# A deep learning architecture for leaf water potential prediction in Populus euramericana ‘I-214’ from hyperspectral reflectance

## TL;DR

This paper introduces a deep learning model that predicts leaf water potential in a specific poplar species using hyperspectral data, offering a non-destructive and efficient method for monitoring tree drought stress.

## Contribution

The novel CIDL framework addresses data imbalance, integrates global-local features, and models target-variable distribution for improved leaf water potential prediction.

## Key findings

- CIDL achieved R2 = 0.78 and RMSE = 0.27 MPa, outperforming traditional and deep learning methods.
- The model provides a reliable non-destructive method for tree water-stress monitoring.
- CIDL shows strong potential for smart forestry applications with small-sample hyperspectral data.

## Abstract

Leaf water potential (Ψleaf) is a fundamental physiological metric quantifying tree water status and forest drought stress, yet its measurement remains labor-intensive and destructive. Hyperspectral techniques show great promise for retrieving plant physiological traits; however, robust Ψleaf estimation remains limited by three critical factors: unbalanced data distributions, the need for global–local feature synergy, and inherent uncertainty in point-based regression.

Here, we propose a deep learning framework (CIDL) that integrates: (1) a conditional generative adversarial network (CGAN) to generate balanced synthetic samples across the full Ψleaf domain; (2) a feature extractor that combines Inception–ResNet with ACmix (IRAC) to capture local absorption features and long-range spectral dependencies jointly; and (3) a distribution-aware regression network (DARN) to explicitly model the target-variable distribution, thereby enhancing predictive reliability. The model was trained and evaluated using a dataset derived from dehydration experiments on leaves of young Populus euramericana ‘I-214’ trees, comprising 229 paired Ψleaf and hyperspectral reflectance measurements, which were further augmented with 500 CGAN-generated synthetic samples to improve model robustness.

CIDL achieved a prediction accuracy of R2 = 0.78 and RMSE = 0.27 MPa on the test set, clearly outperforming traditional machine learning methods (mean R2 = 0.66, mean RMSE = 0.34 MPa) and yielding a modest yet consistent improvement over mainstream deep learning approaches (mean R2 = 0.76, mean RMSE = 0.28 MPa).

These results demonstrate that the proposed CIDL framework provides a generalizable solution for small-sample physiological hyperspectral analysis and offers a reliable, non-destructive pathway for tree water-stress monitoring, with strong potential for applications in smart forestry management.