Abstract: Abstract: Locust hazard is one of the major disasters for farming and animal husbandry in Xinjiang, China. Currently locust disaster monitoring mainly relies on the limited observatory field sites and is not efficient due to Xinjiang's remote geographic location, vast area and inadequate technological support. Fortunately, remote sensing technique offers a valuable tool for locust hazard monitoring and prediction in a large area such as Xinjiang. This study presents a progressive modeling approach for locust hazard risk prediction of the rangeland in Xinjiang. The underlying thought is that the model is to be built based on the key 3 growth stages of locust, namely oviposition, incubation and development, and these processes are heavily affected or even determined by the locust habitats which can be resolved into some key ecological and environmental factors, such as land surface/air temperature, rainfall, soil moisture, soil type, vegetation type and coverage, geographic altitude. The suitability of locust habitat is assessed for these 3 stages using satellite remote sensing data, adopting locust oviposition suitability indicator (OSI), incubation suitability indicator (ISI) and development suitability indicator (DSI). The 3 types of suitability indicators are created mainly based on the derivatives from Terra/MODIS remote sensing data, digital elevation model (DEM) data and ground measured ancillary data. The OSI is created by the weighted combination of 3 sub-indices: soil type factor, soil moisture factor and vegetation factor for oviposition. The ISI is formed from the multiplication of land surface temperature factor and soil moisture factor. And geographic altitude factor, vegetation coverage factor in development stage and vegetation type factor are used to generate the DSI by a weighted combination. Each factor is normalized to the score from 1 to 10, indicating the degree of suitability of this factor. The number 1 represents least suitability and 10 most suitability. Afterwards, the 3 indicators OSI, ISI and DSI are incorporated into locust risk index (LRI) in a multiplicative manner, which is used as a quantitative index to assess the locust hazard risk spatially. The historical data of locust hazard and in-situ measurement data of locust density in 2014 are used to calibrate the model, and consequently the resultant LRI can be further classified into 4 risk levels: low, low-moderate, moderate-high, and high when LRI is less than 100, >100-200, >200-300 and greater than 300, respectively, which can be empirically used to represent the potential severity of locust disaster in the next few months. Considering the variations and interannual fluctuation, a progressive strategy is proposed and incorporated into the modeling process. Two types of modification are applied to the primary model prediction, i.e. oviposition modification and third-instar modification. This strategy allows to making use of the dynamic data acquired by satellite remote sensors, and periodically updates the habitat factors input of the model through quantitative inversion of remotely sensed data. Therefore, the modeled LRI can better reflect the incoming locust hazard possibility and provide more accurate prediction than conventional single input-output model run. The model is subsequently utilized to assess and predict the risk of the locust hazard in the rangeland of Xinjiang in 2010 and 2014. The result indicates that the proposed progressive strategy for the locust hazard risk can reflect the variability of the key habitat factors that affect the locust population development. It also shows that the progressive approach can avoid inaccuracy of one-time prediction, and the modeled results are well correlated to the actual locust hazard severity degree which is classified based on the in-situ measurements of locust density. On the basis of the validation between the modeled risk levels and in-situ measured locust disaster severity classes, the results indicate that approximately 74.4% of the sample sites are completely fallen into the same level, and 94.9% of the samples are different within one level. Thus, the model is useful for early warning of locust hazard and the disaster prevention and relief in Xinjiang area. The model can also be localized and applied in other arid areas of the world.