Objective This study aims to achieve rapid and non-destructive detection of water loss rate in green grapes, control losses, and ensure quality. Based on an LED-induced fluorescence spectroscopy system, a prediction method combining collected spectral data with a multi-model soft voting fusion approach was proposed for green grape water loss rate estimation.Methods Spectral data from 100 green grape samples were collected. Preprocessing was performed using multivariate scatter correction (MSC), standard normal variate transformation (SNV), normalization, and Savitzky-Golay (SG) filtering. Feature wavelength selection was carried out using the successive projections algorithm. Based on this, prediction models for green grape water loss rate were constructed using Random Forest (RF), Partial Least Squares (PLS), and Least Squares Support Vector Machine (LSSVM), as well as a single-model and a multi-model soft voting fusion model, followed by comparative analysis.Results Among the single models, LSSVM showed the best overall performance but exhibited a clear risk of overfitting. A multi-model soft voting fusion model was constructed by assigning weights based on the test set Rp2 and mean absolute error (MAE) index. Compared with the LSSVM single model, the test set Rp2 performance improved by 25.3%, and MAE decreased by 27.6%, with overfitting significantly alleviated. Meanwhile, improvements were achieved in both fitting accuracy and generalization ability.Conclusion The multi-model soft voting fusion strategy exhibits higher prediction accuracy than single models, with stronger stability, generalization ability, and anti-overfitting performance.