Mapping urban scene elements optical properties using a gradient-driven 3D radiative transfer model

Unmixing land cover optical properties (OP) from coarse spatial resolution remote sensing images is a critical methodological challenge for microclimate and energy balance research. We propose the novel Unmixing Spectral method (US-DART) using 3D differentiable radiative transfer model DART to extract endmember OPs from shortwave (SW) mono-/multi-spectral RS images. It has 4 modules: (1) pure pixel selection, (2) linear spectral mixture analysis, (3) gradient iterative refinement, and (4) spectral correlation. It has 3 inputs: a reflectance image, a spatially-classified 3D scene model, and relevant DART parameters (e.g., spatial resolution, skylight ratio, ...). It generates an OP map per type of scene element. It is validated for 3D vegetation and urban scenes quantitatively with DART-simulated satellite image (median pixel reflectance relative error: about 0.1%) and qualitatively with Sentinel-2, PlanetScope, and Google Map Images. Performance is better for opaque surfaces (about 1%) such as roofs than for translucent materials (≈1-5%), such as leaves. By better extracting OPs from coarse resolution SW imagery US-DART substantially advances remote sensing applications, such as simulations of time series of SW albedo and radiation balance.