Deep learning model improvement for building change detection by including local remote sensing training datasets

Document Type : Original Article


1 School of Surveying and Geospatial Engineering, College of Engineering, University of Tehran

2 Faculty of New Sciences and Technologies, Graduate University of Advanced Technology, Mahan, Kerman, Iran

3 End of Haft Bagh Highway Graduate University of Advanced Technology Faculty of Civil & Surveying Engineering



Deep learning networks which have been trained using known datasets normally do not produce convincing results when used on other datasets. In this study, it has been shown that the training of previously developed networks with the newly included local training datasets greatly increases the accuracy. Accordingly, the aim of this study is to improve a DL model by including local data along with existing known dataset for building change detection. The STANet has not been previously used in industrial areas where there are different building characteristics. High spatial resolution satellite images have been used for buildings change detection in this study. The STANet has been implemented using a new dataset comprising from an existing known dataset together with two local building datasets. The training and testing of the STANet is investigated in incremental training data in three stages. In the first stage, STANet network that was trained by already existing LEVIR-CD dataset was tested, but it did not produce reasonable result. In the second stage, the network which was trained by combining LEVIR-CD and a local dataset was tested and obtained a better result. In the third stage, the network which was trained by combining LEVIR-CD dataset, local dataset from stage2 and another more specific and complementary local dataset, was tested and reached a much better result than the previous two stages. As a result, for the trained network with the corresponding data, high average, recall, accuracy and precision achieved. In conclusion, the significance of using local corresponding training dataset on a DL model has been exhibited.


Main Subjects