Water is a crucial element for the growth of watermelon plants, making rapid and non-destructive monitoring of plant water content vital for precision irrigation in watermelon farming. While previous research has demonstrated the sensitivity of short-wave infrared (SWIR) bands to plant water content, their high costs limit widespread application. In contrast, visible and near-infrared (VNIR) spectral instruments offer significant advantages in terms of affordability, compactness, and spectral resolution. However, their potential for predicting the leaf water content (LWC) of watermelon plants has yet to be fully investigated. This study aims to assess the efficacy of hyperspectral reflectance measured with VNIR spectral instruments in estimating the LWC of watermelon plants at various leaf layers. Hyperspectral reflectance data (350−1100 nm) were collected from three leaf layers (upper, middle, and lower) under various drought treatments. Models for estimating LWC were developed using both spectral indices and full wavelength data. The results indicated that the middle leaf layer was the most effective for estimating LWC, and using full wavelength data achieved higher accuracy in LWC estimation. Furthermore, compared to the simple regression model, the AdaBoost-based machine learning model demonstrated superior performance, achieving an R² of 0.9636 in estimating LWC through five-fold cross-validation, which indicates high predictive accuracy. Ensemble learning significantly outperforms traditional methods, providing a substantial improvement in model accuracy. The findings offer important technical assistance for the spectral monitoring of LWC and precision irrigation in watermelon cultivation.