The invention discloses a remote sensing satellite image multi-temporal change detection method based on a three-dimensional convolutional neural network, and provides a three-dimensional U-net model,the model is input into four dimensions of length, width, channel number and time of the image, the length, width and time are operated at the same time by utilizing three-dimensional convolution, and three-dimensional pooling and upsampling operations are used in the same way. According to the method and the device, the relevance among the pictures is controlled by setting the reasonable convolution kernel size of the time dimension, and the relevance of more pictures can be considered by increasing the dimension. For the previous heavy data label problem, the model can set the weight of theloss function of the unsupervised data to be zero in the training process according to a small amount of supervised data, so that the model is directly trained, and the workload of required labels isgreatly reduced.