A Deep Learning Classification Approach Using High Spatial Satellite Images for Detection of Built-Up Areas in Rural Zones: Case Study of Souss-Massa Region, Morocco

A Deep Learning Classification Approach Using High Spatial Satellite Images for Detection of Built-Up Areas in Rural Zones: Case Study of Souss-Massa Region, Morocco Miriam Wahbi1, Insaf El Bakali1, Badia Ez-zahouani1, Rida Azmi2, Abdelmalik Moujahid3, Mohammed Zouiten4, Otmane Yazidi Alaoui1, Hakim Boulaassal1, Mustapha Maatouk1, Omar El Kharki1 1Geoinformation and Territorial Management Unit, FSTT, Abdelmalek Essaadi University, Tetouan, Morocco 2Mohammed VI Polytechnic University, Center of Urban Systems (CUS), Ben Guerir, Morocco 3Department of Mathematics, University of the Basque Country UPV/EHU, Leioa, Spain 4Faculty Polydisciplinary – Geography Department of TAZA, LISA Laboratory of ENSA-Fez, Sidi Mohamed Ben Abdellah University, Fez, Morocco Published in: Remote Sensing Applications: Society and Environment, Volume 29, 2023, Article 100898, Elsevier. DOI: 10.1016/j.rsase.2022.100898. Available online 13 December 2022. Abstract The buildings in the rural areas of Morocco exist in various shapes and sizes and are generally constructed of primary materials such as clay, wood, and tin. Their detection is difficult and often inaccurate with traditional optical satellite imagery and image processing techniques, especially in rural settlements. This study aims to detect and map the settlements in rural areas of the Souss-Massa region using Sentinel-2 satellite images and deep learning algorithms. The baseline architecture tested is the convolutional neural network UNet; its performance was evaluated, the number of filters in the convolution layers was increased, and a deep Residual UNet (ResUNet) was also implemented. Model quality was assessed using precision, recall, F1-score, and ROC curve. Results showed the UNet with increased filters outperformed other models, achieving 87% precision and 54% F1-score. A comparison with other related studies using different machine learning and deep learning algorithms is presented. Findings highlight that deep learning performance for settlement detection in rural areas depends strongly on the number and quality of training images, as well as network design parameters such as the number of filters in convolutional layers. Main Contributions •Proposes and evaluates a deep learning approach (based on U-Net and ResUNet) for mapping and detecting rural settlements in the Souss-Massa region using 10 m resolution Sentinel-2 satellite imagery. •Demonstrates that increasing the number of filters in the convolution layers of the U-Net improves precision and F1-score, outperforming both standard U-Net and deep ResUNet in this context. 1 •Analyses the influence of network architecture and training dataset quality/quantity on performance, emphasizing their crucial role in rural building and settlement detection from remote sensing images. •Provides benchmarking and a critical review comparing results to several state-of-the-art methods and case studies for rural and urban building extraction using deep and classical machine learning approaches. •Shows that for the challenging conditions of rural Morocco, the enhanced U-Net architecture (with increased filters) is particularly effective, reaching up to 87% precision, but identifies limitations due to training database size and image resolution. Impact of the Paper This research provides significant advances for the application of deep learning to the detection and mapping of rural settlements using open-access, high-resolution satellite imagery. The presented methodology improves the accuracy and reliability of rural building extraction, which is traditionally challenging due to the heterogeneity of rural structures and limited data availability. The impact of this work extends to several domains: •Technological innovation: Demonstrates that convolutional neural networks (U-Net and ResUNet) can be effectively applied to Sentinel-2 images for mapping rural settlements with high precision, even under lower spatial resolutions, setting a precedent for the use of freely available Earth observation data. •Territorial management and rural development: Offers practical tools for policymakers, local governments, and geospatial analysts to inventory and monitor rural settlements, supporting better resource allocation, infrastructure planning, and rural connectivity in regions like Morocco where such data are scarce. •Sustainable land management: Facilitates large-scale settlement monitoring and change detection, which is crucial for sustainable planning, agricultural management, and disaster recovery in rural areas, as satellite-derived data can identify trends and anomalies that influence land use and environmental health. •Methodological reproducibility: The approach using open imagery and relatively simple neural architectures can be replicated in other regions facing similar limitations, potentially broadening the benefit to other under-mapped and data-poor rural areas worldwide. By enhancing rural mapping accuracy, this paper helps bridge the information gap faced by many developing regions. The methodology supports future integration with digital rural strategies, satellite-enabled agriculture, and sustainable development initiatives, reinforcing the role of remote sensing and artificial intelligence in societal applications. Reference: Miriam Wahbi, Insaf El Bakali, Badia Ez-zahouani, Rida Azmi, Abdelmalik Moujahid, Mohammed Zouiten, Otmane Yazidi Alaoui, Hakim Boulaassal, Mustapha Maatouk, Omar El Kharki. “A deep learning classification approach using high spatial satellite images for detection of built-up areas in rural zones: Case study of Souss-Massa region - Morocco.” Remote Sensing Applications: Society and Environment, Volume 29, 2023, Article 100898. DOI: 10.1016/j.rsase.2022.100898 2