Fast and nondestructive monitoring of leaf water content (LWC) is critical to crop drought diagnosis and irrigation decision. In order to quantify and predict the LWC with hyperspectral remote sensing data, field experiments of winter wheat with different water-deficit stress levels were conducted for two consecutive years (2016—2017 and 2017—2018). Hyperspectral reflectance was recorded at four growth stages. Then, normalized difference spectral index (NDSI) and ratio spectral index (RSI) were calculated in all possible combinations within 350-2500 nm, and their correlations with LWC were quantified to identify the best indices. Spectral data were also used to build partial least squares regression (PLSR) and competitive adaptive reweighted sampling- partial least squares regression (CARS-PLSR) model to calculate LWC. Two different data forms (original and first derivative reflectance) and two different observation scales (leaf and canopy) were used to explore the suitability of these three algorithms on estimating LWC for winter wheat. Additionally, in order to avoid sampling uncertainty when constructing calibration and validation datasets, a method of increasing the number of sampling times was proposed to improve the robustness of prediction models. The results showed that the best spectral indices for estimating LWC of winter wheat were NDSI (R1 162, R1 321) and RSI（R1 162, R1 321） with R2 of 0.871 and 0.872 respectively, which were both based on original leaf reflectance. RSI models had higher estimation accuracy than NDSI models under the same situation. The PLSR model based on original leaf reflectance obtained the best performance for predicting LWC with R2 of 0.953. CARS-PLSR based on the first derivative leaf reflectance was the most accurate model (R2=0.969; RMSE=0.164; RRMSE=6%). It was indicated that in terms of different forms of hyperspectral data, the original spectral-based models were better than the first derivative spectral-based models in two-band vegetation index and PLSR models, but the results were reversed for the CARS-PLSR model. While for different observation scales, LWC had stronger correlations with leaf reflectance-based models than that of the canopy based models. Overall, CARS-PLSR delivered better performance than the other two methods. In light of this, CARS was a feasible band selection algorithm and the prediction accuracy of CARS-PLSR model for LWC estimation of winter wheat was better than that of the other two models. CARS-PLSR method provides a promising approach for accurate and rapid monitoring of winter wheat drought.