Abstract

Recent advancements in machine learning, combined with the availability of vast remote sensing data, have led to significant progress in ecology and climate science. However, the lack of interpretability in learned representations limits their application to crucial environmental challenges. Understanding the biophysical properties of forests, for instance, is essential in comprehending their role in mitigating climate change. In the field of remote sensing, scientists have attempted to retrieve the biophysical variables by inverting the radiative transfer models (RTMs). However, classical approaches overlook the presence of systematic bias in RTMs, which is particularly problematic when extracting variables from complex forest structures. Motivated by physics-informed machine learning and disentangled representation learning, we propose an innovative approach that integrates the RTM with an auto-encoder-based architecture. Our approach integrates an RTM into a contemporary machine-learning framework and effectively corrects its bias, resulting in improved variable extraction. The developed pipeline holds broad applicability in other machine-learning problems involving physical models. Our research advances the integration of RTMs and machine learning, enabling more accurate analysis of remote sensing data and facilitating a better understanding of forest biophysical properties.

Yihang She is a first-year PhD student in Computer science at the University of Cambridge. His PhD focuses on the development of 3D vision algorithms to enable real-time and low-cost forest carbon estimation.