In forest fire rescue， predicting the final burning area based on the early stages of the fire can effectively guide fire rescue. However， previous studies have used Normalized Difference Vegetation Index （NDVI） as an input indicator， which is sensitive to soil reflectance and has high data noise. Therefore， the Two-band Enhanced Vegetation Index （EVI2） is used to accurately predict the area burned by wildfires. In addition， to address the issue of poor anti-interference ability of a single machine learning prediction algorithm， a Stacking-XRSK model based on stacking ensemble learning is proposed. The results showed that using EVI2 increased $R^{\mathrm{2}}$ by 6.05% compared to NDVI， while reducing MAE and MSE by 0.88% and 0.41%， respectively. Compared with the single model， the Stacking-XRSK model has the highest $R^{\mathrm{2}}$， ranging from 2.8% to 11.06%， and MAE， MSE， and AOC are the lowest. The feasibility and accuracy of using EVI2 instead of NDVI to predict the area of burnt areas have been verified. At the same time， the Stacking model can improve its generalization ability while fully leveraging the advantages of a single base model. This study provides scientific reference for forest fire safety management and timely firefighting.