Vision-language models (VLMs) can segment object categories absent from their training data, offering a route to shoreline extraction that requires no site-specific annotation. Their accuracy in localizing the land–water boundary, and its dependence on prompt formulation, have not been established. This study evaluates four models representing two distinct pipeline families on 3104 aerial images from the Coastal Aerial Imagery Dataset: direct mask prediction models (CLIPSeg, Florence-2) and VLM-coordinate-guided SAM pipelines (GPT-4o, Qwen-VL-Max). Three prompt conditions were compared under two region-level and two boundary-level metrics, stratified by shoreline complexity, shoreline density, coastal landscape, and wave state. Expanding the prompt with detailed rules for discriminating water from land reduced accuracy relative to a minimal instruction in fifteen of sixteen model–metric combinations, with Florence-2’s mean boundary error rising from 107.08 to 120.98 pixels. A flexible anchoring strategy, adding a second positive point and making the negative point optional, raised region accuracy for all four models but improved boundary accuracy only for the segmentation-native pair. Region-level and boundary-level metrics were weakly associated (IoU–AED Spearman correlation between $-0.20$ and $-0.47$), so high overlap scores did not imply accurate boundary placement. Complete failures occurred in 11.2% to 27.4% of images, concentrated on sheltered shoreline settings with high shoreline density rather than wave-disturbed open coast. Although the median boundary error (21.86 pixels for CLIPSeg) suggests that successful zero-shot predictions approach the 2.71–29.55 pixel mean error range of supervised networks, overall reliability remains compromised by long-tail failures, with region accuracy trailing supervised baselines by 3 to 18 IoU points. Zero-shot extraction thus remains less accurate than supervised segmentation, but offers a label-free alternative whose reliability may be improved through architecture-aware prompt design and environment-aware quality control.

Individual-tree-level information on crown distribution and morphology underpins forest inventory, biomass estimation, and carbon accounting. High-resolution Unmanned Aerial Vehicle (UAV) imagery resolves single-tree detail, but automated crown extraction remains difficult: fully supervised segmentation depends on costly pixel-level annotation that generalizes poorly, while the Segment Anything Model (SAM), though training-free, cannot locate trees on its own and existing SAM-based methods restore this ability only by adding task-specific training. We present an end-to-end, annotation-free toolkit for tree crown extraction and ecological analysis. A RetinaNet-based DeepForest detector produces coarse boxes; an adaptive module then removes duplicate boxes and non-vegetation false positives using an intersection-over-union rule and a global greenness index, converting noisy boxes into clean prompts; these prompts drive SAM to decode irregular crown masks without task-specific training; and geometric and texture features are extracted and grouped by principal component analysis and K-means clustering to map ecological patterns. We evaluated the toolkit on multi-biome imagery from the public OAM-TCD dataset. Because pixel-exact metrics are unstable at 10 cm resolution, where wind sway, shadow shift, and small labeling offsets are strongly amplified, we assessed accuracy under an absolute physical-distance tolerance. At a 2.0 m tolerance, consistent with the effective radius of a mature crown, the toolkit reached a precision of 91.25%, a recall of 86.40%, and an F1-score of 88.76%; bootstrap and Monte Carlo resampling confirmed these values are stable. Without manual annotation, it characterized more than 4700 individual crowns and recovered distinct vegetation patterns across geographic settings, offering a highly adaptable, low-cost baseline tool that demonstrates robust performance across the diverse multi-biome scenes within the OAM-TCD dataset.

High-density cities are complex, tightly coupled systems in which ecological infrastructure, disaster-resilience capacity, and transportation networks interact and constrain one another under intensifying climate-driven hazards and urbanization pressure. Urban green spaces sit at the intersection of these subsystems, offering dual ecological and emergency-shelter value, yet their siting is rarely optimized as a system-level problem: coverage, structural safety, and evacuation accessibility are typically treated in isolation, and classical heuristic solvers struggle to reconcile these interdependent, competing objectives at city scale. This study reframes green-space shelter siting as a socio-technical urban system problem and develops a multi-criteria geospatial decision-support framework that couples a Newton–Raphson-based Optimizer (NRBO) with a P-Median (minimum-weighted-distance) formulation, integrating nine quantitative indicators across the three interacting subsystems above into a single decision-theoretic model. Benchmarked against exhaustive enumeration, the framework reproduces identical system-optimal solutions at a computational cost more than two orders of magnitude lower, demonstrating that intelligent optimization can scale system-level decision-making to city-wide planning problems. By applying this framework to the historic core of Nanjing, this study demonstrates how coordinated, cross-subsystem optimization—rather than single-objective design—can resolve conflicting spatial constraints and yield highly resilient urban outcomes. The empirical findings validate the framework’s capacity to significantly enhance coverage, safety, and accessibility simultaneously. The proposed NRBO–P-Median framework is data-driven, computationally scalable, and structurally transferable to other dense urban systems, offering planners and policymakers a generalizable decision-support tool for embedding disaster-resilience functions into intelligent urban green-infrastructure governance.