Abstract: Abstract: The prevalence of aphid in winter wheat field has a significant impact on the production of winter wheat. An effective and timely forewarning of the scope and severity of the disease at a regional scale will not only reduce yield losses but also alert the stakeholders to take effective preventive measures. Forecasting aphid occurrence with Remote Sensing has great advantages over traditional methods and meteorological data, such as lower cost, simpler operation, more real-time and higher resolution. Chinese HJ-1A/1B data which have a high revisit frequency (<4 days )as well as 30 m spatial resolution, in addition to a full band set covering visible, near-infrared, short-wave infrared, mid-infrared and thermal infrared spectral ranges, which thus significantly expand opportunities of remote sensing application in predicting crop diseases and pest. According to the literature research and field investigation found that apart from crops growth status was closely related to wheat aphid occurrence, development, and dispersal, the environmental conditions of field, such as temperature and humidity ,also have a certain impact on its susceptibility to pest. In this study, we retrieve two parts of information from multi-temporal HJ-CCD optical data and HJ-IRS thermal infrared data, including Ratio vegetation index (RVI) and Normalized Difference Vegetation Index (NDVI) representing grow status of wheat, Perpendicular Drought Index (PDI) and Land Surface Temperature (LST) that represent the environmental characteristics. An independent t-test analysis was used to test the difference between pest and healthy samples based on calibration data. Those vegetation and environmental factors that failed to show a statistical significant (p-value <0.001) were eliminated. Five remotely sensed variables (NDVI-T1, RVI-T1, LST-T1, LST-T2, and PDI-T2) were identified as optimal explanatory variables for developing the pest forecasting model. The relevance vector machine (RVM) model was established to predict aphid occurrence of filling stage of wheat in the Beijing suburbs, which is a machine learning algorithm and commonly used to improve business decisions, detect disease, and forecast weather credit to its superior learning ability when on a small datasets. The results obtained from the RVM model are compared with prediction models developed using logistic regression (LR) and support vector machine (SVM) techniques. Goodness-of-fit values of the calibrated models were evaluated through the p-value of Spearman test. In addition, the association between observed outcome of pest occurrence and predicted outcome of pest occurrence was evaluated by using three different measures: Somers'D, Goodman-Kruskal gamma, and Kendall's Tau-c. The RVM model produced a higher Spearman p-value than that the LR model and the SVM model did, indicating that the RVM model had a better performance to predict the aphid than the other two models. Moreover, the values of Somers'D, Goodman-Kruskal Gamma, and Kendall's Tau-c were all higher with the RVM model than those with the remaining model, which further demonstrated the better performance of the RVM model on the training datasets. To further evaluate the difference of performances of the three types of models, we obtain accuracy of models respectively based on validation samples. The results showed that: accuracy by using RVM algorithm is the highest among the three methods, with lower omission and commission than the other two algorithms. Furthermore, the overall prediction accuracy and the kappa coefficient of the RVM model are 87.5%, 0.71respectively, have shown better performance over the SVM model (79.2%,0.55) and the LR model (75.0%,0.44). Additionally, the RVM model needs less computation work and faster prediction speed. These results revealed that the RVM model with high accuracy level and prediction efficiency is much more favorable in forecast the pest occurrence at a regional scale effective and timely. It indicated that the result of this research can be used in different regions.