Auditing visible and structural vegetation signals for urban forest monitoring using street-view imagery and airborne LiDAR

1 Introduction

Urban forests comprise spatially dispersed vegetation resources that must be monitored across public streets, courtyards, institutional grounds, parks, and other ownership contexts. Remote sensing has expanded the scale at which these resources can be mapped, but sensing systems observe different parts of them. Reviews of remote sensing in urban forestry show that aerial, point-cloud, and ground-level observations differ in field of view, structural detail, spatial support, and susceptibility to occlusion (Li et al., 2019; Neyns and Canters, 2022; Velasquez-Camacho et al., 2021). These differences matter when mapped indicators are used to guide tree inventories, field inspection, maintenance assessment, or the identification of locations requiring further management evidence.

Street-level imagery (SVI) provides a pedestrian-facing view that overhead data cannot directly reproduce (Biljecki and Ito, 2021). The Green View Index and subsequent semantic-segmentation methods have enabled repeatable mapping of eye-level vegetation (Li et al., 2015; Chen et al., 2020; Seiferling et al., 2017). Panoramic and multiview indices further reduce directional omission (Wang et al., 2022; Khan et al., 2022). Recent developments extend SVI toward individual-tree detection, geolocation, and fine-grained species classification by combining images with GIS constraints and large-scale benchmark data (Li et al., 2026b,a). For urban forest monitoring, however, a visible-green proportion remains an image-plane condition. It does not by itself establish vegetation height, woody structure, permanence, health, ownership, accessibility, or ecosystem function.

Three-dimensional sensing provides complementary structural information. Mobile and airborne LiDAR can characterize vegetation height and vertical arrangement that image proportions suppress (Tang et al., 2023; Chen et al., 2015), while tree-level point-cloud workflows increasingly support individual-tree segmentation and attribute estimation (Wang and Liang, 2026). Point-cloud semantic segmentation is reviewed by Xie et al. (2020). Yet airborne LiDAR is not a complete urban-tree inventory or a ground truth for street-level experience. Classified returns may combine trees, shrubs, and other vegetation; structural presence does not establish condition, species, management responsibility, or public benefit. Accordingly, the useful management question is not which sensor is universally superior, but where their observations corroborate one another and where disagreement should trigger additional verification.

This need is sharpened by data-quality limitations. SVI coverage can be spatially biased (Fan et al., 2025), and available scenes vary in acquisition date, viewpoint, and temporal continuity (Kim and Jang, 2023). Geographic-identity-preserving generation from satellite views is emerging as one response to street-view coverage disparity (Li et al., 2026c), but observed-image availability and provenance remain central to monitoring design. Systematic review evidence also identifies tool validation, reliability, and multi-source integration as persistent concerns in SVI-based auditing (Dai et al., 2024). Cross-sensor fusion is further affected by temporal lags and mismatched spatial support. Even correctly retrieved and segmented imagery may therefore provide an incomplete representation of nearby vegetation structure. Urban forest monitoring requires both a coverage audit and a construct audit: whether observations exist and whether they support the interpretation assigned to them.

Existing studies demonstrate that street-level and overhead greenness measures are related but non-interchangeable (Helbich et al., 2021; Li et al., 2021). Reviews of greenery assessment similarly caution that indicator meaning depends on viewpoint and spatial support (Labib et al., 2021; Lu et al., 2023; Liu et al., 2025). What remains operationally useful is a transparent framework that preserves these sensor-specific meanings, identifies visible-structural mismatches, and converts them into bounded field-verification questions rather than unqualified management prescriptions.

This study develops such a multimodal audit by aligning street-view semantic indicators and airborne-LiDAR morphology at shared spatial supports. Kazimierz in Krakow is used as a compact historic urban case because narrow streets, courtyards, protected fabric, and public-private edges complicate the relationship between visible and structurally present vegetation (Zachariasz et al., 2023; UNESCO World Heritage Centre, n.d.). The study asks: (1) how strongly do visible and structural vegetation indicators align; (2) where do street-level visibility and structural greenness disagree; (3) how does a second sensing modality change a single-source screen; and (4) how sensitive are the resulting classes to data filters, spatial aggregation, dependence, and classification rules? The contributions are a quality-controlled cross-modal pairing workflow, interpretable verification classes for urban forest monitoring, and explicit uncertainty reporting. The labels are management-screening prompts, not ecological types or intervention priorities, and the workflow has not been validated against independent tree-level or ecological field observations.

2 Materials and methods

2.1 Study area and remote-sensing design

The study area covers the Kazimierz focus area in Krakow, Poland. Candidate street points, retrieved Street View panoramas, and airborne-LiDAR tile coverage are shown in Figure 1. All metric operations were conducted in Poland CS92 (EPSG:2180). The two sensing streams were processed independently and joined only after sensor-specific features had been derived. Figure 2 records acquisition, feature extraction, spatial harmonization, cross-modal classification, and uncertainty assessment. In this study, “urban forest monitoring” refers to screening the visible and structural vegetation resource across the urban landscape. The available classifications do not isolate individual trees or consistently distinguish woody vegetation from herbaceous cover; the workflow is therefore positioned before, rather than as a replacement for, tree-level inventory and field assessment.

2.2 Street-view image acquisition and semantic segmentation

Candidate points were generated at 40 m intervals along eligible OpenStreetMap street geometries. Google Street View metadata were queried at each candidate location, retaining the panorama identifier, acquisition month, panorama coordinates, and query-to-panorama displacement. Four 640 × 640 pixel images at headings of 0, 90, 180, and 270 degrees were retrieved for each successful location. For class c, the heading-level ratio was R_*ich* = N_*ich*/N_*ih*, where N_*ich* is the number of class-c pixels and N_*ih* is the number of pixels in heading h. The sample-point value was the arithmetic mean R̄ic=14∑h=14Rich. Table 1 reports the complete data inventory and analytical hierarchy.

Table 1. Multimodal data inventory and analytical support.

Scenes were segmented with nvidia/segformer-b0-finetuned-cityscapes-1024-1024. The model architecture follows SegFormer (Xie et al., 2021), and its taxonomy follows Cityscapes (Cordts et al., 2016). The 640 × 640 RGB images were processed with the cached model processor: bilinear resizing to 512 pixels, ImageNet mean/standard-deviation normalization, and no explicit crop. Inference used the script's default batch size of four on Apple MPS, with CPU fallback; output logits were bilinearly interpolated to 640 × 640 before the maximum-probability class was assigned. The 512-pixel inference input differs from the checkpoint's 1,024 × 1,024 fine-tuning regime and may disproportionately affect small objects and thin structures such as fine branches and narrow foliage–sky gaps. Because semantic models can also shift across cities and image providers (Tsai et al., 2018), broad semantic classes and explicit image-level checks were retained. This treatment follows cautions against inferring detailed urban-tree attributes directly from generic computer-vision labels (Arevalo-Ramirez et al., 2024).

Vegetation ratio was the share of Cityscapes vegetation pixels and was used for the headline vegetation-to-vegetation correlations. Nature ratio equalled vegetation plus terrain shares and was used as the visible dimension of the screening labels because it retains low green ground cover that is visible but not represented by a vegetation-only class. Facade ratio equalled building plus wall shares. Upper-half vegetation was the vegetation-pixel share within rows 1–320. The two-dimensional enclosure index was E_2*D* = R_*facade*/(R_*facade*+R_*sky*) when the denominator was positive. Sky ratio, road/sidewalk ratio, visual clutter, and the other indicators were calculated from the same masks. Definitions, spatial supports, and roles are listed in Table 2.

Table 2. Core audit indicators, target constructs, and spatial supports.

2.3 Airborne LiDAR processing and morphological metrics

The LiDAR inventory contained 29 LAZ files across source groups 75,174, 78,989, and 83,134, totaling 717,503,084 cached points. Acquisition dates were recovered from adjusted GPS time rather than LAS file-creation dates: 5 October 2021 for group 75,174, 8 September 2023 for group 78,989, and 28 April 2025 for group 83,134. The primary extraction assigned central samples to group 83,134 where a complete 50 m neighborhood was available and used group 75,174 as the fallback, so 17 inventory files contributed to the primary sample-level metrics; group 78,989 remained an inventory and cross-check source rather than a third primary assignment. Coordinates were transformed from WGS 84 to either Poland CS92 (EPSG:2180) or Poland CS2000 zone 7 (EPSG:2178), according to the assigned tile group.

Features were extracted within circular 20 m, 30 m, and 50 m buffers, with 30 m specified as the primary support. The observed LAS classes were 0 (created/never classified), 1 (unclassified), 2 (ground), 3 (low vegetation), 4 (medium vegetation), 5 (high vegetation), 6 (building), 7 (low point/noise), 9 (water), and 12 (overlap). Classes 3–5 were combined as vegetation; class 5 represented high vegetation; class 6 represented buildings; and class 2 provided a local ground-elevation reference. Primary ratios used all returns in the buffer as the denominator. Sensitivity versions removed classes 7 and 12, used only recognized classes 2/3/4/5/6/9, or divided vegetation by ground plus vegetation returns. The high-vegetation ratio was treated as a canopy-oriented structural proxy because it preferentially represents taller vegetation, but it was not interpreted as a validated tree-crown delineation. Heights above the local ground reference were clipped to 0–80 m and summarized in eight bins. The low-open metric was the share of returns below 2 m, the tall-obstruction metric was the share at or above 10 m, and vertical entropy was H=-∑k=18pkln(pk), where p_*k* is the share of returns in height bin k. These metrics operationalize three-dimensional landscape pattern (Wu et al., 2017) and urban spatial structure (Qi et al., 2022).

The dataset contains 2,424 successful Street View sample points. Of these, 2,376 have valid 30 m SVI-LiDAR pairs; 48 contain zero LiDAR points within the primary buffer. The 48 records were retained in coverage accounting and excluded from correlations, audit classes, percentages, and reclassification. No LiDAR values were imputed.

2.4 Spatial pairing, indicator harmonization, and audit labels

Nature ratio and LiDAR vegetation ratio were standardized within the 2,376-record paired sample. Let Vi=z(R̄nature,i) and L_*i* = z(R_{*LiDAR vegetation, i*}). The mutually exclusive green audit label was assigned in the following order:

Ci={legible green structure,Vi≥0.75 and Li≥0.75,latent structural greenness,Li≥0.75 and Vi<−0.25,image-visible green,Vi≥0.75 and Li<−0.25,low-green hardscape,Vi<−0.50 and Li<−0.50,mixed/moderate,otherwise.(1)

Analogous rules classified sky ratio and LiDAR low-open ratio. Single-source screens were separate conditional filters: low visible greenness V_*i* < −0.50, high visible greenness V_*i*≥0.75, low LiDAR vegetation L_*i* < −0.50, low visible sky z_{*sky, i*} < −0.50, and low structural openness z_{*open, i*} < −0.50. Thus an off-diagonal class defined with a counterpart below −0.25 need not be fully contained in a stricter low screen. The labels were specified as transparent screening classes rather than ecological types. Comparisons used a stricter z rule, empirical quartiles, empirical tertiles, a bivariate regression-residual rule, and a five-component Gaussian mixture model (GMM); continuous visible and structural z scores and their difference were also retained.

The paired indicators were interpreted as sensor-specific observations rather than interchangeable measurements of a single ground-truth construct. Street-view semantics describe image-plane visibility from sampled camera positions and headings, whereas airborne LiDAR describes classified three-dimensional returns within a planimetric buffer. LiDAR vegetation was therefore used as a structural counterpart, not as ecological ground truth. The audit labels expose cross-modal agreement and disagreement under explicit spatial supports and threshold rules. For management screening, these labels indicate the next evidence required—for example, tree location, vegetation height, health, ownership, access, or planting feasibility—rather than prescribing an intervention.

2.5 Robustness, spatial validation, and uncertainty assessment

Pearson and Spearman coefficients quantified cross-modal alignment. Robustness checks used panorama-unique records, query-to-panorama displacement filters of 50 and 75 m, image subsets acquired in 2022 or later and 2024 or later, the two primary LiDAR source groups, street-ID means, named-street means, 100 m grid means, and 2,000 street-cluster bootstrap replicates. To compare the two vegetation correlations on the same 2,376 records, a paired bootstrap resampled the 2,208 street identifiers with replacement 10,000 times and retained the within-replicate difference between the upper-image/high-vegetation and full-image/vegetation coefficients; percentile intervals were calculated from the paired differences. A second paired bootstrap used 10,000 resamples of 317 occupied 100 m blocks after transforming all LiDAR coordinates to EPSG:2180; this common-coordinate grid grouped nearby records across source coordinate systems. Both procedures retained all records within each sampled unit and recalculated the two correlations together. Block resampling preserves within-block dependence but does not eliminate dependence across block boundaries. Nominal Fisher-z 95% intervals were also calculated for the Pearson coefficients at each aggregation support, treating those units as independent; these descriptive intervals do not adjust for residual spatial dependence. Time lag was computed from the LiDAR acquisition date and the midpoint of the reported GSV acquisition month, then stratified into ≤ 1, >1–3, >3–5, and >5 year absolute-lag bands. GSV months were also described using cautious leaf-on (May–September), leaf-off (November–March), and transition (April/October) proxies; these month groups were not treated as field-observed phenological states. The paired bi-seasonal TLS design of HuashuTrees illustrates why synchronized leaf-on/leaf-off evidence is materially different from an asynchronous audit (Xie et al., 2026).

Semantic QA had two components. First, two human raters independently scored an existing 200-image sample covering all audit classes, image-date periods, source groups, and displacement bands. The review materials included the images and supporting diagnostics, and each rater completed the assessment separately. Second, the same two raters independently assessed a targeted 60-image subset containing 32 difficult or automatically flagged cases and 28 unflagged cases. For this error audit they viewed each original image beside its SegFormer overlay and were blinded to automated ratios, downstream audit classes, earlier ratings, and each other's scores. They separately rated false-positive and false-negative severity (0 none, 1 minor/local, 2 major/material) for vegetation, sky, and facade; overall quality from 0 to 4; and predefined failure modes. Exact agreement and Cohen's kappa were reported, with quadratic weighting for ordinal scores. These components provided image-level semantic QA, while spatially separated assessment was emphasized because random folds can overstate generalization in spatial data (Roberts et al., 2017). Class-proportion uncertainty was estimated from 2,000 resamples of 410 source-coordinate 100 m blocks. The earlier 100 m aggregation and class-proportion checks retain these source-coordinate groupings; the additional paired-difference check uses the harmonized 317-block grid.

3 Results

The results address four urban forest monitoring questions: whether the two sensing streams cover the same analytical support, how strongly their vegetation indicators align, where cross-modal disagreement occurs, and how robust the resulting verification classes are to filtering, aggregation, and alternative classification rules.

3.1 Coverage and pairing

Metadata queries returned 2,424 successful sample points from 2,478 candidate locations. The 54 failures were spatially structured: their median local street length within 50 m was 316 m, vs. 596 m among successful candidates, and their median distance to the heritage boundary was 300 m, vs. 209 m. Half occurred on footways and 25.9% on service roads. Acquisition year could not be compared because a failed metadata query supplies no panorama date. Four headings per successful point produced 9,696 directional images associated with 1,713 unique panoramas. The LiDAR inventory comprised 29 files and 717.5 million cached points. The primary paired sample contained 2,376 records; 48 Street View points lacked LiDAR support within 30 m. These 48 records were concentrated nearer the sampled-area edge (median distance to the sample-hull boundary 168 m vs. 410 m for paired points) and had higher mean semantic vegetation (0.311 vs. 0.148), indicating that their exclusion may underrepresent greener edge and residential contexts. Among successful metadata queries, 490 panoramas were attached to more than one sample point, covering 1,201 sample records. The median query-to-panorama displacement was 5.86 m, the 95th percentile was 37.43 m, and the maximum was 179.84 m. A total of 2,015 successful metadata records were dated 2022 or later. Primary LiDAR assignment placed 1,610 paired records in source group 83,134 and 766 in group 75,174. Supplementary Figures S2, S3 and Supplementary Tables S8, S9 report the exclusion checks; Figure 3 reports the analytical hierarchy, image dates, displacement, and point density.

3.2 Cross-modal alignment

Table 3 reports the point-level coefficients at 30 m. Semantic vegetation and LiDAR vegetation had Pearson r = 0.694 and Spearman ρ = 0.750. The canopy-oriented comparison between upper-image vegetation and LiDAR high vegetation was modestly stronger (r = 0.717; ρ = 0.770) . In a 10,000-replicate paired street-cluster bootstrap, the correlation differences were Δr = 0.023 (95% interval 0.014–0.032) and Δρ = 0.020 (0.012–0.028). This modest difference is consistent with additional alignment when both measures emphasize vertical vegetation, but it does not isolate vertical position causally and remains a proxy comparison rather than tree-crown validation. Sky ratio and LiDAR tall obstruction had r = −0.747 and ρ = −0.736; sky ratio and LiDAR low-open ratio had r = 0.637 and ρ = 0.631. Facade ratio and LiDAR building ratio had r = 0.610 and ρ = 0.658. Restricting query-to-panorama displacement to 50 m retained 2,321 records (r = 0.715, ρ = 0.753); the 75 m filter retained 2,368 (r = 0.696, ρ = 0.750). The 2022-or-later subset retained 1,976 records (r = 0.699, ρ = 0.753), and the 2024-or-later subset retained 1,495 (r = 0.711, ρ = 0.756). A second paired bootstrap over 317 common-coordinate 100 m blocks gave Δr = 0.023 (95% interval 0.009–0.037) and Δρ = 0.020 (0.008–0.033), retaining positive differences with wider intervals. Supplementary Table S15 reports the paired-correlation comparison.

Table 3. Semantic-LiDAR alignment at 30 m.

The oldest primary LiDAR source was acquired on 5 October 2021; its corresponding retained GSV records ranged from August 2009 to April 2026, spanning 16 years and 8 months, with a median signed lag of 2.69 years after LiDAR. Across both primary sources, the absolute-lag median was 2.69 years. Vegetation alignment remained positive for ≤ 1 year (n = 1, 145, r = 0.709, ρ = 0.760), >1–3 years (n = 699, r = 0.676, ρ = 0.776), >3–5 years (n = 330, r = 0.688, ρ = 0.719), and >5 years (n = 202, r = 0.659, ρ = 0.653). Positive but attenuated correlation in the longest-lag stratum indicates persistence of shared spatial signal; it does not demonstrate robustness to tree growth, removal, construction, or other changes between acquisitions. The leaf-on and leaf-off month proxies had similar coefficients (r = 0.703 and 0.712), whereas the transition proxy was small and temporally unbalanced (n = 22) and was not interpreted as a phenology test. Supplementary Tables S1–S3 report complete date and lag results. Figure 4 plots the paired green and openness phase spaces and the primary classification thresholds.

3.3 Audit classes

In the paired sample, the green audit contained 281 legible-green-structure points, 27 latent-structural-greenness points, 59 image-visible-green points, 711 low-green-hardscape points, and 1,298 mixed/moderate points. The openness audit contained 337 legible-open points, 21 latent-open points, 88 image-visible-sky points, 547 high-enclosure points, and 1,383 mixed/moderate points. Counts and percentages are reported in Table 4, and their locations are shown in Figure 5. These hard classes are accompanied by continuous visible, structural, and difference surfaces in Supplementary Figure S1.

Table 4. Mutually exclusive visible-structural audit classes across all 2,376 valid paired records.

3.4 Reclassification outputs

Table 4 reports the unconditional, mutually exclusive distribution across the complete paired sample, whereas Table 5 conditionally decomposes predefined single-source screens. The low-visible-green screen (z_*nature* < −0.50) contained 974 points: 711 low-green hardscape, 18 latent structural greenness, and 245 mixed/moderate. The total latent class is 27 because its off-diagonal definition requires z_*nature* < −0.25, not the stricter −0.50 screen cutoff. Likewise, the low-LiDAR-vegetation screen (z_{*LiDAR vegetation*} < −0.50) contained 969 points, including 41 of the 59 image-visible-green points defined with the less restrictive counterpart cutoff below −0.25. The same threshold distinction explains the openness counts: 21 records meet the latent-openness definition using the counterpart cutoff z_*sky* < −0.25, of which 15 also satisfy the stricter low-visible-sky screen z_*sky* < −0.50. Table 5 lists every screen threshold, Figure 6 shows class composition, and Table 6 records the evidence required for bounded interpretation.

Table 5. Conditional reclassification of single-source screens within the 2,376-record paired analysis universe.

Low screens use z < −0.50 and high screens use z≥0.75; these filters are distinct from the off-diagonal counterpart cutoff z < −0.25 used in Table 4.

Table 6. Urban forest monitoring and field-verification matrix for bounded interpretation of FL FLthe mutually exclusive FL audit classes FL FLin Table 4; the matrix does not operate on the conditional single-source screens in Table 5FL.

3.5 Sensitivity, spatial robustness, and semantic QA

Across the primary, stricter-z, quartile, and tertile rules, latent structural greenness ranged from 12 to 38 points and image-visible green from 28 to 67. The residual rule returned 159 and 198 points, respectively. Agreement with the primary labels was ARI = 0.845 for stricter z, 0.445 for quartiles, 0.472 for tertiles, and 0.295 for the Gaussian mixture. The GMM partition did not reproduce the rule-based audit classes. This comparison provides no independent cluster-based support for those five labels, but does not determine whether the joint distribution contains natural clusters or how many there might be. The primary classes are transparent rule-based screening categories. Under the 100 m spatial-block bootstrap, latent structural greenness was 1.14% (95% interval 0.64%–1.67%) and image-visible green was 2.48% (1.59%–3.55%). Table 7, Figure 7, and the continuous surfaces in Supplementary Figure S1 jointly show why hard labels must remain screening aids.

Table 7. Threshold sensitivity and spatial-block uncertainty for green audit labels.

Source composition differed materially. Group 75,174 had mean point density 53.66 points/m² and mean overlap share 42.29%, compared with 25.03 points/m² and 14.24% in group 83,134. Nevertheless, source-specific vegetation correlations remained similar (r = 0.731 and 0.702). Excluding noise and overlap returns increased the overall coefficient to r = 0.720 (ρ = 0.754); restricting the denominator to recognized classes gave r = 0.720 (ρ = 0.754), and vegetation divided by ground plus vegetation gave r = 0.698 (ρ = 0.699). Corresponding latent-class counts were 38, 39, and 79 rather than 27, confirming that continuous alignment was more stable than class membership. Complete source composition, class proportions, and denominator checks are in Supplementary Tables S4–S6.

Spatial resampling and aggregation preserved direction but changed magnitude. The 2,000-replicate street-cluster bootstrap gave a 95% interval of 0.666–0.721 for Pearson r and 0.731–0.769 for Spearman ρ. Table 8 shows the point, street-ID, named-street, and 100 m grid results with nominal approximate Pearson intervals; Supplementary Table S7 reports the complete robustness matrix. The named-street estimate of r = 0.847 is based on 141 aggregated units (approximate 95% interval 0.793–0.888), so its apparent increase relative to the point estimate should not be interpreted as a precise gain or a superior estimate.

Table 8. Aggregation-scale sensitivity of semantic-vegetation vs. LiDAR-vegetation correlations.

FL FLPearson intervals are nominal Fisher-z approximations that treat units as independent and do not adjust for spatial dependence. The 100 m grid row retains the earlier source-coordinate grouping; the common-coordinate paired-difference check is reported in Supplementary Table S15. FL Aggregated coefficients describe different spatial supports and are not interpreted as superior estimates.

The original 200-image manual QA covered all five green classes, three image-date periods, both primary LiDAR groups, and all displacement bands. Its ordinal vegetation, sky, facade, road/sidewalk, clutter, segmentation-quality, and visible-character ratings had exact R1–R2 agreement of 71.5%–83.0% and quadratic-weighted kappa of 0.727–0.929. The targeted 60-image supplement included 32 automatically flagged or difficult cases and 28 routine cases; the two raters were blind to numeric outputs and audit classes. Any vegetation error was flagged in 11.7%–45.0% of images across raters, sky error in 3.3%–23.3%, and facade error in 16.7%–20.0% . Class-level error-presence agreement was only 63.3%–80.0%, with kappa 0.111–0.206. These low kappas and the wide vegetation-error range show that a minor-or-worse “any error” judgment was soft and rater-dependent; rater-specific rates are therefore retained. In contrast, both raters reported low major-error prevalence for vegetation ( 3.3%–8.3% ), sky ( 0%–5.0% ), and facade ( 1.7%–5.0% ), and rated most images at least 3 overall ( 83.3% and 93.3% ). For the targeted audit's full 0–4 overall-quality scale, exact score agreement was 38.3%, within-one-category agreement was 91.7%, and quadratic-weighted kappa was 0.367; the high proportions scoring at least 3 describe quality levels rather than strong agreement on exact scores. Full sampling, reliability, and failure-mode results are in Supplementary Tables S10–S13, with linked representative cases in Supplementary Table S14.

4 Discussion

4.1 From scalable sensing to audited urban forest monitoring

The results show that street-view semantics and airborne-LiDAR morphology share substantial vegetation signal but are not interchangeable observations. A high visible-green ratio identifies pedestrian-visible vegetation without proving canopy-equivalent structure; a high LiDAR vegetation ratio identifies classified structural returns without proving tree identity, health, ownership, accessibility, or benefit. The modest bootstrap-supported increase for the upper-image/high-vegetation comparison is consistent with vertically aligned measurement support, but it does not isolate a causal effect of vertical position and remains a proxy comparison rather than crown validation.

For urban forest management, retrieval, date, displacement, and panorama checks address coverage, whereas cross-modal comparison addresses construct correspondence. A monitoring record can pass the first audit and fail the second; neither audit alone authorizes ecological or management conclusions.

Temporal stratification showed positive alignment across lag groups, but persistence of a positive coefficient is weaker evidence than robustness to temporal change. The attenuation beyond 5 years, together with retained imagery extending from August 2009 to April 2026 and the small transition-month subset, indicates where acquisition timing remains most relevant. Bi-seasonal TLS data likewise illustrate that leaf-on and leaf-off observations encode different structural conditions (Xie et al., 2026). Operational use should therefore retain acquisition dates and reserve change interpretation for temporally aligned observations.

4.2 Visible and structural signals in urban morphology

The 281 high–high records support SVI as a measure of pedestrian-visible greenery, whereas 27 latent-structural-greenness and 59 image-visible-green records delimit that interpretation. Vegetation behind walls, inside courtyards, or above the view line can generate high LiDAR vegetation with weak image visibility . Conversely, planters, vines, hedges, low shrubs, and foreground composition can dominate image pixels without canopy-equivalent returns. These labels describe plausible sensor geometry, not verified mechanisms; field evidence is still needed for crown location, woody structure, permanence, condition, and ownership.

Pedestrian visibility depends on observer position and sight lines (Chen et al., 2025), while overhead indicators use a different support. The 3-30-300 rule similarly distinguishes views, canopy, and access (Konijnendijk, 2022). Cross-modal classes preserve those distinctions, but do not turn them into ecological types.

The negative sky/tall-obstruction and positive facade/building associations support morphological plausibility. Image-based enclosure describes what a pedestrian can see (Yin and Wang, 2016), while street-canyon and LiDAR measures describe surrounding three-dimensional structure (Hu et al., 2020). Off-diagonal openness labels should therefore prompt checks of heading, setbacks, courtyards, and overhead form rather than automatic claims about urban quality.

4.3 Spatial support and classification uncertainty

Point, street, and grid coefficients retain their expected direction but change magnitude, consistent with grouping effects and the modifiable areal unit problem (MAUP) (Openshaw, 1984; Holt et al., 1996). A 30 m point-level mismatch therefore cannot be assumed to persist as a street or neighborhood condition, and a camera buffer remains only an approximation of the observer-relevant environment (Kwan, 2012).

Threshold and denominator sensitivity provide a second uncertainty layer. Alternative rules and return denominators changed rare-class membership more than continuous correlations, and the GMM partition did not reproduce the rule-based audit categories. The comparison measures agreement between partitions rather than establishing the existence or number of natural clusters. Continuous maps should therefore be inspected alongside hard labels, especially near decision boundaries. Latent structural greenness and image-visible green are navigation labels whose rule and sensitivity range must accompany any implementation, not stable urban forest types.

4.4 Implications for urban forest monitoring and management

Reclassification is useful because a single-source low-visible-green screen combines 711 joint-low locations, 18 structurally high but poorly visible cases, and 245 moderate cases. These groups require different follow-up: data-quality and planting-feasibility checks for joint-low sites; crown-location, property-edge, and access checks for latent structural greenness; and permanence, height, woody structure, and shade checks for image-visible green. Table 6 organizes those questions for the mutually exclusive audit classes in Table 4, not for the conditional screens in Table 5; the categories are not prescriptions to plant, remove, prune, or open a view.

The rare off-diagonal classes identify cases for focused manual verification. The targeted semantic QA helps direct that verification toward class-specific segmentation issues, while inventories, ownership, condition, utilities, soils, heat exposure, and maintenance records provide the evidence needed for subsequent prioritization.

4.5 Historic urban context, limitations, and transferability

Kazimierz makes visual-structural mismatch especially legible because narrow streets, enclosed courtyards, protected built fabric, and fragmented ownership can separate vegetation presence from street-level visibility. These conditions are useful for testing a monitoring audit, but heritage or planting decisions cannot be reduced to sensor agreement. The workflow identifies where vegetation visibility and structure deserve joint inspection; it does not determine whether planting, retaining enclosure, altering maintenance, or opening a view is culturally or ecologically appropriate.

The study has six main limitations. First, it uses one compact historic district, so cross-city, cross-platform, and cross-morphology transfer remains untested. Because every z score is standardized within this sample, the same numeric threshold will correspond to a different absolute vegetation level in a greener or sparser district; transfer requires recalibration of the reference distribution rather than mechanical reuse of the cutoffs. Second, asynchronous acquisitions leave residual phenological and urban-change uncertainty . The retained GSV imagery spans August 2009 to April 2026 relative to LiDAR acquired in 2021–2025, and positive correlation at long lags does not establish robustness to intervening tree growth, removal, or construction. Third, generic Cityscapes classes contain domain and label error ; 512-pixel inference also differs from the checkpoint's 1,024 × 1,024 fine-tuning regime, and the targeted QA showed that minor-error presence was rater-dependent even though major-error prevalence was low across both raters. Fourth, LiDAR ratios remain sensitive to point density, overlap coding, source processing, denominator choice, and buffer geometry . The fallback group 75,174 contributed 766 records and had a 42.29% mean overlap-return share, compared with 14.24% in group 83,134; similar source-specific correlations are reassuring but do not remove this processing asymmetry. Alternative denominators supported the continuous relationship but changed rare-class membership. Fifth, GSV retrieval failure and missing LiDAR support were spatially structured, so the analytical sample may underrepresent some edge, service-road, and greener residential contexts. Sixth, neither modality observes individual-tree identity, species, health, thermal performance, biodiversity, access, ownership, or resident valuation. Thresholds should be re-estimated or preregistered for other cities, seasons, platforms, LiDAR classifications, and morphologies. Future work can combine synchronized acquisitions and independent urban-tree inventories with emerging individual-tree segmentation, attribute-estimation, and species-mapping resources (Wang and Liang, 2026; Li et al., 2026b,a) to test transferability across cities and seasons.

5 Conclusions

This study reframes urban forest monitoring as a multimodal audit of visible and structural vegetation signals. Across 2,376 valid Street View-LiDAR pairs, image-derived visible indicators and LiDAR-derived structural indicators were correlated but non-equivalent. Their disagreement separated single-source greenness screens into verification classes with different evidence needs. The canopy-oriented upper-image and high-vegetation comparison was strong, but the available classifications do not constitute individual-tree or crown validation.

The management contribution is a reproducible sequence: retain sensor provenance, align spatial and temporal supports, compare observed constructs rather than declaring a universal ground-truth sensor, expose cross-modal disagreement, report classification uncertainty, and connect each class to a specific field-verification need. The workflow can support tree-inventory review and inspection planning while avoiding automatic planting or removal prescriptions. Its transferability remains a hypothesis for subsequent multi-city, multi-season research linked to synchronized acquisitions and independent urban forest field data.

Generative AI statement

The author(s) declared that Generative AI was used in the creation of this manuscript. During the preparation of this manuscript, the author(s) used OpenAI's ChatGPT (GPT-5.6, OpenAI) solely for English-language editing and stylistic polishing, including improvements to grammar, clarity, concision, and academic phrasing. The tool was not used to generate research data, conduct analyses, develop scientific interpretations, draw conclusions, or generate references. All AI-assisted text was carefully reviewed, verified, and edited by the author(s). The author(s) take full responsibility for the accuracy, originality, and integrity of the final manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/ffgc.2026.1960003/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Yashwant Singh Rawat, Federal Technical and Vocational Education and Training Institute (FTVETI), Ethiopia

Reviewed by: Shaoqing Dai, Wuhan University, China

S. Vani, Chaoyang University of Technology, Taiwan

Footnotes

SVI, Street-view imagery; GSV, Google Street View; LiDAR, Light detection and ranging; LAS, LAS point-cloud format; LAZ, Compressed LAS point-cloud format; OSM, OpenStreetMap; QA, Quality assurance; ARI, Adjusted Rand index; GMM, Gaussian mixture model; MAUP, Modifiable areal unit problem; TLS, Terrestrial laser scanning; E_2*D*, Two-dimensional enclosure index.