A Fast Verification Method of Small Building Samples Using Deep Learning

Building extraction with deep learning requires a large amount of samples used as train data. Small differences in building features and big differences between building and background features are important for qualified samples. However, in practice, it is difficult to obtain the required massive samples with good quality which restricts the application of building extraction from remote sensing images based on deep learning seriously. To tackle the problem, this paper proposed a fast verification method of small building samples using deep learning extraction with remote sensing images. The main steps include: firstly, combine the morphological building index (MBI) feature with their original bands to be new input images for deep learning network. Secondly, customize network model parameters based on the U-series semantic segmentation framework, use the mini-batch gradient descent method to achieve training parameters and then obtain a suitable deep learning model. Finally, carry out verification of building samples based on model training and accuracy evaluation. Experiments showed that, compared with the traditional U-Net deep learning building sample verification, the proposed method has the advantages of less sample demand, low cost, high accuracy and stable effect.