Improving Surface Property Retrievals in Boreal Seasonal Snowpacks through Multiscale Modeling of Subgrid Reflectance

Time-series of surface reflectance and derived land cover and snowpack properties, including snow grain size (SGS), Normalized Difference Snow Index (NDSI), and Soil-Adjusted Vegetation Index were analyzed across multiple spatial resolutions using Landsat (30 m), MODIS (500 m), and VIIRS (1 km) observations over a 10-year period in the Western Canadian domain of NASA's Arctic and Boreal Vulnerability Experiment (ABoVE). The objective is to demonstrate a scalable framework for decomposing VIIRS reflectance signals and associated surface property estimates by accounting for the subgrid-scale heterogeneity introduced by mixed landcover types such as forests, wetlands, and open snow fields. The approach is to separate snow and vegetation reflectance contributions within coarse-resolution pixels and leverage this information toward improving retrieval of SGS, NDSI, and SAVI. The analysis showed that differences between VIIRS 1 km coarse scale reflectance and Landsat averaged reflectance at the same resolution are closely linked to subgrid-scale variability in forest fraction, spatial organization of forests (quantified by pixel-based chi square heterogeneity), and the spectral behavior of the subgrid 500 m MODIS and VIIRS bands. To address these discrepancies across scales, a random forest model was trained against high-resolution Landsat data to predict separately the mean and standard deviation of subgrid reflectance within each VIIRS pixel for forested and non-forested areas using VIIRS observations and ancillary data (topography and land-cover). Model results show strong agreement with Landsat for mean NIR reflectance (RMSE < 0.07, R² > 0.90) and good skill in capturing spatial variability (R² ∼ 0.5-0.7), particularly in heterogeneous forested regions. Predicted reflectance distributions were used to estimate NDSI, SAVI, and SGS, demonstrating improved agreement with Landsat-derived values and reducing known biases in VIIRS-based retrievals.