An assessment of social media oil spill reports using transfer learning on remotely sensed images

Abstract

Nowadays, we are increasingly aware that irresponsible human behavior is the main reason for many instances of environmental pollution, including oil spills in the sea. In order to detect such contaminants in a timely manner and take care of them as quickly as possible, in this paper we present a method for automatic oil spill detection through remote sensing over Sentinel-1 SAR images. Presented approach uses social media human reports of oil spill as input information for building a data set and training a model over collected images of actual oil spills. Because of a rather small oil spill data set, we are using two methods for obtaining better results. Namely, we use data set augmentation for enrichment of images in the data set and transfer learning to retrain our collection of images based on trained deep network on ImageNet data set. For classifying images we used four machine learning classification models with different accuracy in oil spill detection. A comparison of accuracies of the individual machine learning model and the whole process of oil spill detection initiated by social media reports and further validated by constructed classifier will be described in this paper,