BUILDING CHANGE DETECTION BY W-SHAPE RESUNET++ NETWORK WITH TRIPLE ATTENTION MECHANISM

Abstract. Building change detection in high resolution remote sensing images is one of the most important and applied topics urban management planning. Different environmental illumination conditions registration problem are error resource bitemporal that will cause pseudochanges results. On other hand, use deep learning technologies especially convolutional neural networks (CNNs) has been successful considered, but usually causes loss shape detail at edges. Accordingly, we propose a W-shape ResUnet++ network which with different enter independently. residual blocks, triple attention blocks Atrous Spatial Pyramidal Pooling. used on both sides to extract deeper discriminator features. This improves channel spatial inter-dependencies, while same time reducing computational cost. After that, Euclidean distance between features computed deconvolution done. Also, dual function designed weighted binary cross entropy solve unbalance changed unchanged data training second part, mask–boundary consistency constraints condition converging edges predicted edge added. We implemented proposed method two datasets then compared results state-of-the-art methods. The F1 score improved 1.52 % 4.22 by using model first dataset, respectively.