Key features of applying unsupervised classification for monitoring spatiotemporal dynamics of landscape cover under conditions of an ecological disaster
- 1Department of Biodiversity and Ecology, Oles Honchar Dnipro National University, Dnipro, Dnipro Region, Ukraine
- 2Department of Department of Botany, Ecology and Gardening, Bogdan Khmelnitsky Melitopol State Pedagogical University, Zaporizhzhia, Zaporizhzhia Region, Ukraine
Abstract
The destruction of the Kakhovka Hydroelectric Power Plant dam in June 2023 triggered a rapid drawdown of the Kakhovka Reservoir and initiated a long-term reorganisation of the Dnipro River floodplain. Such post-disaster landscapes are characterised by the emergence of novel and transitional surface types, whose spectral properties and trajectories are difficult to anticipate, thereby limiting the use of supervised classification. The spatiotemporal dynamics of land cover (surface types) in the floodplain ecosystems of Khortytsia Island, Ukraine, were assessed using an unsupervised remote-sensing framework. A Sentinel-2 time series was analysed using a multi-index approach. Unsupervised classification was implemented via hierarchical clustering with objective selection of the number of clusters. A sequential ordination strategy was applied: initial clustering was conducted in the space of the first two principal components derived from principal component analysis, followed by residual principal component analysis after removal of the dominant components by regression to emphasise ecologically meaningful secondary variation, such as moisture and substrate signals. Cluster maps were optionally extrapolated to full rasters using linear discriminant analysis in the combined principal component analysis (PCA) and residual-PCA feature space. Clusters were interpreted through an expert crosswalk to the European Nature Information System habitat typology, supported by index profiles, spatial context, and seasonal behaviour. The discriminatory value of indices was also quantified at both the cluster level using standardised effect sizes and the global level using normalised importance weights. The surface-type system became progressively more differentiated throughout the growing season, evolving from a minimal early-season set to a more diverse assemblage of surface and habitat types by late season. This progression reflects rapid post-disaster succession and shoreline reconfiguration. Secondary components gained importance during peak vegetation development, indicating the increasing relevance of edaphic and moisture-related gradients in differentiating cover types when canopy biomass obscured non-vegetated signals. The ranking of informative indices shifted seasonally, from wetness and water sensitivity in late winter to early spring, to greenness and productivity in late spring and summer, and finally to substrate exposure and degradation signals in autumn. Unsupervised, sequential PCA-based clustering offers a robust framework for monitoring post-disaster floodplain transformations, as it accommodates novel classes, preserves subtle moisture and substrate signals, and supports interpretable temporal ‘genealogies’ of land-cover types. The results also indicate that relying solely on vegetation indices is insufficient for accurately mapping rapidly changing floodplain habitats following catastrophic disturbances.
Related
- Spatiotemporal coordination between a sports-tourism proxy subsystem and ecological conditions in Zhejiang Province, China, 2010–2024
- Zero-Cycle: First Continuous, Unsupervised, At-Home Monitoring of Spinal Motion as a Per-Cycle Substrate for Machine Learning
- Unsupervised ensemble methods for non-wear detection from triaxial accelerometer data in children following continuous monitoring
- Simple but effective unsupervised classification for specified domain images: a case study on fungi images
- Unsupervised 3D classification and characterisation of osteocyte and chondrocyte lacunae in the murine osteochondral unit
BibTeX
@article{lisovets2026features,
title = {{Key features of applying unsupervised classification for monitoring spatiotemporal dynamics of landscape cover under conditions of an ecological disaster}},
author = {Olena Lisovets and Hanna Tutova and Olha Kunakh and Olexander Zhukov},
journal = {PeerJ},
year = {2026},
month = oct,
doi = {10.7717/peerj.21869},
url = {https://peerj.com/articles/21869},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Key features of applying unsupervised classification for monitoring spatiotemporal dynamics of landscape cover under conditions of an ecological disaster
AU - Olena Lisovets
AU - Hanna Tutova
AU - Olha Kunakh
AU - Olexander Zhukov
PY - 2026
DA - 2026/10/27
JO - PeerJ
DO - 10.7717/peerj.21869
UR - https://peerj.com/articles/21869
AB - The destruction of the Kakhovka Hydroelectric Power Plant dam in June 2023 triggered a rapid drawdown of the Kakhovka Reservoir and initiated a long-term reorganisation of the Dnipro River floodplain. Such post-disaster landscapes are characterised by the emergence of novel and transitional surface types, whose spectral properties and trajectories are difficult to anticipate, thereby limiting the use of supervised classification. The spatiotemporal dynamics of land cover (surface types) in the floodplain ecosystems of Khortytsia Island, Ukraine, were assessed using an unsupervised remote-sensing framework. A Sentinel-2 time series was analysed using a multi-index approach. Unsupervised classification was implemented via hierarchical clustering with objective selection of the number of clusters. A sequential ordination strategy was applied: initial clustering was conducted in the space of the first two principal components derived from principal component analysis, followed by residual principal component analysis after removal of the dominant components by regression to emphasise ecologically meaningful secondary variation, such as moisture and substrate signals. Cluster maps were optionally extrapolated to full rasters using linear discriminant analysis in the combined principal component analysis (PCA) and residual-PCA feature space. Clusters were interpreted through an expert crosswalk to the European Nature Information System habitat typology, supported by index profiles, spatial context, and seasonal behaviour. The discriminatory value of indices was also quantified at both the cluster level using standardised effect sizes and the global level using normalised importance weights. The surface-type system became progressively more differentiated throughout the growing season, evolving from a minimal early-season set to a more diverse assemblage of surface and habitat types by late season. This progression reflects rapid post-disaster succession and shoreline reconfiguration. Secondary components gained importance during peak vegetation development, indicating the increasing relevance of edaphic and moisture-related gradients in differentiating cover types when canopy biomass obscured non-vegetated signals. The ranking of informative indices shifted seasonally, from wetness and water sensitivity in late winter to early spring, to greenness and productivity in late spring and summer, and finally to substrate exposure and degradation signals in autumn. Unsupervised, sequential PCA-based clustering offers a robust framework for monitoring post-disaster floodplain transformations, as it accommodates novel classes, preserves subtle moisture and substrate signals, and supports interpretable temporal ‘genealogies’ of land-cover types. The results also indicate that relying solely on vegetation indices is insufficient for accurately mapping rapidly changing floodplain habitats following catastrophic disturbances.
SN - 2167-8359
ER -