An Ensemble Based Approach for Green Area Segmentation from Satellite Images

Recent statistics show an alarming depreciation in green space in all kinds of regions. Image segmentation using deep learning models is a promising and quick method to determine the green space of an area and aid environment specialists in taking necessary actions. Hence, the objective of this research is to compare state-of-the-art deep learning models and then propose the best approach for segmenting green space from aerial images. Traditional Deep learning and Ensemble Learning techniques were implemented with a publicly available and a self-made data set containing 5108 and 1737 images respectively. All the models were evaluated using suitable performance metrics and the Ensemble Learning approach was found to perform better. The best-performed segmentation model is obtained by incorporating Segnet, UNet and DeepLabV3+ with accuracy, precision, and dice-coefficient of 0.94, 0.96 and 0.91.