Abstract: Due to the low time series of medium and high-resolution remote sensing image data, the time series characteristics of ground objects cannot be accurately recorded during the classification process. In order to increase the temporal change characteristics of ground features, the study uses the spatio-temporal fusion model to reconstruct high-sequence and high-resolution remote sensing images, and analyzes the influence of adding temporal features on the classification results. Taking the central area of Shijiazhuang City, Hebei Province as an example, the study uses three spatiotemporal fusion models to reconstruct high-time sequence 30-meter resolution remote sensing images to add time series classification features. This study uses random forests to classify annual reconstructed time series images, and analyzed the impact of different number of reconstructed images and different time spans on the classification results. Experiments show that the accuracy of image classification by reconstructing annual time series is better than that of single image. The classification accuracy increases with the number of time series images, when the number of images is selected as 12 scenes (one scene per month), the classification accuracy tends to be stable. And different time periods have different influence on classification results. With the introduction of imagery during vegetation change, the classification accuracy is the highest.

Key words: spatio-temporal fusion, data reconstruction, time series classification, random forest, accuracy evaluation