Comparative performance analysis of simple U-Net, residual attention U-Net, and VGG16-U-Net for inventory inland water bodies

Kód výsledku v IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F60076658%3A12520%2F24%3A43908037" target="_blank" >RIV/60076658:12520/24:43908037 - isvavai.cz</a>

Výsledek na webu

<a href="https://doi.org/10.1016/j.acags.2023.100150" target="_blank" >https://doi.org/10.1016/j.acags.2023.100150</a>

DOI - Digital Object Identifier

<a href="http://dx.doi.org/10.1016/j.acags.2023.100150" target="_blank" >10.1016/j.acags.2023.100150</a>

Jazyk výsledku

angličtina

Název v původním jazyce

Comparative performance analysis of simple U-Net, residual attention U-Net, and VGG16-U-Net for inventory inland water bodies

Popis výsledku v původním jazyce

Inland water bodies play a vital role at all scales in the terrestrial water balance and Earth’s climate variability. Thus, an inventory of inland waters is crucially important for hydrologic and ecological studies and management. Therefore, the main aim of this study was to develop a deep learning-based method for inventorying and mapping inland water bodies using the RGB band of high-resolution satellite imagery automatically and accurately. The Sentinel-2 Harmonized dataset, together with ZABAGED-validated ground truth, was used as the main dataset for the model training step. Three different deep learning algorithms based on U-Net architecture were employed to segment inland waters, including a simple U-Net, Residual Attention U-Net, and VGG16-U-Net. All three algorithms were trained using a combination of Sentinel-2 visible bands (Red [B04; 665nm], Green [B03; 560nm], and Blue [B02; 490 nm]) at a 10-meter spatial resolution. The Residual Attention U-Net achieved the highest computational cost due to the increased number of trainable parameters. The VGG16-U-Net had the shortest run time and the lowest number of trainable parameters, attributed to its architecture compared to the simple and Residual Attention U-Net architectures, respectively. As a result, the VGG16-U-Net provided the best segmentation results with a mean-IoU score of 0.9850, a slight improvement compared to other proposed U-Net-based architectures. Although the accuracy of the model based on VGG16-U-Net does not make a difference from Residual Attention U-Net, the computation costs for training VGG16-U-Net were dramatically lower than Residual Attention U-Net.

Název v anglickém jazyce

Comparative performance analysis of simple U-Net, residual attention U-Net, and VGG16-U-Net for inventory inland water bodies

Popis výsledku anglicky

Inland water bodies play a vital role at all scales in the terrestrial water balance and Earth’s climate variability. Thus, an inventory of inland waters is crucially important for hydrologic and ecological studies and management. Therefore, the main aim of this study was to develop a deep learning-based method for inventorying and mapping inland water bodies using the RGB band of high-resolution satellite imagery automatically and accurately. The Sentinel-2 Harmonized dataset, together with ZABAGED-validated ground truth, was used as the main dataset for the model training step. Three different deep learning algorithms based on U-Net architecture were employed to segment inland waters, including a simple U-Net, Residual Attention U-Net, and VGG16-U-Net. All three algorithms were trained using a combination of Sentinel-2 visible bands (Red [B04; 665nm], Green [B03; 560nm], and Blue [B02; 490 nm]) at a 10-meter spatial resolution. The Residual Attention U-Net achieved the highest computational cost due to the increased number of trainable parameters. The VGG16-U-Net had the shortest run time and the lowest number of trainable parameters, attributed to its architecture compared to the simple and Residual Attention U-Net architectures, respectively. As a result, the VGG16-U-Net provided the best segmentation results with a mean-IoU score of 0.9850, a slight improvement compared to other proposed U-Net-based architectures. Although the accuracy of the model based on VGG16-U-Net does not make a difference from Residual Attention U-Net, the computation costs for training VGG16-U-Net were dramatically lower than Residual Attention U-Net.

Druh

J_ost - Ostatní články v recenzovaných periodicích

CEP obor

—

OECD FORD obor

10201 - Computer sciences, information science, bioinformathics (hardware development to be 2.2, social aspect to be 5.8)

Projekt

—

Návaznosti

S - Specificky vyzkum na vysokych skolach

Rok uplatnění

2024

Kód důvěrnosti údajů

S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů

Název periodika

Applied Computing and Geosciences

ISSN

2590-1974

e-ISSN

2590-1974

Svazek periodika

21

Číslo periodika v rámci svazku

neuveden

Stát vydavatele periodika

NL - Nizozemsko

Počet stran výsledku

11

Strana od-do

—

Kód UT WoS článku

001143898700001

EID výsledku v databázi Scopus

2-s2.0-85181001193