Flooding is one of the most dangerous natural disasters that can cause irreparable damage to human societies. Accurate flood prediction is essential for reducing damages and better planning. This study describes the potential application of long short-term memory (LSTM) network, a deep learning neural network architecture, for explicit spatial prediction and mapping of flash flood probability. These models were trained based on hydrological and meteorological data and their performance was evaluated using various criteria. The performance of this model seemed to provide stronger results than classical statistical models, but due to the lack and abundance of inconsistent data in central Iran, which is usually due to the spatial dispersion of stations and measurement errors and the lack of metadata, and their access to them is also somewhat difficult, the presented forecast showed that the estimation of this model faced challenges in peak discharges and tended to greatly underestimate peak discharges due to large values ​​of zero. This shows that although neural network models are a powerful tool for hydrological modeling, they are highly dependent on accurate and regular data and also require long-term time series statistics. However, if the same model is combined with spatial data series, including the location of stations and the topography of the river course and even other spatial features that lead to the creation and amplification of floods, it will definitely provide more satisfactory results.