Wildfire forecasting problems usually rely on complex grid-based mathematical models, mostly involving Computational fluid dynamics(CFD) and Celluar Automata, but these methods have always been computationally expensive and difficult to deliver a fast decision pattern. In this paper, we provide machine learning based approaches that solve the problem of high computational effort and time consumption. This paper predicts the burning duration of a known wildfire by RF(random forest), KNN, and XGBoost regression models and also image-based, like CNN and Encoder. Model inputs are based on the map of landscape features provided by satellites and the corresponding historical fire data in this area. This model is trained by happened fire data and landform feature maps and tested with the most recent real value in the same area. By processing the input differently to obtain the optimal outcome, the system is able to make fast and relatively accurate future predictions based on landscape images of known fires.