From Spectra to Biophysical Insights: End-to-End Learning with a Biased Radiative Transfer Model

TL;DR

This work proposes to integrate RTMs into an auto-encoder architecture, creating an end-to-end learning approach that not only corrects biases in RTMs but also outperforms traditional techniques for variable retrieval like neural network regression.

Abstract

Advances in machine learning have boosted the use of Earth observation data for climate change research. Yet, the interpretability of machine-learned representations remains a challenge, particularly in understanding forests' biophysical reactions to climate change. Traditional methods in remote sensing that invert radiative transfer models (RTMs) to retrieve biophysical variables from spectral data often fail to account for biases inherent in the RTM, especially for complex forests. We propose to integrate RTMs into an auto-encoder architecture, creating an end-to-end learning approach. Our method not only corrects biases in RTMs but also outperforms traditional techniques for variable retrieval like neural network regression. Furthermore, our framework has potential generally for inverting biased physical models. The code is available on https://github.com/yihshe/ai-refined-rtm.git.

Figures (14)

• Figure 1: Learning the inverse.(a) End-to-end learning of the inverse $\mathbf{E_C}$ of RTM $\mathbf{F}$ together with bias correction function $\mathbf{C}_C$. (b) Classical approach --- sampling $D_s$ to train regressive neural network $\mathbf{E_D}$ --- serves as a baseline.
• Figure 2: Distributions of biases learned by AE_RTM_corr. The biases are computed by subtracting the corrected spectra from the originally simulated spectra, given the same set of inferred variables. For bands in the near-infrared and short wave ranges, the RTM tends to under-estimate the spectra for deciduous forest but over-estimate for coniferous forest.
• Figure 3: Superior reconstruction accuracy for AE_RTM_corr, illustrated by spectral band. (a) Reconstruction from NNRegressor displays clear bias (b) AE_RTM_corr reconstructs $X_r$ markedly more accurately.
• Figure 4: Distributions of variables. (a) Application of the NNRegressor to real Sentinel-2 spectra leads to implausible parameter distributions that tend to break out of the preset parameter ranges. (b) Distributions of variables learned by AE_RTM_corr are plausible and distinguish between forest types.
• Figure 5: Pairwise co-distributions of $Z_{r, C}$ learned by AE_RTM_corr. Red: coniferous forest. Blue: deciduous forest. Our model can learn distinct physical patterns.