A physically based differentiable radiative transfer model (DRTM) for land surface optical and biochemical parameters retrieval

The differential path tracing method and automatic differentiation can effectively calculate the derivatives of the loss function, enabling the estimation of surface properties such as reflectivity and transmissivity from sensor images. However, their full potential has not been completely explored in remote sensing. We developed a differentiable radiative transfer model (DRTM) to efficiently simulate and retrieve leaf optical properties, leaf biochemical components, and sensor observation angles from passive remote sensing imagery. The modeling accuracy is verified using various three-dimensional (3D) heterogeneous landscapes, including natural vegetation-covered and artificial urban landscapes. The forward modeling part of DRTM has proved to be faster and more efficient in computer resource usage. In addition, DRTM demonstrated a much more effective adaptation of deep learning than the traditional look-up table method, to better resolve the most challenging inversions from canopy level to foliar level in vegetation remote sensing. In this context, DRTM can potentially address various inverse challenges in remote sensing within a unified framework.