The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Download
Publications Copernicus
Download
Citation
Articles | Volume XLVI-M-2-2022
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLVI-M-2-2022, 97–101, 2022
https://doi.org/10.5194/isprs-archives-XLVI-M-2-2022-97-2022
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLVI-M-2-2022, 97–101, 2022
https://doi.org/10.5194/isprs-archives-XLVI-M-2-2022-97-2022
 
25 Jul 2022
25 Jul 2022

EXTRACTING WATER BODIES IN RGB IMAGES USING DEEPLABV3+ ALGORITHM

A. Harika1, R. Sivanpillai2, V. V. Sajith Variyar1, and V. Sowmya1 A. Harika et al.
  • 1Center for Computational Engineering and Networking (CEN), Amrita School of Engineering, Coimbatore, Amrita Vishwa Vidyapeetham, India
  • 2Wyoming GIS Center, University of Wyoming, Laramie, WY 82072, USA

Keywords: Sentinel 2A/B, Multi-scale Features, Encoder-Decoder, Atrous Spatial Pyramid Pooling, ASPP, Water quality, NIR

Abstract. Deep Learning algorithms are increasingly used for mapping waterbodies in remotely sensed images. DeepLabV3+ is an image segmentation method that includes ASPP and encoder-decoder to retrieve pyramid spatial features at different scales and structural information respectively. Previous studies have shown that DeepLabV3+ can accurately map waterbodies in false colour infrared images. However, ability of DeepLabV3+ for extracting waterbodies in RGB images is unknown. This study tested DeepLabv3+ algorithm to extract waterbodies in the RGB bands. Sentinel 2A/B images (n = 2841) and their corresponding annotations were downloaded from Kaggle (host of public datasets) and subset images (n = 10405) of 100 × 100 pixels were cropped. From these subset images, 8941 were used for training and validation and 1464 were used for testing the trained model. Dice and Jaccard/Intersection over Union (IoU) were used for evaluating the output generated by the model. The network was trained for 50 epochs with 32 iterations in each epoch. The model trained at the end of 30th epoch was selected as final based on minimum information loss (0.0743). The average Dice and Jaccard/IoU scores for the output images were 0.8412 and 0.7169 respectively. The high scores obtained in this study indicate that DeepLabV3+ can be used for identifying waterbodies in RGB or true-colour images.