increasing uncertainty for variables retrieved by SL2P within thematic performance requirements.