Early and real-time detection of rice leaf diseases (RLD) poses a significant challenge for farmers, especially in rural regions with limited access to advanced technology. Conventional deep learning models often require substantial computational resources, rendering them impractical for deployment on mobile or edge devices commonly used in agricultural environments. Although models such as ResNet50, VGG16, and InceptionV3 can achieve high accuracy, they are computationally expensive and may be less suitable for real-time deployment in smart agricultural systems. Furthermore, many existing studies rely on limited datasets, which can restrict model generalizability under diverse real-world conditions. Although transfer learning can improve classification performance, developing lightweight models suitable for real-time IoT deployment remains challenging. To address these limitations, we curated a hybrid dataset of 7,092 rice leaf images by combining self-collected and Kaggle samples and proposed IoT-RiceMobileNet, a lightweight improved MobileNetV2-based model that can be effectively integrated into our developed IoT system. The proposed model outperformed transfer learning and deep learning baseline models, achieving 99.19% test accuracy and 99.20% precision while maintaining low computational complexity. We further confirmed model stability using stratified 5-fold and 10-fold cross-validation on both the constructed RLD and multi-source datasets, achieving mean accuracies of 98.04% and 98.13% on the constructed RLD dataset and 98.11% and 98.58% on the multi-source dataset, respectively. Moreover, our models compact size and 85.17 FPS inference speed support real-time deployment. Finally, the model was integrated into an IoT-enabled mobile, web, and cloud-based inference framework, demonstrating its practical potential for scalable rice disease detection in smart agriculture.

Knowledge graphs (KGs) built from scientific literature capture how concepts are related, but their edges are semantic rather than causal, and large language models (LLMs) asked to extract causal relations from text may conflate correlation with causation or propose unsupported directions. We present a framework that transforms literature-derived KGs into causally meaningful structures. An LLM filters KG nodes to identify candidate causal variables, while statistical causal discovery (SCD), applied to observational datasets from the same domain, independently recovers the causal structure among them. Building on the existing KG pipeline, our toolkit integrates ten causal-discovery methods across independent and identically distributed (i.i.d.) and time-series regimes, including latent-confounder-aware algorithms. On two synthetic benchmarks with known ground truth, the integrated methods reliably recover causal structure, and an end-to-end execution on a real-world polar-science case study confirms that the full pipeline composes correctly, yielding interpretable causal graphs grounded in both literature and data.

Driver behavior detection is essential for improving road safety in Intelligent Transportation Systems (ITSs). This paper proposes an attention-enhanced CNN integrated with Explainable Artificial Intelligence (XAI) to achieve reliable and interpretable driver behavior classification. Furthermore, the model also uses a multi-head attention mechanism to ensure that it captures local spatial features as well as long-range dependencies in the activities of the driver. The State Farm Distracted Driver dataset is experimented with using stratified 5-fold cross-validation. The proposed MHA-CNN model has a mean accuracy of 99.48% on all folds and an overall high levels of precision, recall, and F1-score in all ten distraction classes. Grad-CAM is used to produce attribution maps to improve interpretability by showing semantically important parts of the image, including hands, face, and mobile devices, and can give a visual explanation of why models make decisions. In addition, an ablation investigation is performed with k-fold validation to assess the value of the multi-head attention mechanism. Its model can run at 122 frames per second with 42.92 million parameters, showing that the suggested method is a reasonable tradeoff between accuracy, efficiency, and transparency and can be applied in real-time driver monitoring applications.