TY - GEN

T1 - Land cover mapping with Sentinel-2 imagery using deep learning semantic segmentation models

AU - Honcharov, Oleksandr

AU - Hnatushenko, Viktoriia

N1 - Publisher Copyright: © 2024 Copyright for this paper by its authors.

PY - 2025/2/1

Y1 - 2025/2/1

N2 - Land cover mapping is essential for environmental monitoring and evaluating the effects of human activities. Recent studies have demonstrated the effective application of particular deep learning models for tasks such as wetland mapping. Nonetheless, it is still ambiguous which advanced models developed for natural images are most appropriate for remote sensing data. This study focuses on the segmentation of agricultural fields using satellite imagery to distinguish between cultivated and non-cultivated areas. We employed Sentinel-2 imagery obtained during the summer of 2023 in Ukraine, illustrating the nation's varied land cover. The models were trained to differentiate among three principal categories: water, fields, and background. We chose and optimised five advanced semantic segmentation models, each embodying distinct methodological methods derived from U-Net. Upon examination, all models exhibited robust performance, with total accuracy spanning from 80% to 89.2%. The highest-performing models were U-Net with Residual Blocks and U-Net with Residual Blocks and Batch Normalisation, whereas U-Net with LeakyReLU Activation exhibited much quicker inference times. The findings suggest that semantic segmentation algorithms are highly effective for efficient land cover mapping utilising multispectral satellite images and establish a dependable benchmark for assessing future advancements in this domain.

AB - Land cover mapping is essential for environmental monitoring and evaluating the effects of human activities. Recent studies have demonstrated the effective application of particular deep learning models for tasks such as wetland mapping. Nonetheless, it is still ambiguous which advanced models developed for natural images are most appropriate for remote sensing data. This study focuses on the segmentation of agricultural fields using satellite imagery to distinguish between cultivated and non-cultivated areas. We employed Sentinel-2 imagery obtained during the summer of 2023 in Ukraine, illustrating the nation's varied land cover. The models were trained to differentiate among three principal categories: water, fields, and background. We chose and optimised five advanced semantic segmentation models, each embodying distinct methodological methods derived from U-Net. Upon examination, all models exhibited robust performance, with total accuracy spanning from 80% to 89.2%. The highest-performing models were U-Net with Residual Blocks and U-Net with Residual Blocks and Batch Normalisation, whereas U-Net with LeakyReLU Activation exhibited much quicker inference times. The findings suggest that semantic segmentation algorithms are highly effective for efficient land cover mapping utilising multispectral satellite images and establish a dependable benchmark for assessing future advancements in this domain.

KW - agricultural lands

KW - deep learning

KW - satellite images

KW - Semantic segmentation

KW - U-Net architecture

UR - http://www.scopus.com/inward/record.url?scp=85217281561&partnerID=8YFLogxK

M3 - Conference contribution

AN - SCOPUS:85217281561

T3 - CEUR workshop proceedings

SP - 1

EP - 18

BT - Information Technology and Implementation 2024

T2 - 11th International Scientific Conference "Information Technology and Implementation", IT and I 2024

Y2 - 20 November 2024 through 21 November 2024

ER -