Urban Growth Monitoring Using Temporal Statistical Composites, Deep Features and Max-Tree Spatial Filtering (Case Study: Tehran 2015–2025)

Abstract. This paper presents a method for monitoring urban growth in the Tehran metropolitan area using a decade-long Sentinel-2 image time series (2015–2025, images every two years). For each pixel in the blue band, four temporal–statistical indices including Range, Standard Deviation, Interquartile Range (IQR), and Quartile Coefficient of Dispersion (QCD)—were computed and stacked to create a 4-band composite representing temporal variability. Deep spatial features were then extracted from this composite using a convolutional neural network (CNN) to capture high-level spatial structures. Change detection was performed in the learned feature space and refined through spatial attribute filtering using a Max-Tree to eliminate small or noisy detections such as clouds and outliers. The proposed approach was applied to six epochs (2015, 2017, 2019, 2021, 2023, and 2025) across Tehran. The results revealed approximately 310.7 km² of built-up expansion during the study period. Accuracy assessment showed that the proposed deep-feature filtering achieved an overall accuracy of 92.3% and a Kappa coefficient of 0.89, outperforming the IQR-based statistical filtering (86.5% and 0.80, respectively). The CNN-based deep feature extraction provided a smoother and more spatially coherent detection of urban changes, especially in central and northeastern Tehran. These findings demonstrate that integrating deep spatial features with temporal–statistical indices significantly enhances the robustness and reliability of urban change detection.