Policy-prior learning-based cellular automata for urban expansion simulation via inference-time logit-level policy steering

Learning-based Cellular Automata (CA) can reproduce historical urban expansion, but many trained models remain difficult to steer when planning constraints or development priorities change after calibration. This study develops a policy-prior learning-based CA that steers a frozen ViT-NS-TopK-CA through inference-time logit-space injection. The base model combines a vision Transformer suitability branch with a TopK neighborhood operator to estimate transition logits from historical growth dynamics. Three planning-derived layers then modify these logits before ranked allocation, without changing model weights: a hard ecological constraint mask, a soft cropland/boundary deterrence field, and a positive guidance field representing transport corridors and urban clusters. Evaluation pairs Figure of Merit (FoM) and urban-class Intersection over Union (IoU) with Constraint Violation Rate (CVR) and Guidance Response Rate (GRR). Across five seeds and three five-year steps in Wuhan (2005–2020), the full policy configuration yields zero hard-constraint violations, with high-guidance response above 98% initially and 81.6% in the final step. Relative to the unconstrained baseline, it reduces FoM from 0.439 to 0.434 (a 1.2% cost) and produces fewer, larger, less fragmented expansion patches. Spatial diagnostics reveal fewer hard-zone false positives and more high-guidance true positives, alongside increased high-guidance false positives and omissions in hard and some low-guidance near-urban zones. Post-allocation masking achieves zero hard-constraint violation only by deleting 13.87% of demand or requiring 14.64% refill allocation, with a more fragmented footprint. A same-city 2020–2024 policy-alignment check retains zero CVR and 96.7% final-step GRR, supporting logit-level policy priors as a before-allocation steering interface for urban simulation.

Authors

Institutions

Publication Details

Journal
Geo-spatial Information Science
Published
2026-11-18
DOI
https://doi.org/10.1080/10095020.2026.2753886
Primary Topic
Land Use and Ecosystem Services
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Policy-prior learning-based cellular automata for urban expansion simulation via inference-time logit-level policy steering

Jiabo Xu, Haofeng Xie, Yibing Xiong, Jinhui Liu et al.
Geo-spatial Information Science
Land Use and Ecosystem Services
article

Policy-prior learning-based cellular automata for urban expansion simulation via inference-time logit-level policy steering

Jiabo Xu, Haofeng Xie, Yibing Xiong, Jinhui Liu, Xiangyun Hu, Kai Deng, Quanwei Xing
article en

Abstract

Learning-based Cellular Automata (CA) can reproduce historical urban expansion, but many trained models remain difficult to steer when planning constraints or development priorities change after calibration. This study develops a policy-prior learning-based CA that steers a frozen ViT-NS-TopK-CA through inference-time logit-space injection. The base model combines a vision Transformer suitability branch with a TopK neighborhood operator to estimate transition logits from historical growth dynamics. Three planning-derived layers then modify these logits before ranked allocation, without changing model weights: a hard ecological constraint mask, a soft cropland/boundary deterrence field, and a positive guidance field representing transport corridors and urban clusters. Evaluation pairs Figure of Merit (FoM) and urban-class Intersection over Union (IoU) with Constraint Violation Rate (CVR) and Guidance Response Rate (GRR). Across five seeds and three five-year steps in Wuhan (2005–2020), the full policy configuration yields zero hard-constraint violations, with high-guidance response above 98% initially and 81.6% in the final step. Relative to the unconstrained baseline, it reduces FoM from 0.439 to 0.434 (a 1.2% cost) and produces fewer, larger, less fragmented expansion patches. Spatial diagnostics reveal fewer hard-zone false positives and more high-guidance true positives, alongside increased high-guidance false positives and omissions in hard and some low-guidance near-urban zones. Post-allocation masking achieves zero hard-constraint violation only by deleting 13.87% of demand or requiring 14.64% refill allocation, with a more fragmented footprint. A same-city 2020–2024 policy-alignment check retains zero CVR and 96.7% final-step GRR, supporting logit-level policy priors as a before-allocation steering interface for urban simulation.

Geo-spatial Information Science
Wuhan University (CN), Wuhan University of Science and Technology (CN)
Sustainable cities and communities
Land Use and Ecosystem Services
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon