The increasing availability of pretrained Geospatial Artificial Intelligence (GeoAI) models has accelerated the operational adoption of deep learning within geographic information systems. However, models developed in one geographic region frequently experience performance degradation when deployed elsewhere because of geographic domain shift. Although transfer learning is widely used to improve cross-regional model performance, it remains unclear whether improved predictive accuracy alone provides sufficient evidence for trustworthy deployment. This study investigates this question using rooftop solar photovoltaic mapping in Brisbane, Australia, as an operational case study. Esri’s pretrained US Solar Panel Detection model was locally calibrated through transfer learning and evaluated using both development and independent local test datasets within Brisbane. Local calibration substantially improved detection performance, increasing recall from 38.5
Urban compactness and functional mixing are widely recognized as important carbon mitigation strategies. However, determining appropriate levels of compactness and functional mixing, and their potential carbon reductions, remains challenging due to nonlinear morphology–emission relationships and strong spatial dependence. This study develops a spatially explicit Bayesian optimization (SEBO) framework integrating spatial relationship analysis, nonlinear response-interval identification, and constrained configuration optimization. Using 4,347 grid cells in the Nanjing metropolitan region, spatial econometric models identify sector-specific morphology–emission associations and green-blue moderation effects. Tree-based machine learning then identifies responsive ranges for morphological adjustment, which constrain Bayesian optimization to explore lower-emission configurations and assess mitigation potential. Results reveal substantial sectoral differences in morphology–emission relationships and response intervals. Vertical intensity and functional mix are associated with transportation, residential, and power-related emissions, while horizontal density primarily relates to residential emissions. Optimizing vertical intensity yields average simulated reductions of 14.09
This study investigates the evolution of cognitive load in map navigation scenarios, a topic that carries key theoretical and practical value for optimizing map interface layouts and improving navigation visualization. This study adopts a map target search task as the experimental context and employs a 3 × 4 within-subjects design. Task performance, questionnaire scores, and multimodal physiological data (eye-tracking, photoplethysmography, and electrodermal activity) are simultaneously collected. This study applies two-factor repeated-measures ANOVA and linear mixed-effects model to systematically examine the effects of spatial complexity and target location distribution on extraneous map cognitive load. The results demonstrate that high spatial complexity significantly raises extraneous map cognitive load, while target locations in the upper-right quadrant produce the highest extraneous map cognitive load. All eye-tracking indicators, mean heart rate, interbeat interval, low-frequency power, and numbers of skin conductance responses exhibit statistical sensitivity to extraneous map cognitive load manipulations in the map target search task. Fixation duration, low‑frequency power, and numbers of skin conductance responses can genuinely reflect variations in extraneous map cognitive load. This study enriches the theoretical framework of map cognitive load, extends the application of multimodal physiological measurements within cartography, and provides empirical evidence that supports optimized map interface design and cognitive assistance strategies for navigation tasks.
Map generalization aims to adapt detailed geospatial data to visualization at different map scales while preserving legibility and geometric consistency. Recent learning-based approaches have shown promising results, but they typically require training separate models for each target scale and produce raster outputs that lack geometric regularity. This paper proposes an instruction-guided diffusion framework for raster-based building map generalization that enables scale-adaptive generalization within a single model. By conditioning a latent diffusion model with scale-indicative textual instructions, the framework allows prompt-driven control of generalization behavior across multiple and previously unseen target scales, while implicitly capturing the effects of multiple generalization operators, such as simplification, aggregation, and selection. The raster-based generalization results are subsequently converted to vector maps using frame-field-based vectorization, producing building footprints with preserved orthogonality and parallelism. Experiments on multi-scale building map generalization demonstrate that the proposed method yields vectorized results with substantially improved geometric regularity over existing baselines. Quantitative evaluations at both pixel and polygon levels confirm enhanced boundary regularity, rectangularity, and parallelism. Cross-scale experiments further indicate smooth scale-dependent behavior across the evaluated seen and unseen intermediate scales.
Recent advancements in novel view synthesis, particularly Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS), have driven substantial progress in 3D reconstruction from street-view visual streams. By optimizing continuous or discrete scene representations through differentiable rendering, these computational paradigms have rendered complex urban environment modeling increasingly tractable. Yet the massive volume of crowdsourced street view videos (CSVVs) on social media platforms remains largely intractable for these methods due to their non-standardized nature, which often causes reconstructions to fail after substantial computational resources are wasted. To address this challenge, we introduce Crowdsourced Radiance Fields (CSRF), a framework designed to harness CSVVs. CSRF enhances reconstruction viability and reduces the processing time wasted on failed initializations through a dual-stage architecture: (1) a lightweight quality-assessment module that rapidly filters severely degraded content, and (2) a progressive joint-optimization strategy that reconstructs scenes using locally overlapping TensoRF-based radiance fields while simultaneously estimating camera poses, guided by monocular depth and optical flow priors. To systematically evaluate the capabilities of 3D reconstruction from CSVVs, we present CS3D100, a challenging benchmark comprising 100 diverse user-uploaded CSVVs (drone, dashcam, mobile, and vlog footage) from social media platforms. Extensive experimental evaluations demonstrate that the proposed CSRF framework yields the most competitive quantitative results on the CS3D100 dataset within the scope of the evaluated zero-prior baselines. This performance advantage is validated across a comprehensive suite of criteria, encompassing sequence processing success rate, perceptual rendering quality, computational resource efficiency, and the newly introduced composite metrics, CUR and REI. Furthermore, the framework maintains comparable reconstruction capabilities when applied to standard reference benchmarks. Our framework achieves an absolute reconstruction success rate of 31
