Document Type : Original Research Paper
Authors
1 School of Surveying and Geospatial Engineering, College of Engineering, University of Tehran. Tehran. Iran
2 School of Surveying and Geospatial Engineering, College of Engineering, University of Tehran, Tehran, Iran
Abstract
Background and Objectives: With the rapid expansion of urbanization, the need for automatic updating of change maps has become increasingly important. Accurate and up-to-date spatial information is essential for monitoring construction activities and tracking the development of urban areas. Traditional approaches to change detection are mostly limited to two-dimensional analysis and often lack sensitivity to vertical changes. This shortcoming fails to identify multi-story constructions, thereby limiting the completeness of monitoring outcomes. Recent advances in remote sensing and deep learning have enabled three-dimensional urban change detection, providing superior results compared to classical methods. This study aims to improve the performance of 3D urban change detection by introducing a deep learning approach that integrates multi-source data. The primary objective is to automatically identify and distinguish four types of building-related changes—new construction, complete demolition, height increase, and height decrease—alongside unchanged areas, to generate a comprehensive 3D change map.
Methods: The dataset employed in this research consists of high-resolution RGB aerial imagery and corresponding Digital Surface Model (DSM) data acquired from two different periods over Valladolid, Spain. The input data were prepared by stacking RGB images and DSMs from both epochs into an eight-band input, allowing the network to jointly analyze spectral and elevation information. The dataset was divided into training (90%) and testing (10%) subsets. To increase variability in the training data and reduce overfitting, augmentation techniques such as horizontal and vertical flipping, random rotation, and Gaussian blurring were applied. The proposed model architecture combines a ResNet-34 backbone for feature extraction with a UNet++ decoder for pixel-level change reconstruction. Model parameters were updated using the Adam optimizer. In the first stage, the deep network was trained in a binary setting (change/no-change) and evaluated against classical approaches, including Random Forest, image differencing/ratioing, and a PCA–K-Means hybrid method. In the second stage, the network was retrained for five-class classification, including the four change categories and the unchanged class, using a loss function optimized directly for the Intersection-over-Union (IoU) metric. Model performance was assessed using Accuracy, Recall, Precision, and F1-score.
Findings: In the binary classification stage, after 50 epochs of training, the network successfully identified most real changes while maintaining a low false alarm rate. Evaluation metrics confirmed this performance, with Recall and Accuracy both reaching 98.5% and an F1-score of 0.92, considerably outperforming the classical methods. Unlike traditional approaches, the deep learning model was able to detect almost all small-scale constructions and demolitions. In the five-class stage, the model effectively identified and classified change types, achieving a Recall of 96.32%, an Accuracy of 96%, and an F1-score of 0.95. All newly constructed and fully demolished buildings were correctly labeled in the output maps, and a large proportion of unchanged areas received no misclassification.
Conclusion: The findings demonstrate that combining elevation data with 2D imagery and leveraging deep learning architectures significantly mitigates the limitations of traditional change detection approaches and enhances accuracy. The developed model is capable of detecting not only the location but also the type of change. This approach has strong potential applications in monitoring unauthorized constructions, updating spatial databases, and assessing urban development. However, its effectiveness relies on the availability of accurate DSM data, which may not be consistently accessible for all urban areas. Additionally, the training of deep networks requires extensive labeled datasets and considerable computational resources, which could limit their applicability in operational contexts.
Keywords
Main Subjects
COPYRIGHTS
© 2025 The Author(s). This is an open-access article distributed under the terms and conditions of the Creative Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)
)