Evaluating uncertainties in modeled snow reflectance using UAV-based hyperspectral imaging and multispectral remote sensing

Snow albedo is a key control on surface energy balance in snow-covered environments, yet its estimation from multispectral satellite observations remains uncertain due to limited spectral resolution and spatial heterogeneity in snow reflectance. Thus, accurate surface albedo estimates over snow-covered landscapes are critical for the development of reliable satellite-based albedo products. However, validation data in snowy environments remains scarce, especially at high spatial resolution. This is problematic because within a single satellite scene, snow surfaces often exhibit substantial variability that challenges assumptions of spectral homogeneity of snowpack underlying many reflectance-to-albedo parameterizations.

We present a comparative framework that integrates hyperspectral Unmanned Aerial Vehicle (UAV) observations with multispectral satellite data to evaluate the limitations of derived snow albedo within the spectral configurations of Landsat 8, Landsat 9, and Sentinel 2. Our assessment extended across three distinct snowscapes: alpine, prairie, and taiga in Montana (USA), Montana, and Northern Finland; respectively. Our field-based approach employed two commercial hyperspectral sensors (Resonon Pika L and IR-L), to measure surface reflectance across the VIS–NIR–SWIR range (400-1700 nm; Landsat Bands 1-6; Sentinel 2 Bands 1-11) at high spectral (>250 bands) and spatial (0.3 m) resolution.

We isolated snow-only satellite scenes using a Convolutional Neural Network, enabling the identification of heterogeneous snow surfaces within each snowscape. Hyperspectral reflectance measurements were transformed into Landsat- and Sentinel-equivalent band reflectance using weighted sensor response functions, enabling direct band-wise comparison between hyperspectral and multispectral observations.

Our results highlight systematic discrepancies in Landsat reflectance: notably, strong overestimations in Bands 1, 2, and 5, and a consistent underestimation in Band 6 (SWIR1), with surface reflectance biases reaching up to 17%. The CNN-based classification highlighted the high spatial variability in snow reflectance, underscoring the limitations of assuming homogeneous conditions. These findings demonstrate the need to enhance validation strategies for snow-covered regions and provide a scalable protocol that integrates UAV-based acquisitions, high-resolution spectral measurements, and supervised scene analysis. This work contributes to improved characterization of snow albedo uncertainty and supports refinement of satellite-derived snow albedo products for cryospheric applications.