For the early hyperspectral images of cucumber downy mildew in greenhouses collected in field, it is influenced by environmental illumination and difficult to extract effective features from them. To solve these problems, a novel method of extracting feature bands based on disease difference information was proposed, which improved competitive adaptive reweighted sampling (CARS) and successive projections algorithm (SPA). Besides, an early detection model was built for cucumber downy mildew. Firstly, hyperspectral images were collected for leaves of healthy cucumber and leaves after infection in 12 consecutive days, which were divided into seven categories based on the degree of infection. Then, spectral data was calculated as the average spectrum of region of interest, the difference bands of downy mildew disease were determined by envelope elimination method and feature bands were extracted via CARS for seven different stages of it. SPA was used to perform secondary dimensionality reduction and optimization. Finally, all feature bands were combined to obtain 47 feature bands data. Based on this, a least square support vector machine (LSSVM) was established for disease detection. The disease detection test was performed on a test set of 94 leaf samples. The results showed that Dis-CARS-SPA-LSSVM fused disease difference information can obtain 100% detection rate after 2~12 days infection of disease. The detection rate of the test set infected with disease for 1 day reached 95.83%, the recall rate of infected samples reached 100%, and it avoided the randomness of CARS-SPA feature extraction method which did not fuse the disease difference information due to the interference bands of the nondowny mildew disease feature bands, and the recognition rate was 4.16 percentage points higher than that of CARS-SPA feature extraction model. The experiment results demonstrated that the proposed Dis-CARS-SPA-LSSVM model can effectively achieve early detection of downy mildew disease in greenhouse with a higher accuracy rate.