💻 computer science

Conflated Inverse Modeling to Generate Diverse and Temperature-Change Inducing Urban Vegetation Patterns

This paper proposes a conflated inverse modeling framework that combines a predictive forward model with a diffusion-based generative inverse model to generate diverse, physically plausible urban vegetation patterns capable of achieving specific regional temperature reduction goals, thereby addressing the underdetermined nature of the inverse problem where conventional methods fail to capture solution ambiguity.

Original authors: Baris Sarper Tezcan, Hrishikesh Viswanath, Rubab Saher, Daniel Aliaga

Published 2026-04-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Baris Sarper Tezcan, Hrishikesh Viswanath, Rubab Saher, Daniel Aliaga

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your city is a giant, steaming hot pan on a stove. The concrete and asphalt are the metal, and the sun is the burner. Sometimes, that pan gets so hot it's dangerous to touch. This is the "Urban Heat Island" effect.

For a long time, scientists have been good at predicting how hot the pan will get based on what's on it. They say, "If you add a patch of grass here and a tree there, the pan will cool down by 2 degrees." This is called a Forward Model. It's like a weather forecaster: "Here is the sky, here is the wind, here is the rain."

But city planners have a harder, more confusing question: "I need this specific neighborhood to cool down by exactly 2 degrees. What should I plant?"

This is the Inverse Problem. And here's the catch: There isn't just one answer. You could plant a dense forest in a small corner, or you could scatter trees all over the block, or you could put a big park in the middle. All three different layouts might cool the neighborhood by the exact same amount.

Traditional computer programs are bad at this. They are like a student who only knows one answer to a math problem. If you ask them for a solution, they give you the "average" of all possibilities—a blurry, boring mix of trees and grass that doesn't look like a real city.

The New Solution: The "Infinite Gardener"

The researchers at Purdue University built a new AI system they call Conflated Inverse Modeling. Think of it as a super-smart, creative gardener who can dream up many different ways to solve the same heat problem.

Here is how it works, using a simple analogy:

1. The Two-Part Brain

The system has two distinct parts working together:

The Predictor (The Thermometer): This is the "Forward Model." It's a strict teacher that knows the rules of physics. It looks at a map of trees and buildings and says, "Okay, if you plant this, the temperature will drop by 1.5 degrees." It never guesses; it calculates.
The Dreamer (The Artist): This is the "Inverse Model." It uses a technology called Diffusion. Imagine a sculptor starting with a block of marble that looks like static noise (like TV snow). The sculptor slowly chips away the noise, guided by the "Thermometer," until a beautiful statue emerges.

2. The Secret Sauce: "Blurring" the Goal

Here is the clever trick. If you tell the "Dreamer" exactly where the temperature needs to change down to the pixel, it gets stuck and only makes one boring solution.

So, the researchers told the Dreamer to look at the temperature goal through a frosted glass. They gave it a "coarse" or blurry target.

Analogy: Instead of saying, "Paint a red dot exactly at coordinates (5,5)," they said, "Paint a red area somewhere in this general neighborhood."
Result: Because the goal is slightly vague, the Dreamer is free to be creative! It can paint the red dot in the center, or the corner, or make it a swirl. As long as the overall neighborhood cools down, the Dreamer is happy.

3. The "Check" (The Physics Loss)

Every time the Dreamer creates a new, crazy garden layout, it passes it to the Thermometer.

Thermometer: "Hmm, this layout cools the area by 2.1 degrees. Close enough!"
Thermometer: "Wait, this one only cools it by 0.5 degrees. Try again!"
Thermometer: "This one cools it by 5 degrees. Too much!"

The system keeps generating new, diverse garden layouts until it finds ones that hit the target temperature perfectly.

Why This Matters

Imagine you are a city planner trying to fix a heatwave.

Old Way: The computer gives you one generic map: "Plant trees here." You build it, and it works, but it's the only option you have.
New Way: The computer hands you a menu of 100 different options.
- Option A: A big park in the center.
- Option B: A "green wall" of trees along the main street.
- Option C: Small community gardens scattered everywhere.
- All of them cool the city by exactly the same amount.

This is huge because it gives planners choice. They can pick the option that fits the local culture, the budget, or the existing buildings, knowing that any of them will solve the heat problem.

The Bottom Line

This paper introduces a tool that stops computers from being boring and repetitive. It teaches them that in the real world, there is rarely just one right answer. By combining a strict physics calculator with a creative, "dreamy" AI, they can generate endless, realistic, and diverse ways to cool down our cities, even when we don't have perfect data to start with.

It's like giving city planners a magic wand that says, "Here are a thousand different ways to make your city cooler. Pick your favorite."

1. Problem Statement

Urban areas face increasing thermal extremes due to rapid urbanization and climate change. While forward models (predicting Land Surface Temperature (LST) from vegetation and urban form) are well-established, the inverse problem remains largely unexplored: determining specific spatial vegetation configurations (NDVI maps) that achieve a desired regional temperature shift.

Key Challenges:

Underdetermined Nature: Multiple distinct spatial vegetation patterns can yield similar aggregated temperature responses. A one-to-many mapping exists where a single temperature goal corresponds to many valid vegetation layouts.
Limitations of Current Methods: Conventional regression and deterministic neural networks fail to capture this ambiguity. They tend to converge to "averaged" solutions, producing low-diversity outputs that lack the variability required for realistic urban planning.
Data Scarcity: It is impossible to observe the same urban area under multiple vegetation scenarios and temperature-modification goals simultaneously, making supervised learning of diverse solutions difficult.

2. Methodology

The authors propose a Conflated Inverse Modeling Framework that combines a predictive forward model with a diffusion-based generative inverse model. This approach explicitly encourages both diversity (spatially distinct patterns) and specificity (meeting a target temperature change).

A. Data and Inputs

Inputs: Building Height (BH) maps (proxy for urban morphology) and coarsened Land Surface Temperature (LST) maps.
Target: NDVI (Normalized Difference Vegetation Index) maps representing vegetation density.
Dataset: 20 US cities, split into $3.84 \times 3.84$ km tiles (128x128 pixels at 30m resolution).

B. The Forward Model (Predictor)

Architecture: A U-Net with a ResNet-50 encoder.
Function: Predicts temperature change ( $\Delta T$ ) given an NDVI map and Building Height map.
Role: Used as a "physics" supervisor during the training of the inverse model to ensure generated vegetation patterns result in the correct thermal outcomes. It is trained separately and frozen during inverse model training.

C. The Inverse Model (Generator)

Architecture: A diffusion-based generative model (Score-SDE with NCSN++ backbone, utilizing the EDM formulation).
Conditioning:
- Building Height (BH): Used as a spatial constraint.
- Coarsened LST: Instead of fine-resolution LST (which leads to deterministic, over-constrained outputs), the LST is downsampled and upsampled. This removes fine-scale cues, allowing the model freedom to generate diverse spatial patterns while adhering to the coarse regional temperature structure.
Inpainting Strategy: The model generates NDVI only within a specific Region of Interest (ROI) (e.g., $32 \times 32$ pixels), preserving the surrounding context.

D. Training Strategy: The "Conflated" Loss

The training objective combines standard diffusion loss with a physics-based constraint to enforce temperature consistency at an aggregated scale:

Diffusion Loss ( $\mathcal{L}_{diff}$ ): Standard weighted denoising loss (EDM formulation).
Physics Loss ( $\mathcal{L}_{phys}$ ): An $\ell_1$ $ℓ_{1}$ penalty applied to the pooled mean temperature of the generated patch.
- The forward model predicts the temperature of the generated NDVI.
- This prediction is compared against the ground-truth mean temperature of that region.
- Crucial Design: The constraint is applied at a coarse regional scale (via pooling), not pixel-by-pixel. This prevents the model from overfitting to specific pixel values, thereby preserving spatial diversity while ensuring the aggregate thermal goal is met.
Loss Schedule: The physics weight ( $\lambda_{phys}$ ) follows a warmup and linear ramp schedule to stabilize training, allowing the model to learn basic denoising before enforcing thermal consistency.

E. Inference

Users specify a desired temperature change ( $\Delta_{target}$ ) for a specific ROI.
The diffusion model synthesizes diverse NDVI patterns within the ROI using masked inpainting.
The magnitude of the temperature change is modulated by a gain parameter ( $w$ ).

3. Key Contributions

Conflated Framework: Introduction of a system that achieves the dual goals of diversity (generating multiple valid solutions) and specificity (meeting exact temperature targets) in an underdetermined inverse problem.
Generative Formulation: Formulating vegetation-driven temperature modulation as a generative inverse problem using NDVI, BH, and LST, moving beyond deterministic regression.
Coarse Conditioning Insight: Demonstrating that conditioning diffusion models on coarsened temperature maps (rather than fine-resolution) is essential for generating diverse spatial patterns, while aggregated physics loss ensures the temperature goals are met.
Scalable Tool: A method applicable to data-scarce settings that produces physically plausible vegetation configurations for urban climate adaptation.

4. Results and Evaluation

The method was evaluated on 20 US cities (41,715 sq. km) using four metrics:

Temperature Control Error (CtrlErr): Accuracy of achieving the target temperature change.
Baseline Consistency (BaseErr): Realism of the generated vegetation relative to observed temperatures.
Diversity: Measured by the inverse of Structural Similarity (SSIM) between generated samples.
Surrogate Calibration: Error bound of the forward model itself.

Key Findings:

Performance: The proposed method reduced specificity error (CtrlErr) by 37% and increased diversity by 3.4 times compared to baseline methods (e.g., standard U-Net, fine-resolution conditioning).
Ablation Studies:
- Fine LST Conditioning: Led to low diversity and poor controllability (over-constrained).
- Coarse LST Only: Increased diversity but weakened temperature alignment.
- Coarse + Physics Loss: Achieved the best trade-off, yielding the lowest control error while maintaining high diversity.
Visual Results: The model successfully generated distinct NDVI patterns (e.g., different tree cluster arrangements) for the same city tile that all resulted in the same target temperature reduction (e.g., -2°C).

5. Significance and Future Work

Significance:
This work bridges the gap between forward temperature prediction and generative urban design. It provides planners with a tool to explore multiple feasible solutions for heat mitigation rather than a single deterministic answer. This is critical for urban planning where constraints (infrastructure, aesthetics, cost) may favor one vegetation pattern over another, even if they achieve the same thermal goal.

Limitations & Future Directions:

Constraints: Currently considers only building height. Future work aims to incorporate spatial masks and rule-based constraints (e.g., avoiding roads or protected areas) directly into the diffusion process.
Data Sources: The study relies on Landsat 8. Future work could explore cross-sensor evaluation (e.g., integrating Sentinel-2 for higher resolution or MODIS for broader thermal coverage) to improve generalizability.

In summary, the paper presents a robust, physics-informed generative framework that solves the "one-to-many" inverse problem in urban climate adaptation, enabling the design of diverse, effective, and realistic urban greening strategies.