MLLMs, Street View and Urban Policy-Intelligence: Recovering the Sustainability Effects of Redlining

This repository contains the replication code and data for the paper:

"MLLMs, Street View and Urban Policy-Intelligence: Recovering the Sustainability Effects of Redlining"
npj Urban Sustainability

---

Overview

This study evaluates whether multimodal large language models (MLLMs) can derive neighborhood-level sustainability indicators from Google Street View (GSV) imagery and recover the legacy effects of historical redlining in the Phoenix metropolitan area. We compare MLLM-based inference (GPT-4o) against conventional semantic segmentation (ResNet-based) and authoritative benchmarks (ACS poverty rates, GEIE tree canopy coverage).

---

Repository Structure

MLLM_GeoAI_Redlining/
└── MainScript/Dataset/
    ├── DL_LLM_Census_Analysis.R          # Main analysis script
    ├── DL_LLM_Census_Analysis.RData      # R workspace with processed data
    ├── LLM_Results_All_phx_gilbert_Structured_Update.csv   # MLLM inference results
    ├── merged_summary.csv                # Merged summary statistics
    └── README.md                         # This file

---

File Descriptions

Scripts

File	Description
DL_LLM_Census_Analysis.R	Main R script for all analyses including: data preparation, prediction accuracy comparisons, OLS/ZIP-FE/SAR regression models, bootstrap inference, and figure/table generation

Data

File	Description
DL_LLM_Census_Analysis.RData	R workspace containing processed spatial data, model objects, and intermediate results
LLM_Results_All_phx_gilbert_Structured_Update.csv	Image-level MLLM inference results from GPT-4o, including poverty indicators, tree canopy estimates, and housing classifications for Phoenix and Gilbert GSV panoramas
merged_summary.csv	CBG-level merged dataset combining MLLM predictions, semantic segmentation outputs, and authoritative benchmark data

---

Requirements

R Packages

# Core packages
library(tidyverse)
library(sf)
library(spdep)
library(spatialreg)

# Data retrieval
library(tidycensus)

# Tables and visualization
library(kableExtra)
library(ggplot2)
library(patchwork)

# Additional
library(boot)
library(stargazer)

Software

• R (≥ 4.0)
• RStudio (recommended)

---

Usage

1. Clone the repository:

   git clone https://github.com/[username]/MLLM_GeoAI_Redlining.git

2. Open DL_LLM_Census_Analysis.R in RStudio

3. Load the workspace:

   load("MainScript/Dataset/DL_LLM_Census_Analysis.RData")

4. Run the analysis script to reproduce all figures and tables

---

Key Analyses

The script performs the following analyses:

1. Prediction Accuracy — Correlations between MLLM/ResNet predictions and authoritative benchmarks (ACS poverty, GEIE canopy)

2. Redlining Legacy Effects — OLS, ZIP code fixed effects, and spatial autoregressive (SAR) models comparing measurement approaches

3. Bootstrap Inference — Nonparametric bootstrap for robust confidence intervals

4. ACS Margin of Error Analysis — Diagnostic analysis of measurement uncertainty by HOLC classification

---

Data Sources

• Google Street View — Panoramic imagery sampled at 100m intervals within CBGs
• American Community Survey (ACS) — 5-year poverty estimates (2018–2022)
• GEIE Tree Canopy — Tree Equity Score canopy coverage data
• HOLC Maps — Historical redlining boundaries from the Mapping Inequality project

---

Citation

If you use this code or data, please cite:

@article{howell2025mllm,
  title={MLLMs, Street View and Urban Policy-Intelligence: Recovering the Sustainability Effects of Redlining},
  author={Howell, Anthony and [co-authors]},
  journal={npj Urban Sustainability},
  year={2025}
}

---

License

This project is licensed under CC BY 4.0, consistent with npj Urban Sustainability's open access policy.

---

Contact

For questions about the code or data, please open an issue or contact the corresponding author.