# Integrating PROSPECT-D physics and adversarial domain adaptation resnet for robust cross-ecosystem plant traits estimation

## TL;DR

This paper introduces PPADA-Net, a new method combining physics-based modeling and machine learning to accurately estimate plant traits across different ecosystems using remote sensing data.

## Contribution

The novel PPADA-Net framework integrates PROSPECT-D radiative transfer modeling with adversarial domain adaptation for robust cross-ecosystem plant trait prediction.

## Key findings

- PPADA-Net outperforms traditional and data-driven models with R² values of 0.72 (CHL), 0.77 (EWT), and 0.86 (LMA).
- The model achieves high-precision spatial mapping in a real-world farmland dataset with nRMSE of 0.07 for LMA.
- The framework enhances spectral-trait modeling under data scarcity, enabling scalable ecosystem monitoring.

## Abstract

Plant functional traits, including chlorophyll content (CHL), equivalent water thickness (EWT), and leaf mass per area (LMA), are critical indicators for assessing ecosystem functioning, functional diversity, and their roles in the Earth system. Hyperspectral remote sensing serves as a pivotal tool for multi-trait mapping; however, existing methods exhibit limited generalizability across ecosystems, land cover types, and sensor modalities. Challenges such as data heterogeneity, domain shifts, and sparse in situ measurements further hinder model generalization. To address these limitations, this study developed PPADA-Net, a novel framework integrating PROSPECT-D radiative transfer modeling with adversarial domain adaptation for robust cross-ecosystem plant trait prediction. In a two-stage process, a residual network is pretrained on synthetic spectra from PROSPECT-D to capture biophysical links between leaf traits and spectral signatures, followed by adversarial learning to align source and target domain features, reducing domain shifts. The model’s performance is validated on four public datasets and one field-measured dataset. PPADA-Net outperforms traditional partial least squares regression (PLSR) and purely data-driven models (e.g., ResNet), achieving mean R² values of 0.72 (CHL),0.77 (EWT), and 0.86 (LMA). Additionally, PPADA-Net demonstrates practical utility in a real-world farmland dataset (D5), achieving high-precision spatial mapping with an nRMSE of 0.07 for LMA. By merging physical priors with adaptive learning, PPADA-Net enhances spectral-trait modeling under data scarcity, offering a scalable tool for ecosystem monitoring, precision agriculture, and climate adaptation.