A Condition-Aware Shading Domain-Adaptive Framework for Robust Chlorophyll Inversion Across Shade Managements in Hopea hainanensis .

Abstract

Shade management, which is widely adopted in cultivation and understory regeneration, alters plant light environments, thereby degrading the trait inversion performance and posing a key challenge in plant phenotyping. To address this issue, this study reframed chlorophyll retrieval of Hopea hainanensis under shade management as an illumination-regime-dependent conditional domain shift problem, and developed a condition-aware domain adaptation framework (CAI-DAI) tailored to this setting. The results showed that chlorophyll content increased with shading intensity, accompanied by clear differences in canopy spectral distributions among shading levels, supporting the presence of condition-dependent variation under shade management. Model comparisons showed that CA-IE and CAI-DAI, which integrate conditional encoding and conditional alignment, performed better than the comparative models across fine-tuning ratios from 30% to 70%. Among them, CAI-DAI achieved the best and most stable performance, with test MAE ranging from 4.355 to 4.774 μg·cm−2 and nRMSE ranging from 16.4% to 18.2%, and R2 ranging from 0.456 to 0.585. Further evaluation at individual shading levels (S1–S4) showed that CAI-DAI produced narrower error ranges than CA-IE. It also showed smaller error fluctuations under most fine-tuning ratios. These results demonstrate that the proposed framework effectively improves robustness under heterogeneous shading conditions and limited labeled samples, providing methodological support for chlorophyll monitoring and decision-making related to shade management.

Plant phenotyping relevance

植物のクロロフィル含量を推定する条件認識型ドメイン適応フレームワークを開発・評価しており、表現型取得・推定手法が研究の中心である。

Code and data availability

The paper's hyperspectral/chlorophyll phenotyping data are explicitly not public (available only on request), and the supplement contains only a feature-selection figure, not datasets, images, code, or models. No paper-specific public asset qualifies.

No evidence-backed public reproduction asset is currently recorded.