Abstract: Abstract: Wheat yellow rust is a deadly disease of winter wheat and its timely and accurate detection at regional scale is critical to ameliorate infectious yield loss and safeguard wheat production. With the development in remote sensing technology, satellite imagery with high spatial resolution and high revisiting frequency has become increasingly accessible. Remote sensing data has a unique advantage over traditional method in detecting crop diseases and controlling their spreading, including simple operation, real-time detection, high spatiotemporal resolution and targeting specific-disease, especially the multispectral satellite imagery which covers a wide range of wave bands and provides rich information related to crop diseases at regional scale. Compared to conventional broad band satellite imagery, the Sentinel-2 is a sensor with three wave bands within the edge of the red light which are sensitive to crop diseases. In this study, a Sentinel-2 image acquired in May 12, 2018 was used to extract 26 characteristic variables related to wheat yellow rust, including 3 visible bands (blue, green and red) reflectance variables, one near infrared band, 3 red-edge bands, 14 broad-bands and 5 red-edge vegetation indices. An approach combining K-means and ReliefF algorithm is proposed to filter these features. We first used the RelieF algorithm to calculate the weight of each feature and filter out 10 features most relevant to the disease. The feature with maximum weight was then taken as the initial center of the K-Means algorithm, and other features were added to form a cluster in descending order of their weight. The combination of features with the highest clustering accuracy was taken as the final input variable to the model. The optimal features, including enhanced vegetation index (EVI), structure intensive pigment index (SIPI), simple ratio index (SR), normalized red-edge2 index(NREDI2), normalized red-edge3 index (NREDI3), three wide-band vegetation indices and 2 red edge band vegetation indices were fed into the yellow rust severity monitoring model as input. The back propagation neural network (BPNN) method was a widely used nonlinear artificial neural network and can learn, implicitly, the relationships between inputs and outputs via a multi-layered network. Network training is a process of continual readjustment of weights and thresholds by reducing the network error to a pre-sent minimum or pre-set training steps. We used BPNN to calculate severity of wheat yellow rust (healthy, slight, sever) in Ningqiang County, Shaanxi province, by using the broad-band vegetation indices and the red-edge band vegetation indices as inputs. The results showed that the BPNN model considering broad-band and red-edge vegetation indices as inputs worked better than model using only a single broad-band vegetation indices, improving accuracy by more than 10% and commission accuracy and kappa coefficient reached by 83.3% and 0.73, respectively. The BPNN model includes more information in its input parameters, thereby improving the accuracy of detecting crop diseases. It is more suitable for detecting wheat yellow rust at regional scales and has a wide implication in monitoring and controlling crop diseases at regional scale.