An evaluation of the effectiveness of spectral bands and indices on semantic segmentation with Attention U-Net using Sentinel-2A imagery and ESA WorldCover products

Abstract

In remote sensing applications, semantic segmentation is applied to assign a semantic label
(e.g., settlement, forest, and meadow) to the pixels of an image. Due to the diversity of natural
and unnatural landscapes, semantic segmentation remains a difficult task for remotely sensed
imagery. In this study, the Attention U-Net model was trained as a deep learning network for
semantic segmentation of a Sentinel-2A image together with the WorldCover reference
dataset to generate the land cover map of the study site covering Bolu, Duzce and Zonguldak
provinces of Turkey. The Attention U-Net was applied to two datasets (original Sentinel-2A
dataset with 10 spectral bands and Sentinel-2A dataset with additional spectral indices
including Normalized Difference Vegetation Index-NDVI, Soil Adjusted Vegetation Index-SAVI,
Normalized Difference Build-Up Index-NDBI and Normalized Difference Water Index-NDWI)
and performance comparison was performed. The performances of deep learning models
were investigated using Intersection Over Union (IoU) for segmentation model training and
mean IoU for model results. The results show that the use of spectral indices as auxiliary data
in semantic segmentation increases the mean IoU by approximately 10%. The results indicate
that spectral indices have a higher degree of effectiveness in accuracy assessment than the
original spectral bands.