Performance Improvement of Water Body Segmentation by DeeplabV3+Using Two Dimensional Variational Mode Decomposition

Monitoring changes in surface water bodies and other earth surface features uses remote sensing data. The DeeplabV3+ network is an encoder decoder based deep neural network that is widely used to segment images with good precision. To improve the predictions made by the DeeplabV3+ model, a novel technique based on the two-dimensional variational mode decomposition (2D-VMD) is proposed in the present work. Sentinel 2A/B images dataset from Kaggle is used for this study. The images and their corresponding annotations are also available. The masks were obtained using the index known as Normalized Water Difference Index (NDWI). From 2841 images, we cropped 100 x100 subsets resulting in 1,50,204 images. The proposed method is found to be effective in improving the predictions made by DeeplabV3+ model. With respect to the images considered for the study, the average F1 score increased from 0.33 to 0.47. The average Jaccard score increased from 0.21 to 0.34.