TorchRTM: A high-performance vegetation RTM package for deep learning integration

Abstract

The surge in earth observation data availability has exponentially enhanced the potential for monitoring, analysis, and modelling of biodiversity, habitat degradation, and climate impacts. However, to effectively drive conservation efforts, this data deluge requires computationally-efficient advanced interpretation methods. Deep learning is a popular tool, that excels in predictive tasks, yet its interpretability and usefulness for mechanistic ecological and environmental modelling could be further improved by integration with mechanistic radiative-transfer models (RTMs). However, processing such hybrid models is computationally intensive and requires robust software frameworks for efficient implementation of both deep learning and RTMs. Here, we present TorchRTM, a GPU-supported multi-level vegetation RTM with highly parallelized computation supporting inverse and forward modes. TorchRTM is fully compatible with the deep learning frameworks PyTorch, as it leverages both data-level and instruction-level parallelism to enhance computational efficiency. All biophysical modules in TorchRTM support GPU acceleration, it can be scaled to integrate with additional ecological system models, combined with ecological remote sensing observations. Its modular architecture allows seamless integration into deep learning models, facilitating accurate radiative transfer parameter estimation and improving the understanding of underlying physical processes. TorchRTM is demonstrated its high-speed simulation capabilities through two model inversion examples including, 1) a classical look-up table, and 2) an inverse deep learning framework supporting hybrid prediction. Our software offers a promising solution for remote sensing data retrieval via hybrid modelling, combining the interpretability of mechanistic models with the computational power of deep learning. We believe this makes it a powerful tool for the scientific community.

Authors 3

1. Peng Sun Aachen

   Leiden University

   Affiliation as printed

   Leiden University

2. Peter M. van Bodegom Aachen

   Leiden University

   Affiliation as printed

   Leiden University

3. Marco D. Visser Aachen

   Leiden University

   Affiliation as printed

   Leiden University