Yacine Bouroubi

Vegetated riparian buffers play a critical role in maintaining ecological health and water quality, yet efficient characterization and large-scale monitoring remain challenging due to the resource demands of traditional field campaigns. This study, conducted in an agricultural setting, introduces a straightforward, image-based methodology for riparian buffer characterization, exploiting advancements in deep convolutional neural networks (DCNN) and very high spatial resolution satellite imagery. Leveraging a large Riparian Strip Quality Index (RSQI) field dataset, the proposed approach adapts a Multi-View DCNN (MVDCNN) architecture, originally developed for 3D object recognition, to correlate satellite images of riparian strips with RSQI metrics. Of the seven spectral band combinations, multiple input views, and two training modes evaluated, the configuration using four views with RGB bands from a pretrained network achieved the best results. However, the alternative spectral band combinations produced similar levels of performance, suggesting that texture and shape information are key factors in the model’s effectiveness. Comparisons with a conventional workflow involving object-based land cover classification followed by RSQI calculation indicate that the trained MVDCNN achieves stronger correlations between imagery and RSQI scores (average RMSE = 7.35, R ² = 0.93 using RGB bands) compared to the object-based method (RMSE = 11.25, R ² = 0.87). To our knowledge, this is the first direct application of DCNNs to riparian buffer quality assessment. Requiring minimal preprocessing and no photo-interpretation expertise, the proposed approach leverages existing field data to facilitate more accessible, scalable, and adaptable riparian buffer monitoring, with potential for application in diverse environmental contexts.

Operational crop mapping requires classifiers capable of robust generalization across years. While feature importance is routinely used for model optimization, its temporal stability has rarely been systematically investigated, creating a critical gap in deploying reliable monitoring systems. This study moves beyond identifying “most important” features to systematically evaluate and quantify their inter-annual stability for enabling automated classification. Using six agricultural years (2018, 2019, 2020, 2023, 2024 and 2025) of Sentinel-1 and Sentinel-2 data over Morocco, we extracted 156 multi-sensor features across 12 monthly composites and analyzed their importance stability through statistical metrics, clustering, and novel composite indices: the Reliability Index (RI) and Automatic Selection Score (AuSS). This framework automates feature selection by ranking features with RI and AuSS and then applying Pareto optimization to identify a minimal stable feature set—without requiring annual retraining or expert intervention. Our analysis confirms a fundamental tension: the most discriminative features (e.g., NDVI, VH, VV) are also the most volatile, while stable features (e.g., NDRE, MSI, NDMI) offer modest predictive power. Hierarchical clustering revealed four behavioral typologies (Dominant Stable, Performant Volatile, Stable Minor, and Noise), guiding strategic feature management. Crucially, a Pareto analysis demonstrated that a refined portfolio of 6 indices (VH, VV, NDVI, NDRE, GCVI, RVI) captures 57.2% of cumulative predictive importance, filtering out inter-annual noise while preserving discriminative signal. The Voting Ensemble leveraging this Stable Portfolio maintained consistent high accuracy (87.4% accuracy, 87.2% F1-score) with minimal performance degradation during temporal transfer, while models based on volatile top features exhibited significant drops. Entropy analysis confirmed that all features in the Stable Portfolio provide consistent informational certainty, indicating that stability-driven selection does not increase model uncertainty. We conclude that feature stability is not merely a diagnostic metric but a foundational criterion for operational design. We propose a practical, metrics-driven framework for constructing automated crop classification systems that are more resilient to inter-annual climate variability.

Satellite observations provide critical data for a myriad of applications, but automated information extraction from such vast datasets remains challenging. While artificial intelligence (AI), particularly deep learning methods, offers promising solutions for land cover classification, it often requires massive amounts of accurate, error-free annotations. This paper introduces a novel approach to generate a segmentation task dataset with minimal human intervention, thus significantly reducing annotation time and potential human errors. ‘Samples’ extracted from actual imagery were utilized to construct synthetic composite images, representing 10 segmentation classes. A DeepResUNet was solely trained on this synthesized dataset, eliminating the need for further fine-tuning. Preliminary findings demonstrate impressive generalization abilities on real data across various regions of Quebec. We endeavored to conduct a quantitative assessment without reliance on manually annotated data, and the results appear to be comparable, if not superior, to models trained on genuine datasets.

Digital twins are increasingly gaining popularity as a method for simulating intricate natural and urban environments, with the precise segmentation of 3D objects playing an important role. This study focuses on developing a methodology for extracting buildings from textured 3D meshes, employing the PicassoNet-II semantic segmentation architecture. Additionally, we integrate Markov field-based contextual analysis for post-segmentation assessment and cluster analysis algorithms for building instantiation. Training a model to adapt to diverse datasets necessitates a substantial volume of annotated data, encompassing both real data from Quebec City, Canada, and simulated data from Evermotion and Unreal Engine. The experimental results indicate that incorporating simulated data improves segmentation accuracy, especially for under-represented features, and the DBSCAN algorithm proves effective in extracting isolated buildings. We further show that the model is highly sensible for the method of creating 3D meshes.

Digital twins are gaining in popularity for simulating complex natural and urban environments. In this context, accurate segmentation of objects within 3D urban environments is of crucial importance. The aim of this project is to develop a methodology for extracting buildings from textured 3D meshes. To this end, PicassoNet-II, a semantic segmentation architecture is employed. The methodology also incorporates Markov field-based contextual analysis to assess post-segmentation features. In addition, building instantiation is performed using cluster analysis algorithms. Training this model to fit various datasets requires a large amount of annotated data, both from Quebec City, Canada, and from simulated data. Experimental results show that the use of simulated data improves segmentation accuracy, and the DBScan algorithm proves effective in extracting isolated buildings. This project paves the way for improved applications in 3D urban modeling, offering opportunities in fields ranging from urban planning to resource management. The positive results with simulated data reinforce the impact of this research on improving digital models of our ever-changing urban environments.

Knowledge of demographic data is valuable information for planning initiatives. Typically, census, survey, and population projection exercises provide this information. In some developing countries, these operations pose a variety of economic and logistical challenges, thereby depriving authorities of accurate and timely information on their populations. To provide approaches for solving this situation, our study evaluates a population estimation method that is based on detection of residential geo-objects (houses) on very-high-resolution (VHR) satellite images using convolutional neural networks (CNN). The approach would be applicable to countries where a complete census is difficult to perform due to resource constraints or political instability. A 2008 VHR satellite image of Sudan is annotated according to seven classes of buildings to create a dataset that was used to train an object detection model, faster region-based CNN, by transfer learning. The model obtained mean average precision of 79% and 99% during training and validation, respectively. This unusual difference is due to the dominance of well detected classes in the validation dataset. The model was fine-tuned to detect the same building classes on images in 2021. A link between residential geo-objects and population size was established using 2008 population data and available field data. Subsequent characterization of the current population should assist in preparation of the 2023 census. Limitations of this approach were raised, but it could be used to improve the framework for population data collection in developing countries.