Landfill Detection In Satellite Images Using Deep Learning

Open landfills are known to cause a great deal of detrimental effects on the environment and public health by producing harmful gasses and spreading waterborne diseases.Therefore,the efficient environmental management is of utmost importance.Remote sensing data might be a great asset to alleviate the waste management burdens,by being able to accurately detect and keep track of the landfills.However,object detection in remote sensing data,specifically optical satellite images,is a very ambitious and challenging task due to the factors like cluttered environments,complex backgrounds and low resolution.Moreover,if the objects to be detected are small,then it makes it significantly harder to accurately localize them.With the advances made in a computer vision domain since the deep learning was introduced,object detection tasks have witnessed remarkable improvements.Despite reaching human level accuracy in object detection for optical images with deep learning algorithms,object detection in satellite images is still in strong need of enhancements.This work proposes a modified version of state-of-art deep learning architecture called RetinaNet.RetinaNet uses a novel loss function that solves the class imbalance problem.To be more precise,it down-weights the background class so that the model is more focused on actual regions of interest.This paper proposes to use the DenseNet architectures that are good at learning diversified object features,as the feature extraction network,which will act as a backbone to RetinaNet architecture.The average precision achieved for landfill detection is 84.7 %,which outperforms the baseline model and other state-of-art deep learning methodologies.