Planning Site Formation

• Planning Site Formation is a process that defines explicit spatial constructs using heterogeneous evidence and context-specific constraints.
• It leverages diverse methods such as adversarial generative models, variational graph auto-encoders, and self-organizing maps to structure planning units.
• The approach integrates spatial encoding, evaluation metrics, and human oversight to ensure feasibility, coherence, and operational usability in planning tasks.

Planning site formation, as realized in recent computational planning literature, denotes the formation of spatial entities that planners act upon: land-use configuration tensors for unplanned urban areas, contiguous demand-oriented regions for adaptation planning, excavation workspaces and dig sequences, screened or ranked infrastructure sites, and machine-readable site boundaries reconstructed from planning documents (Wang et al., 2020, Noorani et al., 13 Nov 2025, Terenzi et al., 2023, Radu et al., 2021, Kim et al., 14 May 2026, Degen et al., 1 Jun 2026). Across these settings, the planning object is not merely a label or forecast but an explicit spatial construct with geometry, adjacency, feasibility, and operational implications. The resulting methods span adversarial generative models, Variational Graph Auto-Encoders, representative initialized spatially constrained self-organizing maps, linear programming, SCUC/SCED-based screening, coverage planning, and multimodal geospatial reconstruction.

1. Problem classes and formal planning objects

The literature instantiates planning site formation in several technically distinct ways. In automated urban planning, Wang et al. define a land-use configuration as a longitude-latitude-channel tensor and cast planning as generating that tensor for an unplanned area from surrounding contexts (Wang et al., 2020, Wang et al., 2021). In disaster adaptation planning, Noorani et al. define a “demand-oriented planning site” as a spatial unit that is a contiguous collection of grid cells, generated dynamically from high-resolution socioeconomic and environmental data rather than from pre-existing administrative boundaries (Noorani et al., 13 Nov 2025). In autonomous excavation planning, the planning object is a sequence of excavator base poses whose local workspaces cover the entire dig region while preserving machine mobility (Terenzi et al., 2023). In renewable generation site selection and planner-initiated data-center siting, the planning object is a screened or ranked subset of candidate sites subject to system-level feasibility criteria (Radu et al., 2021, Kim et al., 14 May 2026). In document-grounded planning-record reconstruction, the object is a valid GeoJSON boundary recovered from indirect spatial evidence (Degen et al., 1 Jun 2026).

Planning setting	Planning object	Formal representation
Automated urban planning	Land-use configuration	$M \in \mathbb{R}^{n \times n \times m}$ or $M \in \mathbb{R}^{n \times n \times C}$
Adaptation regionalization	Demand-oriented planning site	$S=\{S_1,\dots,S_K\}$ partition of $N$ cells
Autonomous excavation	Coverage-valid work sequence	$B_1,\dots,B_N$ over a 2.5D elevation map
Renewable screening	Relevant RES sites	$\mathcal N^n_{\mathrm{SITE}}$
Flexibility-aware siting	Pre-qualified bus-envelope pairs	$\mathcal Q_f^{(1)}$
Planning-record reconstruction	Site boundary	GeoJSON Feature, Polygon or MultiPolygon

This variety shows that “site” is domain-specific. In one class of problems, the site is the area to be configured; in another, it is the unit to be delineated; in another, it is the location to be selected; and in another, it is the boundary to be reconstructed. A plausible implication is that the unifying technical issue is the translation of heterogeneous evidence into spatially explicit planning units that satisfy domain constraints.

2. Representational primitives and input modalities

A consistent theme is the use of structured spatial encodings. In LUCGAN, a central site $R$ is a $1\,\mathrm{km}\times 1\,\mathrm{km}$ square divided into an $n\times n$ grid, with $M \in \mathbb{R}^{n \times n \times C}$0 equal to the number of POIs of type $M \in \mathbb{R}^{n \times n \times C}$1 whose $M \in \mathbb{R}^{n \times n \times C}$2 falls into cell $M \in \mathbb{R}^{n \times n \times C}$3; one reported implementation uses $M \in \mathbb{R}^{n \times n \times C}$4, while the later systematic study uses $M \in \mathbb{R}^{n \times n \times C}$5 by default and experiments with $M \in \mathbb{R}^{n \times n \times C}$6 (Wang et al., 2020, Wang et al., 2021). The surrounding context is represented by eight adjacent context squares together with feature blocks for monthly housing-price change, POI-category ratios, bus metrics, and taxi metrics, assembled into an attributed spatial graph for VGAE-based embedding (Wang et al., 2020, Wang et al., 2021).

The hierarchical urban-planning formulation in IHPlanner replaces a direct grid-only view with a three-level representation: a Spatial Attributed Graph $M \in \mathbb{R}^{n \times n \times C}$7 for surrounding regions, latent zone labels $M \in \mathbb{R}^{n \times n \times C}$8, and fine-grained grid assignments $M \in \mathbb{R}^{n \times n \times C}$9 (Wang et al., 2022). Human instructions are embedded as $S=\{S_1,\dots,S_K\}$0 and concatenated with the graph-derived context embedding, so regulatory targets such as greening rate, FAR, and setback enter the generative process as explicit conditioning variables.

In RepSC-SOM, the study area is discretized into $S=\{S_1,\dots,S_K\}$1 grid cells, each cell $S=\{S_1,\dots,S_K\}$2 carrying a feature vector $S=\{S_1,\dots,S_K\}$3 such as flood-risk, income, or population density, and a planning site is a partition whose components are spatially contiguous and internally homogeneous in those features (Noorani et al., 13 Nov 2025). In excavation planning, the site is first mapped from LIDAR point clouds into a georeferenced point cloud in the ECEF frame and then rasterized into a 2.5D elevation map at $S=\{S_1,\dots,S_K\}$4 resolution, with elevation and occupancy/traversability layers; final excavation geometry is encoded through user-defined cut and dump polygons with target elevation $S=\{S_1,\dots,S_K\}$5 (Terenzi et al., 2023).

The energy-siting papers adopt network-level abstractions. Radu et al. represent candidate renewable sites by capacity variables $S=\{S_1,\dots,S_K\}$6, dispatch $S=\{S_1,\dots,S_K\}$7, and technical potential bounds $S=\{S_1,\dots,S_K\}$8 in a two-stage CEP formulation (Radu et al., 2021). The flexibility-aware data-center framework defines standardized load envelopes $S=\{S_1,\dots,S_K\}$9 over 24 hours for $N$0, together with contingency-feasibility indicators $N$1 and pass rates $N$2 (Kim et al., 14 May 2026). Plan2Map starts from raw planning PDFs and constructs a structured record containing site address, postcodes, British-Grid references, road names, place names, map scale, and page-level metadata before registration and segmentation (Degen et al., 1 Jun 2026).

3. Generative formation of urban land-use sites

The adversarial urban-planning line begins from the claim that land-use configuration can be treated as a deep generative learning problem conditioned on surrounding context (Wang et al., 2020, Wang et al., 2021). In the 2020 LUCGAN formulation, the surrounding area is encoded as a 9-node graph $N$3 with adjacency $N$4, node-attribute matrix $N$5, and a VGAE objective

$N$6

with two GCN encoder layers, a decoder $N$7, global average pooling over node embeddings, and reported settings $N$8 and $N$9 (Wang et al., 2020). The generator maps the context embedding $B_1,\dots,B_N$0 through a fully connected layer, reshape, and a cascade of ConvTranspose2d layers with BatchNorm and ReLU to an $B_1,\dots,B_N$1 nonnegative tensor $B_1,\dots,B_N$2, while the discriminator applies Conv2d layers with stride $B_1,\dots,B_N$3, LeakyReLU activations, and sigmoid output (Wang et al., 2020). The discriminator is trained against three classes of tensors: excellent plans, terrible plans, and generated plans. The quality score used for excellent-versus-terrible labeling is

$B_1,\dots,B_N$4

where $B_1,\dots,B_N$5 is total check-ins in $B_1,\dots,B_N$6 and $B_1,\dots,B_N$7 is POI-type Shannon-diversity (Wang et al., 2020).

The later systematic treatment by Wang et al. keeps the tensorial formulation but introduces LUCGAN$B_1,\dots,B_N$8, which performs conditioning augmentation on the context embedding,

$B_1,\dots,B_N$9

and concatenates this with additional noise to form the generator input $\mathcal N^n_{\mathrm{SITE}}$0 (Wang et al., 2021). On Beijing data comprising 2,990 residential communities, 328,668 POIs in $\mathcal N^n_{\mathrm{SITE}}$1 classes, taxi GPS traces, bus smart-card transactions, housing prices, and Weibo check-in logs, the paper reports that at $\mathcal N^n_{\mathrm{SITE}}$2, LUCGAN$\mathcal N^n_{\mathrm{SITE}}$3 outperforms the best baseline by KL $\mathcal N^n_{\mathrm{SITE}}$4, JS $\mathcal N^n_{\mathrm{SITE}}$5, Hellinger $\mathcal N^n_{\mathrm{SITE}}$6, and Wasserstein $\mathcal N^n_{\mathrm{SITE}}$7, and improves over original LUCGAN by $\mathcal N^n_{\mathrm{SITE}}$8, $\mathcal N^n_{\mathrm{SITE}}$9, $\mathcal Q_f^{(1)}$0, and $\mathcal Q_f^{(1)}$1, respectively (Wang et al., 2021).

IHPlanner extends this line by making site formation hierarchical and human-instructed rather than directly tensor-conditional (Wang et al., 2022). The model first generates latent zone labels $\mathcal Q_f^{(1)}$2, then computes functionality projections

$\mathcal Q_f^{(1)}$3

and finally models peer dependencies among the $\mathcal Q_f^{(1)}$4 projections using multi-head self-attention,

$\mathcal Q_f^{(1)}$5

A planning layer reshapes the resulting representation into $\mathcal Q_f^{(1)}$6, and training minimizes a total objective combining stage-1 GAN loss, KL regularization, and stage-3 reconstruction loss (Wang et al., 2022). On the Beijing case with five green-rate levels $\mathcal Q_f^{(1)}$7, IHPlanner attains the lowest AVG_KL, AVG_JS, AVG_HD, and AVG_Cos across all levels and remains robust over $\mathcal Q_f^{(1)}$8 (Wang et al., 2022).

These urban models share three commitments: explicit quantification of the planning object, learned context encoders that preserve spatial dependence, and evaluation against both distributional criteria and planner-defined quality proxies. At the same time, the papers explicitly state limits: the framework is “purely statistical,” does not explicitly enforce domain-specific zoning rules such as setback or FAR, and still recommends final approval by a human planner (Wang et al., 2021, Wang et al., 2020).

4. Spatially constrained regionalization and infrastructure siting

In adaptation planning, site formation becomes the generation of dynamic planning units rather than land-use tensors. Noorani et al. define a demand-oriented planning site as a partition $\mathcal Q_f^{(1)}$9 of $R$0 grid cells such that each $R$1 is spatially contiguous and internally homogeneous in features relevant to the planner’s disaster-adaptation objectives (Noorani et al., 13 Nov 2025). RepSC-SOM extends classical SOM with representative seed initialization, adaptive geographic filtering of Best Matching Unit assignments, and post-SOM region-growing refinement. The SOM update follows Kohonen: $R$2 and geographic filtering restricts BMU search to nodes whose geographic distance is within threshold $R$3, where $R$4 is chosen from the semivariogram “range” of the selected features (Noorani et al., 13 Nov 2025). After convergence, regions with the same BMU label are merged and refined; the paper gives an illustrative similarity score

$R$5

Evaluation includes within-region variance, a silhouette coefficient adapted to contiguous clusters, a fragmentation index, and comparison against baseline units such as census tracts (Noorani et al., 13 Nov 2025).

Radu et al. treat site formation as model reduction through renewable generation site selection in capacity expansion planning (Radu et al., 2021). Their two-stage method begins with SITE, a screening LP that keeps only those RES sites whose omission would significantly change system cost or design. The objective is

$R$6

subject to feed-in target, dispatch-limit, and technical-potential constraints; any site with $R$7 in the SITE solution is deemed relevant and retained for the reduced CEP problem (Radu et al., 2021). On a 33-country ENTSO-E case with 1,740 candidate sites, SITE consistently retrieves over $R$8 of the optimal RES sites, yields objective difference $R$9 between reduced and full LP, and reduces peak memory by $1\,\mathrm{km}\times 1\,\mathrm{km}$0–$1\,\mathrm{km}\times 1\,\mathrm{km}$1 and total solver time by $1\,\mathrm{km}\times 1\,\mathrm{km}$2–$1\,\mathrm{km}\times 1\,\mathrm{km}$3 (Radu et al., 2021).

The 2026 flexibility-aware framework moves from screening candidate generators to planner-initiated siting of large flexible loads (Kim et al., 14 May 2026). Stage 1 defines the N–1 pass rate

$1\,\mathrm{km}\times 1\,\mathrm{km}$4

with default threshold $1\,\mathrm{km}\times 1\,\mathrm{km}$5. Stage 2 evaluates each qualified $1\,\mathrm{km}\times 1\,\mathrm{km}$6 pair by year-hourly SCUC/SCED. Stage 3 ranks alternatives by entropy-weighted multi-criteria scoring and TOPSIS closeness coefficient $1\,\mathrm{km}\times 1\,\mathrm{km}$7 (Kim et al., 14 May 2026). The three planner-issued envelopes are firm, pause, and shift, defined by Eq. (14) with a within-day ramp-rate limit $1\,\mathrm{km}\times 1\,\mathrm{km}$8. On the synthetic 2,000-bus Texas system, operational flexibility expands the pre-qualified set from 193 firm buses to 226 pause buses and 210 shift buses at $1\,\mathrm{km}\times 1\,\mathrm{km}$9, and from 57 firm buses to 68 pause buses and 69 shift buses at $n\times n$0, while median all-hour mean LMP remains $n\times n$1 for the $n\times n$2 cases (Kim et al., 14 May 2026).

Taken together, these methods replace inherited planning units or exhaustive candidate sets with generated, screened, or ranked sites whose admissibility is explicitly tied to homogeneity, reliability, or system-cost fidelity.

5. Geometric site formation, excavation, and boundary reconstruction

In autonomous excavation, planning site formation has a literal geometric meaning: the system must form a valid sequence of workspaces and bucket motions that transform present topography into the target geometry (Terenzi et al., 2023). The site is first mapped by LIDAR and ICP-based SLAM with GPS-RTK loop closures into a georeferenced point cloud and then rasterized into a 2.5D elevation map with elevation and occupancy/traversability layers. The user specifies target cut and dump polygons in Google Earth Pro with target elevations $n\times n$3, and the planner identifies where $n\times n$4 (Terenzi et al., 2023). The global planner performs Boustrophedon decomposition along orientation $n\times n$5, constructs a quotient graph, solves a Minimum-Branching-Vertices Spanning Tree Problem, and uses dynamic programming with recurrence

$n\times n$6

to choose coverage subroutines and entry/exit corners. Local excavation planning partitions the workspace into five radial-angular zones and selects dump zones by minimizing

$n\times n$7

The digging planner then optimizes the attack point $n\times n$8 under reach constraints using Bayesian optimization with a Gaussian-process surrogate and Expected Improvement, requiring about 30 evaluations rather than 150 naive grid samples (Terenzi et al., 2023). The reported system excavates a $n\times n$9 pit in $M \in \mathbb{R}^{n \times n \times C}$00, achieves $M \in \mathbb{R}^{n \times n \times C}$01, and improves final grade absolute error from about $M \in \mathbb{R}^{n \times n \times C}$02 before refinement to about $M \in \mathbb{R}^{n \times n \times C}$03 after refinement (Terenzi et al., 2023).

Plan2Map addresses a different geometric problem: reconstructing valid geospatial site boundaries from planning documents that contain only indirect spatial evidence (Degen et al., 1 Jun 2026). The benchmark contains 208 UK planning-record cases, and GeoPlanAgent decomposes the task into evidence extraction, localisation, map registration, boundary segmentation, projection, and verification. The Reader produces a structured record from raw planning PDFs; the Locate sub-agent uses signals such as site address, postcodes, grid references, house-number road pairs, and visible map labels to infer candidate coordinates; and the Worker performs auto-rotation with a 4-class ResNet50 classifier, sliding-window tile matching with MINIMA-LoFTR plus RANSAC, boundary segmentation with SAM 3 plus LoRA adapters, affine projection to WGS84, vectorisation, and GeoJSON assembly (Degen et al., 1 Jun 2026). The segmentation model uses

$M \in \mathbb{R}^{n \times n \times C}$04

with focal, Dice, surface, classification, and presence terms, LoRA rank 16, $M \in \mathbb{R}^{n \times n \times C}$05, dropout $M \in \mathbb{R}^{n \times n \times C}$06, and AdamW with learning rate $M \in \mathbb{R}^{n \times n \times C}$07 (Degen et al., 1 Jun 2026). On the full benchmark, GeoPlanAgent + Critic reports mean IoU $M \in \mathbb{R}^{n \times n \times C}$08, median IoU $M \in \mathbb{R}^{n \times n \times C}$09, IoU $M \in \mathbb{R}^{n \times n \times C}$10 on $M \in \mathbb{R}^{n \times n \times C}$11 of cases, median error $M \in \mathbb{R}^{n \times n \times C}$12, [email protected] $M \in \mathbb{R}^{n \times n \times C}$13, cost $M \in \mathbb{R}^{n \times n \times C}$14 per document, and time $M \in \mathbb{R}^{n \times n \times C}$15 per document (Degen et al., 1 Jun 2026).

These two systems show complementary meanings of geometric site formation. Excavation planning generates executable geometry-changing actions from terrain state and machine constraints. Boundary reconstruction generates machine-readable spatial extents from documentary evidence and map imagery. In both cases, the output must be geometrically valid and operationally usable rather than merely semantically plausible.

6. Evaluation regimes, human oversight, and recurring limitations

Evaluation in planning site formation is strongly task-specific, but the reported metrics cluster around four concerns: plausibility, internal coherence, system feasibility, and geometric accuracy. For urban land-use generation, Wang et al. use KL divergence, JS divergence, Hellinger distance, Wasserstein distance, and a separate quality-scoring model trained on real configurations (Wang et al., 2021). The earlier LUCGAN report also reserves $M \in \mathbb{R}^{n \times n \times C}$16 of hand-labeled excellent/terrible maps to train a random-forest scorer, obtaining average generated-map scores of about $M \in \mathbb{R}^{n \times n \times C}$17 for LUCGAN, $M \in \mathbb{R}^{n \times n \times C}$18 for VAE-only, $M \in \mathbb{R}^{n \times n \times C}$19 for AVG, and $M \in \mathbb{R}^{n \times n \times C}$20 for MAX, while noting that MAX is context-agnostic (Wang et al., 2020). IHPlanner uses AVG_KL, AVG_JS, AVG_HD, and AVG_Cos across green-rate levels (Wang et al., 2022). RepSC-SOM evaluates within-region variance, an adapted silhouette coefficient, fragmentation, and hazard-aligned homogeneity versus baseline units (Noorani et al., 13 Nov 2025). Renewable site screening uses spatial reduction $M \in \mathbb{R}^{n \times n \times C}$21, screening accuracy $M \in \mathbb{R}^{n \times n \times C}$22, TSCE, capacity differences, problem-size reduction, peak memory, and solver time (Radu et al., 2021). Flexibility-aware siting emphasizes pass rate, mean LMP, P95–P5 price dispersion, binding hours, congestion rent, and TOPSIS closeness coefficients (Kim et al., 14 May 2026). Excavation planning reports coverage fraction, planning-time success, cycle time, soil-movement rate, and grade error (Terenzi et al., 2023). Plan2Map uses IoU, centroid-distance error, [email protected], runtime, and cost (Degen et al., 1 Jun 2026).

Human oversight remains explicit across the literature. LUCGAN is described as an assistant rather than a full replacement, and final approval by a human planner is still recommended (Wang et al., 2020). Noorani et al. place the planner inside the loop for feature selection, spatial constraint override, desired region count, and interactive exploration (Noorani et al., 13 Nov 2025). The flexibility-aware siting framework is planner-initiated and produces ranked, pre-certified catalogues rather than automatic commitments (Kim et al., 14 May 2026). GeoPlanAgent includes an optional Critic that can approve, switch candidates, or request relocalisation (Degen et al., 1 Jun 2026).

The limitations are similarly recurrent. The urban generative models are Beijing-centric and do not explicitly encode zoning rules such as setback and FAR, although a later LUCGAN implementation notes that fine-tuning on as few as 200 labeled excellent/terrible samples may support transfer to a new region when the POI taxonomy is comparable (Wang et al., 2021, Wang et al., 2020). RepSC-SOM requires good-quality fine-grained data, imposes computational overhead for very large $M \in \mathbb{R}^{n \times n \times C}$23, and may require legal or political buy-in if dynamic units replace official zones (Noorani et al., 13 Nov 2025). SITE depends on manual choices of $M \in \mathbb{R}^{n \times n \times C}$24 and $M \in \mathbb{R}^{n \times n \times C}$25, and its screening LP only approximates storage and transmission effects (Radu et al., 2021). The data-center framework assumes standardized flexibility envelopes and DC-OPF-based reliability screening (Kim et al., 14 May 2026). Plan2Map shows that direct VLM-to-GeoJSON prediction remains unreliable, with remaining errors concentrated in localisation and map registration (Degen et al., 1 Jun 2026). Excavation planning, although fully autonomous in the reported system, still depends on accurate mapping, target geometry specification, and navigation/execution integration (Terenzi et al., 2023).

A plausible implication of these convergent findings is that planning site formation is becoming a multi-stage discipline in which representation design, spatial constraint handling, and post hoc verification are at least as important as the core learning or optimization module. The cited work does not present a single universal framework, but it does define a coherent computational agenda: explicit spatial objects, heterogeneous context ingestion, domain-specific feasibility filters, and measurable outputs that can be inspected, revised, and deployed.