Empowering Urban Design: Leveraging Deep Learning and Crowdsourced Data for Spatial Justice — a Mixed-Methods Case Study Of Constantine, Algeria

Rapid urbanisation in North Africa frequently neglects the diverse needs of citizens, resulting in spatial injustice. This research investigates the disconnect between top-down planning and actual space utilisation in Constantine, Algeria, by testing an integrated framework that combines citizen-generated data with deep learning techniques. Based on critical realism, a convergent parallel mixed-methods design was employed over a six-month period (January–June 2024), involving the recruitment of 300 residents (100 per district) through stratified quota sampling across the central, intermediate, and peripheral zones. Participants utilised a trilingual (Arabic/French/Berber) mobile application to submit geotagged observations of public space usage; a total of 8, 742 observations were collected. Concurrently, a ResNet-50 convolutional neural network was fine-tuned on Sentinel-2 imagery (10-meter resolution; eight land-use classes) using stratified 5-fold cross-validation, a class-weighted cross-entropy loss function, and Bayesian hyper parameter optimisation. Civic engagement was assessed through a validated 20-item composite index at baseline and conclusion; paired-sample t-tests with Cohen's d and bootstrapped 95% confidence intervals quantified changes. The composite civic behaviour score increased from 2. 14/10 to 7. 79/10, representing a 264% relative rise (95% CI: 241–287%; d = 2. 87; t (272) = 18. 42, p < .001; n = 273 completers), which should be interpreted as an association rather than causation given the single-group pre-post study design. The CNN achieved an overall accuracy of 88. 4% (95% CI: 86. 9–89. 8%), macro- precision of 86. 7%, macro- recall of 85. 9%, and a macro- AUC of 0. 94 on an independently held- out test set validated against ground- truth field observations. Peripheral Zone C displayed a 48. 3% discrepancy between planned and actual land use, compared to 26. 1% in the central zone, with significant spatial clustering (Moran's I = 0. 68, p