Abstract
<jats:p>Flash-drought monitoring by classical indices averages predictive signal across calm and stressed years, blunting the signal exactly where it matters. We show empirically that the coupling between precipitation, soil moisture, vapor pressure deficit (VPD), and NDVI is stress-conditional: the hydrological pathway is visible when vegetation is already under stress (Spearman rho of about 0.18 for 30-day precipitation) and near-noise otherwise (about 0.04). Building on that observation, we introduce WxCond, a compact (about 3k-parameter) flash-drought detector that couples MODIS NDVI with ERA5-Land anomalies through a learned scalar stress gate; the gate produces a per-row, inspectable routing between a hydrological and a vegetation feature-family expert. Across six US Corn Belt states (587 counties, 2000-2023) and the weekly US Drought Monitor D2+ label, WxCond improves Average Precision by roughly 12 percentage points over the strongest classical baseline under leave-one-state-out cross-validation; on the 2012 flash drought it fires a median of 16 days before USDM D2+ onset across 511 counties (a shorter lead than SMVI's 32 days at matched precision, so a recall-for-lead-time operating trade-off rather than a strict win). An ablation shows the gated architecture is within LOYO fold-variance of an ungated variant and a plain single-MLP baseline: the gate contributes interpretability and per-row pathway decomposability rather than raw accuracy, which is the framing we adopt. SHAP attribution is consistent with the precipitation to soil-moisture to VPD to NDVI pathway. The result is a small, transparent, spatially generalising template for stress-conditional drought monitoring in tabular agricultural machine-learning.</jats:p>