Interpretable machine learning and causal inference for maize LAI estimation from UAV multimodal imagery

Abstract

Accurate retrieval of leaf area index (LAI) is vital for crop monitoring and genetic breeding. Although multi-modal unmanned aerial vehicle (UAV) remote sensing has advanced LAI estimation, conventional empirical models often overfit on small breeding populations and cannot disentangle the true causal effects of genetic backgrounds from confounding factors within a statistically rigorous framework. This study presents a robust framework for plot-scale maize LAI estimation across 800 breeding plots from four genetic subgroups: doubled haploid (DH), mixed, temperate (TEM), and tropical/subtropical (TST). From UAV RGB and multispectral imagery acquired at three phenological stages, we extracted 76 multi-modal features comprising point-cloud structural metrics, spectral vegetation indices, and texture features. Following a dual-criterion mutual information and multicollinearity filter, four algorithms, including traditional tree-ensembles and the Tabular Prior-data Fitted Network (TabPFN), were evaluated using nested validation. TabPFN achieved superior generalization performance, yielding a mean test R2 of 0.778 ± 0.031, an RMSE of 0.264 ± 0.024, and an MAE of 0.203 ± 0.022, significantly outperforming tree-ensemble models (p < 0.01). Across growth stages, retrieval accuracy peaked at the expanded bell-mouth stage (R2 = 0.802) and successfully captured the unimodal trajectory of canopy development. SHAP-based attribution showed that spectral indices contributed most to the predictions (55.9%), followed by canopy texture (26.3%), with the spatial heterogeneity metric Tex_Entropy being the most influential single feature (22.8%). When embedded as the nuisance estimator within a Double Machine Learning framework for causal inference, TabPFN confirmed that, relative to the TEM subgroup, only the DH genetic background exerted a consistent and significant negative causal effect on LAI (ATE = −0.070, p = 0.030). These results establish TabPFN as a reliable and extensible tool for non-invasive, high-throughput phenotyping in precision breeding.

Plant phenotyping relevance

UAVマルチモーダル画像からトウモロコシLAIを推定する計算・画像解析ワークフローを開発・比較検証し、高スループット表現型解析への応用を示しているため、方法が中心的である。

Code and data availability

公開本文の所在を確認できませんでした。非公開または購読が必要な可能性があります。

No evidence-backed public reproduction asset is currently recorded.