Estimating cotton leaf area index under shallow subsurface drip irrigation using TabPFN and UAV-based multisource remote sensing

Abstract Leaf area index (LAI) is a key indicator of cotton canopy development and crop water status, yet its accurate estimation under shallow subsurface drip irrigation in arid regions remains challenging because spatial variability in soil moisture and salinity increases canopy heterogeneity. This study integrated UAV-based hyperspectral and RGB imagery with Tabular Prior-data Fitted Network (TabPFN) to estimate cotton LAI under small-sample conditions. Spectral vegetation indices, colour indices, and texture features were extracted, and stable predictors were selected using recursive feature elimination with cross-validation. Stage-specific models for the squaring, flowering, boll-development, and boll-opening stages were developed from hyperspectral, RGB, and fused feature sets. TabPFN was compared with XGBoost, random forest, and support vector regression, and SHapley Additive exPlanations (SHAP) quantified feature contributions. Across the 12 growth-stage–feature-set combinations, TabPFN performed best overall (mean R ² = 0.781 ± 0.190). Multisource fusion was strongly stage dependent: relative to the best-performing single-source feature set, fusion increased R ² by 1.2%, 7.1%, and 12.2% at squaring, flowering, and boll development, respectively, but performed slightly worse than hyperspectral features at boll opening. SHAP revealed clear stage-specific shifts in feature importance: texture dominated at squaring, photochemical indices were most influential at flowering and boll opening, and pigment-related and structural features jointly contributed at boll development. Field-scale LAI maps generated with the optimal models revealed pronounced spatial differences among irrigation treatments. Combining TabPFN with UAV-based multisource remote sensing can provide accurate and interpretable cotton LAI estimates under small-sample conditions, thereby supporting crop monitoring and precision irrigation in arid regions.