Oil spills, which may be generated by various sources and can enter the ocean via various entry sites, are a significant pollutant that significantly influences marine ecosystems and can have far-reaching and disastrous consequences. The health of marine and coastal ecosystems is severely threatened by oil spills, regardless of whether they result from an accident or ships cleaning their tanks. Satellite synthetic Aperture Radar (SAR) devices have a high degree of accuracy in identifying oil spills. These systems function regardless of cloud cover or sunlight and can differentiate oil from a stable sea surface. We present the OSD-DNN framework for monitoring the seas for signs of oil leakage. A convolutional neural network (CNN) with improved SPP Net architecture was employed for training and Testing during the analysis of the proposed method. After that, an improved cross-entropy Adam optimization is used on the model compilation. A strategy that employs Non-adaptive thresholds was applied to image denoising. The Contrast-limited adaptive histogram equalization (CLA.HE) method was used to normalize the histograms. The median threshold canny approach is used for the process of image segmentation. The CNN method was used for the process of feature extraction. The deep convolutional neural network (DCNN) method can see and identify oil spills. According to the experiment's findings, using a deep neural network to detect oil spills improved accuracy, and the Deep CNN performed better than the existing methods.

Data processing plays a significant role in improving the performance of machine learning models. In an IoT network, data generated by sensor nodes will be multivariate and consist of complex patterns and correlations. Such data must be carefully preprocessed before applying ML algorithms for classification. In this paper, we propose a study that improving anomaly classification accuracy on various IoT sensor datasets. In this context, we propose a system that deals with various data distributions and detects anomalies more efficiently. The proposed study uses multiple data transformation methods to prepare the data for further analysis. The data transformation facilitates data conversion, making it more suitable for machine learning models. We demonstrate the efficacy of the proposed method on widely used IoT sensor datasets. We also demonstrate the results of the proposed system with the various machine learning models using performance metrics such as accuracy, AUC Score, F1 score, and Mean Square Error (MSE). It is observed that the proposed system is more effective on various IoT datasets.

Wireless Sensor Networks (WSNs) heavily depend on energy resources it crucial to enhance their longevity to achieve better performance. Clustering serves as an effective technique for controlling WSN topology. However, selecting an appropriate cluster head plays a critical role in the success of this approach. Cluster head selection involves considering various parameters, such as energy levels, node locations, and the number of neighbors, thereby rendering it a complex task. To address this complexity, we propose a cluster head selection method based on fuzzy logic, which emulates human decision-making by simultaneously handling multiple parameters. Our approach utilizes the Mamdani Inference engine and fuzzy parameters, including residual energy, node centrality, and distance from the base station. We evaluate our proposed method using heterogeneous networks featuring two and three levels of hierarchy. Through simulations, we demonstrate that our approach outperforms the SEP in a two-level network and the SEP-E protocol in a three-level network in terms of the network's lifetime.

Due to severe deterioration and writing style differences, keyword spotting from historical handwritten documents remains challenging. This paper uses query-by-example (QBE) and a segmentation-based technique to investigate keyword spotting in historical documents. To match the image of the query to those in a reference database, features extracted from word images by a set of textural features such as Local Directional Number Pattern (LDNP), Complete Local Binary Patterns (CLBP), and Completed Robust Local Binary Pattern (CRLBP) are employed. The process of classifying data involves minimizing a similarity criterion that is derived from the distance between two feature vectors. High performance is achieved by a series of evaluations utilizing various combinations of distance measurements, and these are compared with the approaches used in the ICFHR 2014 word spotting competition.

Emotion recognition has become important for easier and more effective interaction between humans and computers. It could be possible for machines to enhance and improve human communication by better understanding the variety of emotions. Therefore, constructing an emotion-specific brain-computer interface based on EEG might be the first step in overcoming that limitation by producing a neuroscientific medical tool to help such patients regain or maintain a good quality of life. Here, we used the Wavelets Multi-resolution Analysis method to extract various wavelet features from the EEG signal. Different mother wavelet functions like db, sym, coil, and haar are used to decompose EEG signals. Diverse features for five frequency bands are used, such as Power, Relative Power and Power Spectral Density, Wavelet Entropy, and statistical parameters. It has been observed that “Db6,” with six decomposition levels, which yields five separate frequency bands, gives good classification accuracy. It has been observed that electrodes T7, T8, CP1, CP6, F3, F4, FP1, FP2, POZ, and F3 gives good classification accuracy. The performance is tested using various classifiers like SVM, k-NN, DT and RF. It has been observed that time domain Hjorth parameters alone have good classification accuracy compared to other features. Maximum classification accuracy of 73.42%, 71.25%, and 67.84% is achieved for Positive, Negative, and Neutral emotion using the k-NN algorithm using FP1, FP2, F3, F4, T7, T8, CP1, CP6, and POZ channels.

This study presents a comprehensive bibliometric analysis of research in facial emotion recognition. Utilizing the Scopus database, a total of 2492 articles published in English were meticulously analyzed. The study employed various bibliometric indicators, including publication year, source, author, country, and keyword analyses. Tools such as Vos viewer and Excel were used for performance analysis. Our findings highlight a significant growth in publications over the years, with a notable concentration in computer science and engineering. The study identifies key contributors, prevalent research trends, and emerging facial emotion recognition research topics. The insights gained from this study provide a valuable overview of the field's evolution and its current landscape, facilitating future research directions and collaborations.

Recently, the economy has been hit with a recession that has resulted in increasing job losses and heightened job insecurity. This makes predicting job trends crucial for employers and employees alike. Social media is an innovative technology to mine for sentiment analysis to provide for nuanced and data-driven insights into the current employment status. The proposed work uses VADER, a lexicon and rule-based sentiment analysis tool for calculating sentiment scores for all tweets individually, and classifying them as positive, negative or neutral on the basis of these scores. The primary focus of the work is to 1) Calculate sentiments for tweets and analyse the impact of changes in trajectories in job trends and status of layoffs across various countries and 2) Perform geospatial analysis and machine learning algorithms to represent the current status of layoffs and compare accuracies of various models to highlight the most efficacious one respectively. Three algorithms were used in the study, namely Random Forest, Logistic Regression, and K-Nearest Neighbours, out of which Logistic Regression yielded the highest accuracy of 91.88%.