Water resources face increasing pressure from climate change and human activities. This study presents a Smart Water Resilience System (SWRS) for Morocco’s Oum Er-Rbia Basin, integrating GeoAI, GIS, and multi-source remote sensing data to monitor, forecast, and assess hydrological and vegetation dynamics. Analysis of satellite-based indicators from 2015 to 2022 showed that Random Forest land cover classifications achieved high reliability, with overall accuracy from 0.93 to 0.98 (93–98 ^2 in water surface area during this period, reflecting both seasonal and interannual variability. Three deep learning models—GRU, Attention-LSTM, and CNN-LSTM—were trained on spring-season (March–May) data from 2015 to 2022 to forecast key indicators, including NDVI, precipitation, land surface temperature, turbidity, chlorophyll, groundwater change, runoff proxy, discharge, SPI, CWSI, resilience score, flood risk, IWVI, VHI, evapotranspiration, and soil moisture for 2023–2030. Among these, CNN-LSTM achieved the highest predictive accuracy at the basin scale. The derived Integrated Resilience Index (IRI) highlighted spatial and temporal vulnerabilities, providing a comprehensive basis for interactive water management and resilience planning.
We present an ML pipeline for early dropout prediction in Moodle using 7 GB of logs from 3,120 students, 62 courses, across 5 academic terms. Weekly features (activity, achievement, participation) feed classical classifiers (RF, GBM, LR, SVM, NN) and an LSTM for temporal dynamics. To avoid leakage, features are restricted to information available up to prediction week t, and evaluation uses temporal hold-out (earlier terms → later term) and cross-course splits. On the temporal test set at week 4, the best ensemble achieves F1 = 0.82 (95
Personal Health Knowledge Graphs (PHKGs) are transforming the landscape of personalized healthcare by integrating a variety of data sources, including electronic health records (EHRs), clinical databases, wearable device data, and environmental information, into comprehensive, semantically rich frameworks. This article provides an in-depth survey of the methodologies behind PHKG construction, highlighting the integration of diverse data, the use of standardized ontologies, and the application of advanced knowledge extraction techniques. It examines the significant applications of PHKGs in clinical decision support, health monitoring, personalized recommendations, and disease prediction. By identifying key challenges, synthesizing current advancements and mapping future research opportunities, this study aims to establish PHKGs as a cornerstone for achieving precision medicine and proactive care.
Sustainable fisheries management increasingly relies on predictive insights, which can help anticipate changes in marine resources. Although remote sensing technologies now offer continuous monitoring of ocean conditions, the complex relationships between environmental variability and fish population dynamics remain insufficiently captured by current models, especially in regions with contrasting marine ecosystems. In this study, we introduced an integrated predictive framework that combines satellite-derived environmental parameters with machine learning algorithms to predict fish catch across Moroccan waters. The model was validated using official data from the Moroccan Ministry of Fisheries and Maritime Affairs. To achieve this, we implemented a comprehensive analytical pipeline for the period 2005–2023 using MODIS-Aqua parameters: sea surface temperature (SST), chlorophyll-a (CHL-A), particulate backscattering (BBP), diffuse attenuation coefficient (KD490), photosynthetically active radiation (PAR), particulate organic carbon (POC), and normalized fluorescence line height (NFLH). Eight regression models were trained and optimized separately for the Atlantic and Mediterranean regions, including both linear baselines and tree-based algorithms. The analysis revealed that ensemble models, trained on feature-selected parameters, predicted fish catch better than linear approaches. Random Forest achieved the best performance in the Atlantic (R2 = 0.599, RMSE = 23,813, MAE = 17,158), while Gradient Boosting outperformed in the Mediterranean (R2 = 0.568, RMSE = 698, MAE = 547). The Atlantic’s dynamic conditions favor nonlinear ensemble learning, whereas the more stable Mediterranean ecosystem benefits from gradient-based optimization. Overall, this research demonstrates the strong potential of combining remote sensing with machine learning to improve marine resource assessment.
The rapid expansion of social media, in the absence of strict regulatory oversight, has significantly transformed online communication dynamics. This evolution has also facilitated the proliferation of automated and malicious accounts, commonly referred to as bots. Cryptocurrency-centered communities, which constitute the focus of this study, are particularly exposed to such threats, often manifesting through scams, misleading investment schemes, and disinformation campaigns. This research presents a methodological framework for the proactive identification of suspicious accounts in multilingual cryptocurrency-related discussions. The proposed approach combines behavioral indicators with textual features, leveraging machine learning paradigms. A multilingual dataset comprising over 10,000 Facebook comments was curated and enriched with linguistic and behavioral metadata. Two complementary architectures were evaluated: fastText lexical embeddings coupled with a Light classifier and DistilBERT contextual embeddings combined with a Multi-Layer Perceptron (MLP). After optimization, the first model achieved an AUC of 96
