Droughts in Africa's drylands threaten regional food security and global agricultural markets. Early-warning systems increasingly rely on Earth Observation (EO), yet precipitation-based indicators often fail to detect emerging vegetation water stress. With new low-Earth-orbit missions, evapotranspiration (ET), which represents actual land-surface water flux, and ET-derived metrics such as the Evaporative Stress Index (ESI) have become essential. ECOSTRESS provides ∼70 m sub-daily land surface temperature observations for ET estimation via the Priestley–Taylor Jet Propulsion Laboratory (PT-JPL) model. However, PT-JPL often exhibits positive ET bias in drylands, increasing the risk of drought omission errors. We evaluated whether hyperspectral vegetation indices (HVIs) can reduce these biases using multi-year field spectrometry, eddy covariance fluxes, and EnMAP/PRISMA imagery in a Kenyan dryland experiment. On the independent validation subset, the standard PT-JPL overestimated ET by 21.8% (mean observed latent heat = 4.86 MJ m−2 d−1). Incorporating HVIs reduced bias to 3.5% when constraining soil evaporation and to −6.4% when applied to both canopy and soil components, while also improving other goodness-of-fit metrics. Bias reduction occurred through two mechanisms: (i) alleviating NDVI saturation, which strengthened canopy constraints under wetter conditions, and (ii) reformulating the soil-moisture constraint using hyperspectral reflectance, thereby limiting soil-evaporation inflation under humid and transitional conditions. These improvements were consistent across hydrological periods and sensor platforms. The findings demonstrate that narrowband spectral information enhances ET partitioning and directly support upcoming narrowband–thermal missions (e.g., CHIME, Landsat Next, LSTM, S2NG, SBG) by improving ET-based drought early-warning in moisture-limited environments.