Efficient 3D convolutional neural networks for Sentinel-2 land cover classification with limited ground truth data

Kód výsledku v IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F00216208%3A11310%2F25%3A10504234" target="_blank" >RIV/00216208:11310/25:10504234 - isvavai.cz</a>

Výsledek na webu

<a href="https://verso.is.cuni.cz/pub/verso.fpl?fname=obd_publikace_handle&handle=gDLXnXS3_Q" target="_blank" >https://verso.is.cuni.cz/pub/verso.fpl?fname=obd_publikace_handle&handle=gDLXnXS3_Q</a>

DOI - Digital Object Identifier

<a href="http://dx.doi.org/10.1080/22797254.2025.2512531" target="_blank" >10.1080/22797254.2025.2512531</a>

Jazyk výsledku

angličtina

Název v původním jazyce

Efficient 3D convolutional neural networks for Sentinel-2 land cover classification with limited ground truth data

Popis výsledku v původním jazyce

This paper focuses on an innovative application of deep learning (DL) techniques, particularly 3D convolutional neural networks (CNNs), for land cover classification using multispectral Sentinel-2 (S-2) data. In this study, we evaluated the performance of window-pixel-wise 2D, 3D, and 3D Multiscale CNN architectures for land cover classification. 3D and 3D multiscale CNNs were using the spectral dimension as the third dimension for convolutions. Methodology was applied to classify large area (23,217 km2) in Czechia according to the Land use, land-use change and forestry (LULUCF) categories, a key sector in greenhouse gas inventories. The input dataset included S-2 data, along with NDVI, NDVI variance, and SRTM elevation data, all resampled to the 10 m S-2 grid and forming multi-dimensional input 5 x 5 pixel patches. The results show that a 3D CNN with 3 x 3 x 3 spatial-spectral filters and classical training achieved the best F1 score of 0.84, outperforming other proposed CNN architectures and a baseline Random Forest classifier. The study highlights the ability of 3D CNNs to integrate spatial-spectral information, making them highly effective for multispectral data analysis, even with limited (small) training ground truth datasets. This approach provides valuable information for researchers seeking to optimize DL methods for land cover classification, particularly for applications aligned with the LULUCF frameworks.

Název v anglickém jazyce

Efficient 3D convolutional neural networks for Sentinel-2 land cover classification with limited ground truth data

Popis výsledku anglicky

This paper focuses on an innovative application of deep learning (DL) techniques, particularly 3D convolutional neural networks (CNNs), for land cover classification using multispectral Sentinel-2 (S-2) data. In this study, we evaluated the performance of window-pixel-wise 2D, 3D, and 3D Multiscale CNN architectures for land cover classification. 3D and 3D multiscale CNNs were using the spectral dimension as the third dimension for convolutions. Methodology was applied to classify large area (23,217 km2) in Czechia according to the Land use, land-use change and forestry (LULUCF) categories, a key sector in greenhouse gas inventories. The input dataset included S-2 data, along with NDVI, NDVI variance, and SRTM elevation data, all resampled to the 10 m S-2 grid and forming multi-dimensional input 5 x 5 pixel patches. The results show that a 3D CNN with 3 x 3 x 3 spatial-spectral filters and classical training achieved the best F1 score of 0.84, outperforming other proposed CNN architectures and a baseline Random Forest classifier. The study highlights the ability of 3D CNNs to integrate spatial-spectral information, making them highly effective for multispectral data analysis, even with limited (small) training ground truth datasets. This approach provides valuable information for researchers seeking to optimize DL methods for land cover classification, particularly for applications aligned with the LULUCF frameworks.

Druh

J_imp - Článek v periodiku v databázi Web of Science

CEP obor

—

OECD FORD obor

10508 - Physical geography

Projekt

—

Návaznosti

S - Specificky vyzkum na vysokych skolach<br>I - Institucionalni podpora na dlouhodoby koncepcni rozvoj vyzkumne organizace

Rok uplatnění

2025

Kód důvěrnosti údajů

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

Název periodika

Europen Journal of Remote Sensing

ISSN

2279-7254

e-ISSN

2279-7254

Svazek periodika

58

Číslo periodika v rámci svazku

1

Stát vydavatele periodika

IT - Italská republika

Počet stran výsledku

20

Strana od-do

2512531

Kód UT WoS článku

001502969900001

EID výsledku v databázi Scopus

2-s2.0-105007604034