Landfill Identification in Satellite Imagery Using Deep Learning

In the modern world, there is an acute problem of environmental pollution, which is associated with human activities. Our country is no exception. One of the most important garbage problems in ecology today is unauthorized landfills. According to official figures, there are a total of about 17,000 unauthorized landfills in the country, as well as about 13,000 unauthorized waste disposal sites with a total area of ​​4 million hectares. When litter is released into the environment, there are many potential risks to the land and animals nearby. Some of the potential problems are: soil and water pollution, wildlife disturbance, increased risk of natural disasters. Unauthorized landfills, unlike authorized ones, are not equipped with special systems to protect the soil near and under the landfill, so it is very important to monitor the emergence of new illegal landfills. Therefore, the goal of this work is to implement a solution for determining household waste landfills from satellite images using deep learning. To achieve this goal, the task of forming a dataset was solved, the task of classifying satellite images into two classes "polygon" and "non-polygon", and the results were analyzed. For the task of classifying satellite images, a transfer learning approach was chosen, several well-known architectures of convolutional neural networks were used with different classifiers and various options for additional training. The best result was achieved in the model with a convolutional base Inception with a fully connected classifier and fine tuning.