Abstract: Cucumber downy mildew can spread rapidly, even causing significant yield reduction in greenhouses. Existing disease detection mostly starts after the symptoms of the disease appear, resulting in delayed early warning. It is often required to determine the optimal period of prevention window. A disease prediction model can be constructed to consider the growth status of the crops, environmental, and pathogen information. Early warning of downy mildew can be expected to promote the green production of greenhouse vegetables. This study aims to overcome the delayed warning and passive prevention of cucumber downy mildew in greenhouses. An LSTM-BPNN hybrid neural network was proposed to timely predict the severity of cucumber downy mildew in greenhouses. Multi-source information was also integrated with crop growth, environment, and airborne spores. A series of experiments was conducted in a Venlo-type greenhouse at Jiangsu University in China. The test material was taken as the cold-resistant and weak-light-tolerant variety of "Jinyou 1" cucumber. Different concentrations of nutrient solution were set for treatment. A continuous 66-day observation was conducted during measurement. A total of 198 valid samples were obtained after the experiment. Phenotypic parameters of the cucumber plant were collected, such as plant height, stem diameter, and leaf area. Environmental data were also captured in the greenhouse, such as average temperature and relative humidity. At the same time, spore-capturing devices were used to capture spores of the downy mildew pathogen in the air. The spore concentration was obtained after microscopic counting. The leaf images were processed using the PP-LiteSeg semantic segmentation network. The total leaf area and lesion area were calculated without loss, thereby calculating the severity of cucumber downy mildew. All samples were preprocessed using maximum value normalization. The correlation between each factor and the severity of disease was analyzed using the Pearson correlation coefficient. The results showed that cucumber plant height, stem diameter, and leaf area were negatively correlated with the severity of the disease, with correlation coefficients of -0.37, -0.37, and -0.52, respectively. The better the plant growth condition was, the stronger the disease resistance was, and the lower the disease severity was. The average temperature, average relative humidity, and the number of spores in the air were positively correlated with the severity of the disease, with correlation coefficients of 0.28, 0.41, and 0.38, respectively. Especially in the early stage of the disease (severity <20%), the correlation coefficient between spore number and disease severity reached 0.63, which was the key indicator for early warning. The plant height, stem diameter, and leaf area were normalized to serve as dimensionless growth characteristics Q, thus reducing the feature dimension. The memory mining of LSTM for time series data was combined with the excellent nonlinear mapping of BPNN. An LSTM-BPNN fusion model was constructed to predict the gradient disappearance of the sigmoid function for small disease severity samples. The activation function of the output layer in the BPNN was replaced with ELU to optimize the network hyperparameters. The number of neurons was set to 5 and 8, respectively, in the LSTM and BPNN hidden layers. Adam optimizer was used for model training. The dataset was divided into training set, validation set, and test set in an 8:1:1 ratio. An early stopping strategy was introduced to prevent overfitting. The test results showed that the average absolute error (MAE) of the LSTM-BPNN model was 0.0133, the mean square error (MSE) was 0.0068, the root mean square error (RMSE) was 0.0825, and the determination coefficient (R²) reached 0.942. All evaluation indicators were superior to those of the single LSTM and BPNN models. Compared with the LSTM-CNN and GRU-BPNN hybrid models, 99.40% and 95.44% accuracy rates were achieved, respectively, for 3- and 7-day disease prediction. The prediction performance was better than that of the comparison models. The prediction error of the model accumulated with increasing prediction time suitable for short-term rolling prediction within 7 days. The accuracy of prediction was obtained by iteratively outputting disease risks using daily updated growth, environmental, and spore data. Three types of information - host growth resistance, environmental stress, and pathogen spore sources- were integrated into the prediction framework. Disease warning was based on plant growth indicators or pathogen spore information, rather than before. A precise early warning technical solution was provided for cucumber downy mildew in greenhouses. Disease prevention and control were promoted to reduce pesticide abuse and yield losses caused by the disease. The findings can also offer theoretical references and technical support for the intelligent early warning and prevention of airborne diseases in facility agriculture.