Temperature is a crucial indicator for studying climate, as well as the social and economic changes in a region. Temperature reanalysis products, such as ERA5-Land, have been widely used in studying temperature change. However, global-scale temperature reanalysis products have errors because they overlook the influence of multiple factors on temperature, and this issue is more obvious in smaller areas. During the cold months (January, February, March, November, and December) in the Yellow River Basin, ERA5-Land products exhibit significant errors compared to temperatures observed by meteorological stations, typically underestimating the temperature. This study proposes improving temperature reanalysis products using deep learning and multi-source remote sensing and geographic data fusion. Specifically, convolutional neural networks (CNN) and bidirectional long short-term memory networks (BiLSTM) capture the spatial and temporal relationships between temperature, DEM, land cover, and population density. A deep spatiotemporal model is established to enhance temperature reanalysis products, resulting in higher resolution and more accurate temperature data. A comparison with the measured temperatures at meteorological stations indicates that the accuracy of the improved ERA5-Land product has been significantly enhanced, with the mean absolute error (MAE) reduced by 28.7% and the root mean square error (RMSE) reduced by 25.8%. This method obtained a high-precision daily temperature dataset with a 0.05° resolution for cold months in the Yellow River Basin from 2015 to 2019. Based on this dataset, the annual trend of average temperature changes during the cold months in the Yellow River Basin was analyzed. This study provides a scientific basis for improving ERA5-Land temperature reanalysis products in the Yellow River Basin and offers theoretical support for climate change research in the region.