Human activities have fundamentally altered the functioning of most terrestrial ecosystems, particularly grazing ecosystems where vegetation-dependent processes underpin both human well-being and ecological resilience. Despite the importance of ground cover for sustaining these processes, the long-term dynamics of vegetation cover—and the degree to which temporal variability is explained by climatic drivers—remain poorly quantified at farm and paddock scales. Existing approaches typically focus on photosynthetic vegetation or rainfall alone, with limited integration of temperature effects, non-photosynthetic components, or consistent benchmarking across environmental and land-use contexts. Here, we applied a remote-sensing framework to quantify long-term relationships between total vegetation cover (i.e., the sum of photosynthetic and non-photosynthetic components) and climate, expressed as interactions between accumulated rainfall and temperature using Landsat-derived fractional cover and SILO gridded climate data. Using 40 years of Landsat-derived fractional cover data, we characterised vegetation cover dynamics using three complementary metrics: (1) long-term mean vegetation cover, (2) the coefficient of determination (R2), defined as the proportion of temporal vegetation-cover variability explained by the selected climate predictors, and (3) temporal trends in climate-adjusted residuals indicating vegetation-cover change not explained by the climate model. We further introduced a bioregional benchmarking approach that contextualises paddock-scale patterns against comparable land uses within shared environmental settings. Across more than 140,000 km2 of Australian grazing landscapes, interactions between monthly temperature and accumulated rainfall (1–18 months) frequently explained a large proportion of vegetation cover variability (R2 > 0.5). Significant residual trends were detected across more than one-quarter of the landscape, with decreasing climate-adjusted vegetation-cover signals more prevalent than increasing signals. These residual trends identify areas requiring contextual interpretation rather than direct attribution to degradation. All three metrics varied markedly among bioregions and land use classes, highlighting the importance of region-specific benchmarking for interpretation. Our results demonstrate the value of integrating long-term Earth observation data with climate modelling to detect ecosystem change signals and support evidence-based decision-making in grazing ecosystems.
Mapping potential vernal pools (PVPs) through remote sensing is important for guiding the management of these key habitats. Photointerpretation of PVPs from springtime aerial imagery is time consuming and prone to errors of omission (e.g. from obscuration by coniferous trees) and commission (e.g. from tree shadows). We tested a hybrid mapping approach that uses LiDAR- and image-based geospatial indicators to guide photointerpretation of PVPs. We created 5 indicators, based on established and novel approaches, to detect characteristics of PVPs. These characteristics included depressions in the elevation model, voids in ground-level LiDAR returns, low intensity of ground-level LiDAR returns, and very low brightness in 4-band (RGB, NIR) multispectral imagery. The hybrid approach had omission and commission errors of 14% and 19%, respectively, for aquatic depressions. A photointerpretation-only approach (for a different study area) had omission and commission errors of 10% and 32%, respectively. Subjectively, we found the hybrid approach to be more efficient compared to the photointerpretation-only approach. PVPs associated with more than 2 indicators had very low commission errors (<4%) which suggests that co-occurring indicators can be used identify PVPs that are more likely to be actual vernal pools. The hybrid approach combined the ability of LiDAR and imagery to detect vernal pool characteristics with the ability of a human photo-interpreter to reduce commission errors.
Remote sensing and deep learning enable efficient post-event landslide mapping to support rapid disaster assessment and subsequent hazard analysis, yet most studies rely on binary landslide labels that obscure type-related geomorphological differences. To address this limitation, this study constructed two multi-type landslide inventories and supervised segmentation datasets from PlanetScope and SPOT-6 data acquired over the Noto Peninsula, Japan, after the 2024 earthquake and subsequent rainfall event. Using expert stereoscopic interpretation aided by publicly available government landslide information, we mapped 1679 post-earthquake and 2505 post-rainfall landslides as falls, slides, and flows. Both inventories contained substantial class samples and broad object-size ranges, providing a basis for supervised learning. We evaluated 13 representative semantic segmentation architectures to assess inventory learnability, segmentation quality, foreground probability, predictive entropy, computational complexity, and practical model selection. Across three training runs, most models successfully learned the multi-type landslide information, supporting the learnability of the constructed inventories. DRANet and SwinUNet showed the strongest overall performance, with foreground mIoU of 0.88 ± 0.02 and 0.86 ± 0.01 under PlanetScope, and 0.86 ± 0.02 and 0.86 ± 0.01 under SPOT-6, respectively. Compared with DRANet, SwinUNet achieved similarly strong segmentation performance at substantially lower computational cost. Exploratory pairwise Wilcoxon comparisons were broadly consistent with these patterns, while additional analyses showed lower performance for smaller patch-level components and greater confusion between falls and flows. This study provides an exploratory framework for remote sensing- and deep learning-based multi-type landslide mapping, though its generalization remains to be evaluated across independent regions and sensor conditions.
