Global agricultural productivity reduced significantly due to frequent outbreaks of plant disease. To reduce crop losses, timely identification is essential. With advancements in artificial intelligence, especially in deep learning, the identification of plant diseases has seen significant improvements, with image segmentation enabling precise localization of disease symptoms. The study reviews segmentation techniques from classical image processing, machine learning, and deep learning-based methods. Further, deep learning methods are broadly divided into semantic, instance, and panoptic segmentation approaches. Key models are identified in these categories, and their relative strengths and weaknesses are discussed. An analysis of 106 studies from 2020 to 2025 is conducted, focusing on their objectives, strategies of implementation, and segmentation methodologies applied. The finding show U-Net and Mask R-CNN are widely adopted architectures because of their robustness across complex visual scenes. Hybrid models that employ multiple segmentation techniques have shown promising results in improving accuracy and diagnosis reliability. Several challenges still exist such as the lack of annotated datasets, high computational overhead, and difficulties in generalizing to unseen conditions. Future research directions are suggested with the aim of surmounting the challenges identified and encouraging further development. This work integrates current spectrum of segmentation techniques while discussing the practical constraints that hamper the development of an efficient architecture in precision agriculture. The study provides a foundational framework for developing scalable and reliable segmentation models for smart agriculture.

Agriculture is essential to the economy and social well-being of any community. Early diagnosis of plant diseases is critical to crop protection and sustainable agriculture. Deep learning (DL) has demonstrated outstanding performance in disease detection, but its complexity has made field implementation difficult. To overcome this issue, lightweight DL models have been developed. This paper presents a comparative analysis of lightweight DL models for plant disease classification on three publicly available data sets: PlantDoc, Mango, and Soybean Leaf Disease. The PlantDoc dataset has 28 classes with different crops and disease conditions captured under real field conditions. The Mango data set has images of 7 disease classes and 1 healthy class, and the Soybean data set has 2 disease classes and 1 healthy class, with multiple leaves per frame, depicting real-world scenarios. This study compares lightweight classification models like ShuffleNetV2, MobileNetV2, and MobileNetV3Small and object detection models YOLOv8n and YOLOv11n, modified for classification via bounding box class filtering. The models were evaluated using conventional evaluation metrics such as accuracy, precision, recall, F1 score, and Cohen’s kappa score; additionally, model size and number of parameters are considered in comparison. Experiments demonstrate that MobileNetV2 achieves the highest accuracy of 89.86

The exponential growth of the Internet of Things (IoT) has significantly broadened the attack surface for cyber threats, necessitating robust and adaptive security frameworks. This study introduces a novel suite of hybrid deep learning-based Intrusion Detection System (IDS) models tailored for large-scale IoT environments. Specifically, we propose and optimize three IDS models-CNN-LSTM, GRU-AE, and Bi-LSTM-CNN-leveraging hybrid architectures to address the challenges of real-time threat detection, computational limitations, and adaptive learning. The models are evaluated on two benchmark datasets, BoT-IoT and CICIDS2017, achieving average accuracies of 97.8%, 98.3%, and 98.6%, and F1-scores of 97.5%, 98.1%, and 98.4%, respectively. These results demonstrate a performance improvement of 6%-12% in accuracy and 8%-15% in F1-score over conventional IDS methods, with false positive rates consistently below 2.5% and detection rates exceeding 98.7%. The proposed models are optimized for lightweight deployment and low-latency detection, reducing inference time by up to 35% and energy consumption by 22% compared to baseline deep learning models-ensuring feasibility in resource-constrained environments. Furthermore, the study explores mechanisms for continuous learning, enabling a 15%-20% improvement in adaptability to previously unseen attack patterns. Interpretability is enhanced through the integration of SHAP (SHapley Additive exPlanations) values, with over 92% of critical prediction decisions explained by the top 10 contributing features. A comparative analysis highlights not only performance gains but also the models' robustness, scalability, and resilience under adversarial conditions. The results confirm that the proposed hybrid IDS models offer a scalable, efficient, and interpretable solution for enhancing IoT network security. This work contributes a comprehensive and deployable framework to address the evolving landscape of cyber threats in the IoT ecosystem.

Soil fertility assessment is a key component of soil fertility enhancement for better crop productivity and sustainable farming. The traditional methods for assessing soil fertility, such as chemical analysis techniques, are frequently expensive, time-consuming, and unsuitable for widespread application in agricultural monitoring. To overcome these constraints, this paper introduces Soil Sense Predictor, a machine learning-based soil fertility classification system based on the Random Forest algorithm. Several soil parameters were used for the fertility assessment using data preprocessing, feature analysis, model development and validation. The standard data partitioning and cross-validation approaches were used to train and validate the model for reliable predictive performance. The proposed framework was evaluated against some widely used supervised machine learning models, such as Naïve Bayes, Decision Tree, K-Nearest Neighbours, Support Vector Machine and Logistic Regression. The best overall performance was achieved by Random Forest with accuracy of 89%, precision of 90%, recall of 97% and an F1-score of 93% in the experimental results. The results suggest that machine learning has potential to help soil fertility assessment as a reliable decision-support system in precision agriculture. The proposed approach can support farmers to manage the use of fertilizers more efficiently, enhance crop productivity, reduce resource losses, and increase sustainable farming practices.

Soil plays a foundational role in sustaining agricultural productivity and ecological stability, yet traditional soil analysis methods remain labour-intensive, slow, and often inadequate for real-time decision-making in modern precision agriculture. With the rise of Agriculture 4.0, machine learning (ML) and Internet of Things (IoT) technologies offer transformative potential for accurate, scalable, and data-driven soil assessment. Current research in this domain remains relatively fragmented, with a lack of clarity in models, sensing methodologies, and algorithmic strategies that produce the highest accuracy and operational value across diverse agricultural contexts. To address this gap, this research used a PRISMA-guided systematic literature review to systematically examine contemporary machine learning (ML) and IoT-based soil analysis methodologies. The review utilized sophisticated search queries across databases: Scopus, IEEE Xplore, ACM Digital Library, ScienceDirect, and Google Scholar, for identifying studies published between 2019 and 2024. After rigorous screening that involved removing duplication and full-text assessment of the retrieved entries, 77 high-quality articles were identified that met the eligibility criteria from an initial set of 180 entries. Data extraction was performed under descriptive and thematic synthesis, thereby allowing the comparative evaluation of supervised and unsupervised learning models, IoT sensing frameworks, soil parameters, dataset characteristics, evaluation metrics, and deployment constraints. Comparative analysis revealed that the use of supervised models, such as Random Forest (RF), Support Vector Machines (SVM), Gradient Boosting Machine (GBM), Convolutional Neural Networks (CNN), and deep ensembles, produces higher accuracy in the classification of soil quality, fertility, pH, and nutrient levels, especially in structured datasets like the Soil Fertility Dataset. IoT-based sensing systems significantly improve the reliability of predictions by offering continuous and detailed measurements of soil moisture, nutrient status, and environmental conditions.

Low visibility, less contrast, and warped object boundaries make current detection algorithms far less effective, making it impossible to find objects in foggy environments. This work builds a novel integrated architecture that includes temporally optimized deep feature extraction, haze-aware image clarity, and an enhanced YOLOv8 detector specifically intended for foggy environments to address these limitations. Differential characters of images that are hazy are compared to other natural images which restrict generalization of present Image Quality Assessment (IQA) algorithms. Initial intention of this work is designing and developing techniques for IQA and object detection with hazy outdoor images. Here, hazy input image is first directed deal to propose Haze aware Structural Pixel Neighbor (HSPN) features, Color rendition, and Mean Subtracted Contrast Normalized (MSCN) coefficients. Also, assessing image quality is done using Deep Convolutional Neural Network (DCNN) that is trained by proposed optimization algorithm, namely Chronological Chimp optimization Algorithm (CChOA). Moreover, this developed CChOA is formed by integrating the Chronological concept with Chimp Optimization Algorithm (ChOA). Furthermore, assessed image quality is further tends to object identification using You Only Look Once stage-9 (YOLOv8), which is based on Neural Network (NN). Finally, object is identified and performance of this model is enhanced by evaluating with three performance metrics, such as Mean Seismic Data Structural Similarity (MSDSS), Signal-to-Noise Ratio (SNR), and Structural Similarity Index (SSIM) with values of 0.944, 50.769, and 0.925, respectively.

The advent of blockchain technology has brought about an enormous shift in financial transactions, while the extensive integration of cloud computing in IT systems has experienced substantial growth due to its efficiency and availability. Cloud computing continues to face challenges, although there have been efforts made to address issues around privacy and security. This paper examines the potential of blockchains, renowned for their decentralized nature and consensus-based software architecture, to provide effective solutions for distributed applications, and hence address issues commonly associated with centralized systems. The study examines the legal and technical dimensions to investigate how blockchains can effectively tackle the issues of cloud computing. In addition to proposing solutions to improve security and privacy, the current use of blockchain technology in cloud storage applications is examined by comparing all security techniques with simulated blockchain. By leveraging the decentralized attributes of blockchain technology, we aim to overcome the limitations of centralized cloud systems. It investigates the potential of this technology to effectively circumvent the constraints of traditional cloud computing platforms.

The relentless growth of connected devices is transforming industrial, urban and domestic environments, yet it also expands the attack surface for distributed denial of service (DDoS), unauthorized access and data manipulation. Centralized security architectures struggle to cope with the scale and heterogeneity of the Internet of Things, creating single points of failure and privacy risks. This review takes a close look at how blockchain and artificial intelligence (AI) can work together to solve these problems. Blockchain plays an important role in decentralizing trust, maintaining data integrity, and enabling transparent audit trails. AI subfields such as machine learning (ML), deep learning (DL), reinforcement learning (RL), and multi-agent systems (MAS) enhance these benefits. They enable real-time anomaly detection, predictive analytics, and adaptive policy control. A seven axis Blockchain–AI Security Integration Schema (BASIS) is proposed to classify solutions by security objectives, intelligence modalities, trust primitives, deployment choices, scalability techniques, privacy controls and interoperability mechanisms. In this study also review Layer-2 consensus protocols, federated learning and lightweight deep learning models that address energy and computational constraints. Case studies from supply chains, healthcare and smart grids illustrate the benefits and limitations of current deployments. The evidence suggests that while AI improves the accuracy and responsiveness of threat detection, blockchain offers tamper-proof data provenance. However, there are still issues in achieving scalability, reducing computational overhead, and striking a balance between auditability and privacy. Hybrid on-chain/off-chain architectures, quantum-safe cryptography, and standardized frameworks to guarantee adoption and interoperability are some future research avenues.

An Intrusion Detection System (IDS) is a commonly employed security mechanism for detecting, mitigating, and minimizing the impact of concealed and unrecognized intrusions in the Internet of Things (IoT). This study proposes a novel hybrid Deep Maxout–Quantum Neural Network (Maxout–QNN) intrusion detection system to address high-dimensional data, class imbalance, slow convergence, and premature stagnation in IoT security applications. Two optimisation-based variants are introduced: Hybrid + Self-Upgraded Cat and Mouse Optimisation (SUCMO) and Hybrid + Seagull-Adopted Elephant Herding Optimisation (SAEHO). The proposed IDS operates through three fundamental stages: preprocessing, feature extraction, and classification. The input data are first subjected to enhanced Z-score normalization during preprocessing. Afterward, relevant features—including statistical and higher-order statistical measures, enhanced entropy-based, and correlation-based attributes—are extracted. Finally, classification is performed using a hybrid Deep Maxout–QNN model based on the extracted features. Existing IoT intrusion detection systems often struggle with high-dimensional data, class imbalance, and limited scalability. To overcome these challenges, this work introduces a hybrid Deep Maxout–QNN model optimized using the SUCMO algorithm. The proposed SUCMO and SAEHO algorithms dynamically adjusts exploration–exploitation parameters to accelerate convergence and prevent premature stagnation. This mechanism distinguishes SUCMO and SAEHO from conventional metaheuristics, ensuring faster convergence and higher accuracy. The proposed Hybrid + SUCMO model achieved 96.65

Abstract Noise removal means eliminating noise from a noisy image, thereby improving the quality of original image. Elimination of noise from the input signal remains a major issue for investigators. With the increasing number of digital images captured day-to-day, the requirement for more accurate perceptibly appealing image is enhancing. However, images taken by contemporary cameras are automatically deteriorated by noise, resulting in degraded visual quality. In general, retrieval of essential data from noisy images in the procedure of denoising to acquire the best quality of images is a big issue today. Therefore, work must be done to eliminate noise in the image without falling image characteristics, like edges, corners, and other sharp frameworks. Hence, this research paper overviews numerous techniques for image denoising and image quality enhancement. This overview investigates 50 Research papers related to noise removal and image quality enhancement, and developed technique-wise reviews, namely spatial domain-based techniques, optimization-based approaches, transform domain-based approaches deep learning (DL)-based techniques and machine learning (ML)-based techniques. An investigation participate in a survey based upon classifying experiment approaches, datasets, year of publication, toolset employed, effectual metrics for image denoising and image quality improvement. Finally, the challenges of overviewed techniques are illustrated to concentrate investigators for developing various efficient techniques for image denoising and image quality improvement.

Pneumonia remains a serious worldwide health concern, particularly in low-resource countries, where prompt diagnosis is challenging. Early detection relies on chest radiography; however, data privacy rules and patient data fragmentation make it difficult to build AI models. Federated Learning allows for collaborative model training without sharing patient data, a promising solution. Standard federated learning methods, such as FedAvg, suffer from data heterogeneity and significant communication overhead. To overcome these constraints, this research proposes an upgraded federated framework with FedProx, which mitigates client drift in non-IID contexts by proximal optimization and Low-Rank Adaptation, a parameter-efficient fine-tuning technique that minimizes communication costs. Vision Transformers are utilized as the backbone architecture for chest X-ray categorization because they capture the global visual context more effectively than convolutional models. The proposed technique was validated for a pneumonia classification job utilizing the publicly available Chest X-Ray Images dataset, which was distributed across simulated clients to replicate real-world healthcare organizations. The model’s performance is measured using accuracy, precision, recall, F1-score, AUC, and system-level measures, including communication cost per round and convergence rate. Under conditions of non-IID heterogeneity of data, the proposed FedProx+LoRA framework demonstrated a classification accuracy of 88.5 which was higher compared to the centralized baseline (63.9

The rapid increase in fake social media accounts, particularly X platform (Twitter), has been a significant issue. Fake profiles can deceive people or groups of people, disseminate falsehoods, destroy credibility, sell fake news, support evil actions, and control online interactions. Fake accounts are highly developed, thus necessitating automated, sophisticated methods to detect them manually. This paper presents a powerful deep learning model to identify false profiles in multimodal Twitter data. The architecture fuses LSTM to analyse text, CNN to analyse visual information, and ANN for metadata characteristics with Adam optimizer setting a state-of-the-art accuracy of 97.2060

Recent advancement in artificial intelligence and deep learning technologies has made automatic diagnosis of plant disease a reality, which offers a better alternative to traditional manual methods. However, the large size of many existing deep learning models limits their deployment on resource-constrained platforms such as IoT or mobile devices. This study proposed LiteCShuffle (LCS), a lightweight model based on the standard ShuffleNetv2 model, to solve this issue. The LCS model utilized channel attention mechanisms and channel shuffle in its inverted residual block for improving the model’s ability to extract relevant features at various stages. The proposed model performance is assessed on the PlantVillage dataset, which includes 39 classes across 13 crop types. The LCS model shows a notable enhancement in plant disease classification, attaining an accuracy of 99.86% while having only 0.15 million trainable parameters and a total size of 0.58 MB. In comparison to ShuffleNetV2, which attains an accuracy of 99.68% with 1.29 million parameters and a model size of 15.4 MB. The proposed model uses fewer parameters and is less complex. Additionally, it surpasses other frequently employed models, such as VGG16, EfficientNetV2s, DenseNet201, SqueezeNet, and MobileNetV2. The model’s effectiveness for real-time applications is demonstrated by its 28 millisecond latency on the Nvidia Jetson Nano, which is far lower than ShuffleNetV2 models. Code at: https://github.com/Dsangeeta97/Litechsuffle/.

An enhanced version of the YOLO-NAS object detection network model has been presented in this paper, and MISH activation and Artificial Bee Colony (ABC) optimization algorithms are integrated. MISH functional adds non-monotonic behavior, which at the same time enhances the feature representation and complements the gradient flow. ABC optimization that assists in the optimization of the hyperparameters to a ground truth and resistance to the models. The given model is tested on the dataset that is introduced by the researchers themselves, and it shows better results compared to baselines based on the YOLO-NAS variants in precision, recall, and mean average precision (mAP) measures. Experiments prove the fact that a combination of a biologically inspired optimizer and a contemporary activation function helps to make training more stable and predictions more accurate. The results show that the proposed fine-tuned YOLO-NAS model outperformed the other tested models, that is, YOLOv6, YOLOv7, and YOLOv8, in the three metrics of accuracy, recall, precision, F1 score, and mAP at 0.50, 0.75, and 0.95 on the test dataset. The proposed model achieved an accuracy of 98% while recognizing real-time objects.

Smart grids integrate real-time communication, computational intelligence, and physical energy infrastructure to improve operational efficiency and adaptability. However, this interconnectivity increases vulnerability to cyber intrusions capable of disrupting control signals, compromising data integrity, and causing large-scale outages. This study presents an architectural synthesis that combines a long short-term memory network with an attention mechanism for temporal-saliency feature extraction, followed by an ensemble of gradient-boosting classifiers (XGBoost, LightGBM, and CatBoost) for robust decision-making. Experiments on the benchmark smart grid intrusion detection dataset demonstrate that the integrated framework achieves 79.8% average cross-validation accuracy and 75.67% overall test accuracy, with a normal-class recall of 98.26% in the baseline configuration. To address severe class imbalance, a combination of synthetic minority oversampling technique and focal loss was applied, which improved minority attack-class recall from 1.43 to 64.3% and increased its PR-AUC from 0.2884 to 0.791. The balanced configuration yielded an receiver operating characteristic curve (ROC-AUC) of 0.928 for both classes, demonstrating substantial gains in minority-class detection while maintaining high precision for majority classes. These results highlight the potential of strategically combining temporal modelling, attention-driven interpretability, and ensemble diversity, augmented with targeted imbalance mitigation, to develop effective, scalable, and interpretable intrusion detection systems for critical energy infrastructure.

Present MLL (Multimodal Large Language) models do exceptionally well in computer vision tasks, such as answering virtual questions and captioning images. However, they are not up to par in important perceptual tasks like object detection. To overcome this constraint, an entirely novel research work is presented for contextual object detection, which aims to comprehend observable items in various human-AI interaction scenarios. This study investigates two widely used scenarios: captioning the images and virtual question answering. For human-machine interaction, a brand-new CODNet (Context-based Object Detection Network) is presented that can locate, recognize, and connect visual objects with spoken inputs. The proposed network is capable of end-to-end modeling of visual language contexts. This network unifies language and vision functions by interpreting images in a foreign language and coordinating language functions with vision-centric jobs that may be managed and created flexibly using language instructions. The network comprises three sub-modules: an encoder that extracts the required features from the context, a pre-trained large language model meta AI (Llama) for decoding the context, and a decoder for predicting anchor boxes around the contextually recognized objects in images. The features are extracted with the help of a convolutional neural network merged with multiplicative LSTM. A fine-tuned YOLOv6 model detects and locates the required objects in the image provided. The network is trained on the MS-COCO dataset, evaluated on real-time images, and achieved over 95% average weighted accuracy.

Cloud computing has transformed modern IT architecture by enabling scalable and flexible computing resources. As the popularity of cloud-based services evolves, so do security concerns, making anomaly detection an important field of research. The purpose of this paper is to provide a comprehensive review of cloud anomaly detection methods, with a particular emphasis on machine learning-based approaches. In order to validate the performance of machine learning over anomaly detection on cloud, ten different machine learning algorithms are applied on three datasets. Accuracy, precision, recall, and F1-score are calculated. Out of these ten techniques AdaboostM1 performs best in dataset 1 with values of 0.965449, 0.945588, 0.950879 and 0.945588 for precision, recall, F1-Score and accuracy respectively. While decision tree, hoeffeding tree and $\mathbf{J 4 8}$ performed best with same values in dataset $\mathbf{2}$ in terms of all measurement metrics with accuracy 0.997929, precision 0.997987, recall 0.997929 and F1-Score 0.997946. In dataset 3 AdaboostM1, NaiveBayes and VotedPerceptron performed best in terms of recall, F1-score, and accuracy with values 0.831081, 0.754413 and 0.831081 respectively, while in terms of precision decision tree, hoeffeding tree and J48 performed best with same value of 0.711735. The research reveals that different techniques are well suited in cloud security on different datasets and creates opportunities for future scalable and adaptable anomaly detection systems.

Video Analytics is widely used by the internet-based platforms to govern the mass consumption of videos. Traditionally, it is carried out from the decoded format of the videos. This requires the analytics server to perform both decoding and analytics computation. This process can be made fast and efficient if performed over the compressed format of the videos as it reduces the decoding stress over the analytics server. The field of video analytics from the binarized formats using modern deep learning techniques is still emerging and needs further exploration. This proposed work is based on the same notion. In this work, two analytics tasks that is, classification and object detection are carried out from the binarized videos. The binarized formats are produced by using an already-designed end-to-end video compression network. The experiments have been carried out over standard datasets. The proposed MobileNetv2-based classification network shows an accuracy of 66% over the YouTube UGC dataset and the YOLOX-S-based detection network shows mAP of 45% over IMAGENet datasets. The proposed work shows competitiveness and improvement in the detection outcomes on compressed data and also provides further motivation for the adoption of deep learning-based video compression in practical analytics domains.

Early identification and management of plant diseases are paramount for sustaining crop health, ensuring optimal yields, and safeguarding food security in agricultural systems. Left untreated, diseases caused by fungi, bacteria, viruses, and pests can significantly diminish agricultural output, posing a threat to global food production. While recent research has explored machine learning-based techniques for early disease detection, many proposed models are resource-intensive, characterized by large model sizes, and millions of trainable parameters. Recognizing resource-constrained devices' needs, recent studies have developed lightweight models, but their shallow structure may hinder accurate disease identification. This study proposes the RTR_Lite_MobileNet model, an enhanced version of the original MobileNetV2 model designed for efficient deployment on resource-constrained devices. Different attention techniques, such as Squeeze-and-Excitation Networks (SENet), Efficient Channel Attention (ECA), and Triplet Attention, are added to reduce the model's computational footprint while boosting its ability to capture complicated disease patterns. Extensive experimentation validates the efficacy of RTR_Lite_MobileNet, consistently outperforming MobileNetV2 with top accuracies across multiple datasets: 99.92 % on Plant Disease, 82.00 % on PlantDoc, 97.11 % on PaddyDoctor, 90.84 % on Coffee, 100 % on Wheat, 96.78 % on Soybean, and 96.67 % on Sugarcane. Deployment on edge devices such as Raspberry Pi 4 and 5 demonstrates its computational efficiency, as evidenced by lower latency and memory consumption. Research results indicate that RTR_Lite_MobileNet is a practical and effective option for real-time plant disease diagnosis, paving the way for additional uses in agricultural monitoring and IoT applications.

Plant leaf diseases (PLDs) can continue to be a significant problem in the agricultural sector, leading to significant losses in production and jeopardizing food security. Early detection is essential, and recent achievements in the domain of deep learning (DL) have made automated high-accuracy solutions possible. The most popular and commonly used of these is the You Only Look Once (YOLO) family of object detection models, which have been proposed to detect plant diseases in real time. This review presents a new and in-depth synthesis of YOLO-based methods, including YOLOv1 to YOLOv10 and the domain-specific variants, including CTB-YOLO (coriander), BED-YOLO (YOLOv10n), and RAG-augmented YOLOv8 (coffee). This work compares to previous surveys in that (i) it presents a structured dataset catalog containing information on size, resolution, disease classes, and limitations (such as imbalance and annotation problems); (ii) it provides comparative benchmarking analysis of performance measures (accuracy, precision, recall, F1-score, mean Average Precision, and frames per second) across versions of YOLO to illustrate trade-offs between speed and accuracy; and (iii) it gives forward-looking discussion on how (ii) open challenges and (iii) future research directions, including lightweight YOLO models to run on mobile. This review presents a summative reference and a new contribution to the progress of the YOLO-based PLD detection approach to sustainable agriculture.