Optimizing neglected and underutilized taro crop water status estimations using UAV‑acquired thermal and multispectral remotely sensed data

Accurate estimation of crop water status is crucial for enhancing agricultural productivity and sustainability, particularly in rainfed smallholder farming systems. Despite the recognized importance of crop water monitoring, limited research has explored growth‑stage-specific estimation of water status in neglected and underutilized crops such as taro (Colocasia esculenta (L)), particularly using integrated thermal, multispectral, and topographic information. This study presents a novel optimization framework combining UAV-derived spectral and thermal information with terrain characteristics and advanced feature selection techniques to estimate taro crop water status across distinct growth stages. Specifically, stomatal conductance and leaf temperature were evaluated as proxies for crop water status during the emergence, vegetative, and maturity stages. Deep Neural Network models were developed using individual feature subsets and subsequently optimized using hyperparameter tuning and Least Absolute Shrinkage and Selection Operator-based feature selection. The findings revealed that multispectral bands and thermal data, together with derived thermal indices, consistently outperformed individual feature subsets, producing R2 values ranging from 0.61 to 0.84. Combining all feature groups further improved predictive performance, yielding R2 values exceeding 0.86 across all growth stages. Following optimization, models achieved R2 values ranging from 0.91 to 0.95 and rRMSE values below 17%, for both stomatal conductance and leaf temperature. The vegetative growth stage exhibited the highest predictive performance, highlighting the importance of growth‑stage-specific modelling for crop water status estimation. This study provides the first growth‑stage-specific optimization framework for estimating smallholder taro water status through complementary UAV and terrain-derived predictors. The findings demonstrate the value of combining high-resolution proximal sensing and advanced machine learning optimization to support precision water management and climate-resilient cultivation of neglected and underutilized crops in resource-constrained farming systems.