Wine grapes are generally harvested in batches, and their quality is affected by harvest time. Conventional methods mainly rely on the test of physical and chemical indicators of samples in laboratories, such as testing phenol and sugar, to determine the maturity of harvest. However, if multiple fields are required to be continuously monitored before harvesting, it will be difficult to ensure the quality of grapes due to large batches, high costs, heavy workload of sampling and analysis and lower timeliness. In this paper, Cabernet Gernischt taken as the study object, a novel method using the near-ground spectral images by Unmanned Aerial Vehicles (UAVs) to determine maturity was proposed. A multispectral camera, ADC Micro, was carried by a four-rotor UAV, DJI Phantom, and the grape images of nine sampling points were taken in-situ with an S-shaped sampling route. Meanwhile, grape samples were collected. Then, a multispectral image processing software, PixelWrench2 x64, was employed for image processing to obtain the values of red (R), green (G) and near-infrared (NIR) index of each image. In addition, grape juice was obtained by pressing and total sugar was selected as the characteristic of maturity determination due to detection duration, cost and representativeness. A handheld sugar meter, PAl-1, was applied to detect the total sugar of the juice. Furthermore, the significance between R, G and NIR components and sampling date were respectively analysed, illustrating that the R component of leaf-dense areas (the local areas) had the most significant relation with and date (with P-value=5.314 44×10^-4 and Adjusted R²=0.815). Therefore, the local R component was selected as the maturity characteristics of modelling. According to the principle that the model set and validation set should be 4:1, the models between total sugar and local R component were respectively developed using linear and logarithmic regression. The results showed that compared with the linear model, there was a very significant logarithmic relation between them (with p-value=5.124 07×10^-10, adjusted R²=0.970 62). The mean of prediction errors of the model was less than or equal to 1.388%, the maximum prediction error of the model was less than or equal to 4.6% and the pre-harvest prediction error was ±0.46%. It was demonstrated that the logarithm model had high accuracy of detection. As a consequence, before harvesting, the multispectral images of Cabernet Gernischt could be gathered in-situ in fields by using UAVs to collect spectral images to obtain the local R-component value. Then, the value could be taken into the logarithmic model to predict the content of total sugar. Based on the standard that total sugar should be 22%±0.46%, Cabernet Gernischt maturity could be determined. Hence, it is convenient and feasible to use spectral images of fields to predict wine qrapes’ quality and harvest time, which provides a novel idea for the application of spectral images in agricultural production.