SIAMIXFORMER: A FULLY-TRANSFORMER SIAMESE NETWORK WITH TEMPORAL FUSION FOR ACCURATE BUILDING DETECTION AND CHANGE DETECTION IN BI-TEMPORAL REMOTE SENSING IMAGES

Sep 21, 2026

Uploaded by Jiven

Building detection and change detection using remote sensing images supports urban planning and rescue decision-making, and also enables building damage assessment after natural disasters. Existing approaches often rely on a single pre-disaster image because post-disaster damage can degrade performance. This work introduces SiamixFormer, a Siamese architecture with two hierarchical transformer encoders and temporal fusion to combine pre- and post-disaster features for improved accuracy.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages16

Reading Time40 min

AccessFree to download

Abstract

1 Introduction

Disaster management and the need for building change assessment

Traditional feature-based methods and early machine learning

CNN-based and fusion approaches for building change detection

Bi-temporal comparison strategies: segmentation, feature, and input levels

Transformer architectures for remote sensing and building-focused variants

arXiv :2208 .00657v2 [ cs .CV] 21 Jul 2023 SIAMIXFORMER: A FULLY-TRANSFORMER SIAMESE NETWORK WITH TEMPORAL FUSION FOR ACCURATE BUILDING DETECTION AND CHANGE DETECTION IN BI-TEMPORAL REMOTE SENSING IMAGES Amir mohammadian, Foad Ghaderi Human-Computer interaction lab, Electrical and Computer Engineering Department Tarbiat Modares University Tehran, Iran July 24, 2023 ABSTRACT Building detection and change detection using remote sensing images can help urban and rescue planning. Moreover, they can be used for building damage assessment after natural disasters. Currently, most of the existing models for building detection use only one image (pre-disaster image) to detect buildings. This is based on the idea that post-disaster images reduce the model’s performance because of presence of destroyed buildings. In this paper, we propose a siamese model, called SiamixFormer, which uses pre-and post-disaster images as input. Our model has two encoders and has a hierarchical transformer architecture. The output of each stage in both encoders is given toa temporal transformer for feature fusion in a way that query is generated from pre-disaster images and (key, value) is generated from post-disaster images. To this end, temporal features are also considered in feature fusion. Another advantage of using temporal transformers in feature fusion is that they can better maintain large receptive fields generated by transformer encoders compared with CNNs. Finally, the output of the temporal transformer is given to a simple MLP decoder at each stage. The SiamixFormer model is evaluated on xBD, and WHU datasets, for building detection and on LEVIR-CD and CDD datasets for change detection and could outperform the state-of-the-art. 1 Introduction Natural disasters such as earthquakes, floods, and tsunamis have always been a human concern that take lives of many people every year and impose many financial losses on different countries. Highly affected by the changes in global climate, natural disasters have become stronger and more frequent in recent years [1].One of the most important tasks in response phase of disaster management is assessing damages to buildings and roads from satellite images as fast as possible. To this end, it is essential to detect buildings and identify the level of changes in them. Detecting building and their changes can also be used for urban planning, rescue planning, and preparation before disasters. Traditionally, these analyses are performed by experts, which is a time-consuming task [2] . With recent advances in computer vision, it is possible to analyze satellite images automatically with high accuracy and speed. In traditional methods, hand-crafted features such as color, shadow, edge, and roof texture were extracted from images, and a feature vector was generated for each sample. Then classification or clustering was done using classic machine learning algorithms [3, 4, 5, 6] . In more recent techniques, like convolutional neural networks (CNNs), features are extracted in a data-driven manner. Much effort has been done on utilizing CNN models and fusing deep and shallow features, in this problem. For example, a dual-stream network (DS-Net) that adaptively captures local and long-range information is proposed in [7] . The authors of [8] introduced a multi-scale fusion network to tackle the problem of different scales of buildings in remote sensing images. A Siamese fully connected network is proposed in [9] and it is shown that using two inputs can help the model obtain better segmentation accuracy. A PREPRINT-JULY 24, 2023 Using bi-temporal images is a common approach in problems such as change detection or building damage assessment. In this approach, images that are taken at two different times are compared in order to assess the difference between them. In general, combining the features extracted from bi-temporal images is performed at different levels as follows: 1. Segmentati

FAQs

Why do many existing building detection models use only the pre-disaster image?

They assume that using post-disaster imagery can reduce performance because destroyed buildings are present in the scene.

What is the core idea of SiamixFormer?

It is a Siamese model that takes both pre- and post-disaster images and uses a hierarchical transformer with a temporal transformer to fuse features across time.

How are temporal features used in feature fusion?

Queries are generated from pre-disaster features, while keys and values come from post-disaster features, so temporal dependencies are incorporated during fusion.

Sep 21, 2026

Uploaded by Jiven