A Cascade Deep Neural Network Approach for Oil Spills Classification and Detection

Oil spills represent a growing environmental challenge that poses a significant threat to living organisms. Moreover, the treatment of oil spills, especially in severe cases, has serious economic repercussions and requires substantial labor and time. Therefore, the effective detection of oil spills has become an important research problem. Traditional methods for detecting oil spills, such as manual patrolling and dynamic sensors, are often limited in accuracy and coverage. As a result, the automation of oil spills detection has emerged as a critical global imperative in scientific research. The aim of this paper is to employ deep learning technology to achieve effective detection of oil spills based on aerial images. Our approach is composed of two phases. In the first phase, a Deep Convolutional Neural Network (DCNN), namely ResNet50, is trained on a large dataset containing images showing oil spills at a seaport. The trained DCNN is used to classify the input image as "Oil Spill" or "No Oil Spill". In the second phase, the images classified as "Oil Spill" are analyzed using a deep learning detection model, namely You-Only-Look-Once (YOLOv4), to localize the oil spills. The results indicate the capability of the proposed method to achieve effective oil spill detection. In particular, the classification accuracy obtained by the ResNet50 model is equal to 98%. Moreover, the YOLOv4 model was able to obtain effective localization of the oil spills with mean-average precision of 62%.