---
title: 'GSM-UTCI: City-Scale Urban Microclimate Model'
url: https://www.emergentmind.com/topics/gsm-utci
type: topic
---

# GSM-UTCI: City-Scale Urban Microclimate Model

Searching arXiv for the specified paper to ground the article in the source paper.
GSM-UTCI is a multimodal deep learning model for **Geometric–Semantic–Meteorological prediction of the Universal Thermal Climate Index**. It maps 1 m–resolution spatial data—normalized Digital Surface Model (nDSM) and 1 m land cover—together with hourly meteorological conditions to a 1 m resolution map of **daytime-average UTCI** for 08:00–19:00. The framework is trained in supervised fashion on SOLWEIG-derived UTCI fields and is designed as a city-scale surrogate for computationally intensive urban microclimate simulation, with the specific objective of retaining near-physical fidelity while making scenario-based planning workflows operational at full-city extent [2507.23000].

## 1. Definition and problem setting

GSM-UTCI addresses a specific bottleneck in urban heat-stress modeling: the mismatch between the physical realism of high-resolution microclimate models and the computational demands of applying them iteratively over an entire city. Traditional models such as SOLWEIG and ENVI-met can compute mean radiant temperature $T_{mrt}$ and UTCI at high spatial resolution, but the reported constraints are substantial: they are computationally expensive for city-wide domains, require complex 3D and meteorological input preparation, and are too slow for iterative “what if I change the land cover here?” workflows [2507.23000].

The target variable is **daytime-average UTCI**, rather than a single instantaneous thermal field. In the formulation used here, UTCI functions as a biometeorological index integrating air temperature, humidity, wind speed, and $T_{mrt}$ into a single “feels-like” temperature linked to heat stress categories. The focus on the 08:00–19:00 average reflects sustained exposure relevant to health, labor, outdoor activities, transit, and equity analyses. The choice of **1 m resolution** is equally central: it resolves street canyons, tree shade, parks versus parking lots, vacant impervious parcels, and other forms of intra-urban heterogeneity that are invisible at 30–1000 m satellite land-surface-temperature scales. This scale therefore supports parcel- and block-scale interventions such as tree planting and pavement change.

Formally, the model maps spatial inputs and meteorological inputs to a single raster target:
- spatial inputs: nDSM and land cover at 1 m,
- meteorological input: $12$ hours $\times 18$ variables for a typical hot summer day,
- output: a 1 m map of daytime-average UTCI.

A plausible implication is that GSM-UTCI is best understood not merely as a predictor, but as a **surrogate planning model**: it compresses the behavior of a physical simulation pipeline into an inference-time engine suitable for repeated counterfactual evaluation.

## 2. Data model, preprocessing, and physical training labels

The study area is the entire city of Philadelphia, Pennsylvania, USA, selected to include dense central business district morphology, rowhouse neighborhoods, parks, industrial zones, and waterfronts [2507.23000]. The semantic spatial input is a 1 m land use/land cover raster from Pennsylvania Spatial Data Access, derived from high-resolution aerial imagery and LiDAR with semi-automatic classification. The classes are tree canopy, grass, bare earth, water, buildings, roads, and impervious surfaces. These are encoded as integers $0$–$6$ and rescaled to $[0,1]$ by dividing by $6$, yielding a one-channel raster
$$
\mathbf{X}_{lc} \in \mathbb{R}^{1 \times H \times W}.
$$

The geometric input is a single-band nDSM,
$$
\mathbf{X}_{dsm} \in \mathbb{R}^{1 \times H \times W},
$$
constructed from USGS 3D Elevation Program LiDAR point clouds using PDAL to derive DEM and DSM, then combined with city building footprints and tree canopy layers to obtain heights above ground for vegetation and buildings. This representation is intended to capture the 3D structure most relevant to shading and radiative exposure.

Meteorological input is drawn from the National Solar Radiation Database for August 2020 and organized as a multivariate hourly time series over 08:00–19:00:
$$
\mathbf{X}_{met} \in \mathbb{R}^{T \times N}, \quad T=12,\; N=18.
$$
The variables include air temperature, relative humidity, Global Horizontal Irradiance, Direct Normal Irradiance, and Diffuse Horizontal Irradiance, together with additional atmospheric and radiation parameters not enumerated in full in the provided summary.

For training, the city is split into $512 \times 512$ pixel tiles at 1 m resolution, each covering approximately $0.26\ \text{km}^2$, for a total of $12{,}642$ tiles. nDSM and all meteorological variables are standardized by z-score,
$$
\hat{x} = \frac{x - \mu}{\sigma},
$$
with $\mu$ and $\sigma$ computed from the training subset, while land-cover codes are scaled to $[0,1]$.

The target labels are generated using **SOLWEIG v1.x** as part of UMEP. SOLWEIG is used to compute hourly $T_{mrt}$ and then UTCI for every 1 m pixel using the land-cover map, building height model, tree canopy height model, and hourly meteorological forcing. The shortwave and longwave radiative exchange is summarized through the radiant flux expression
$$
R_{str} = \zeta_k \sum_{i=1}^{6} K_i F_i + \varepsilon_p \sum_{i=1}^{6} L_i F_i,
$$
where $K_i$ and $L_i$ are directional shortwave and longwave fluxes, $F_i$ are view factors, $\zeta_k = 0.70$, and $\varepsilon_p = 0.97$. UTCI is then computed via the official polynomial approximation after deriving vapor pressure from relative humidity and air temperature. Hourly UTCI fields are produced for all days in August 2020 and averaged over 08:00–19:00 to form a single “typical August daytime UTCI” map per pixel. This map serves both as the supervised target $Y_{UTCI}$ and as the baseline for scenario comparison.

## 3. Network architecture and multimodal fusion

GSM-UTCI is a **three-branch multimodal neural network** composed of a geometric encoder, a semantic encoder, a meteorological encoder, and a FiLM-based fusion mechanism, followed by a convolutional prediction head [2507.23000]. The architectural division corresponds closely to the physical determinants of outdoor heat stress: urban form, surface material/cover, and atmospheric forcing.

The **geometric encoder** ingests the nDSM. A stem convolution first maps the one-channel raster to three channels to match ImageNet-pretrained weights:
$$
\tilde{\mathbf{X}}_{dsm} = f_{stem}(\mathbf{X}_{dsm}), \quad
\tilde{\mathbf{X}}_{dsm} \in \mathbb{R}^{3 \times H \times W}.
$$
A Vision Transformer backbone, specifically `vit_tiny_patch16_224`, then tokenizes the raster into patches and processes them with multi-head self-attention. The coarse output feature map
$$
\mathbf{F}_{geo} \in \mathbb{R}^{C' \times H' \times W'}
$$
is projected to 16 channels using a $1 \times 1$ convolution and upsampled back to full resolution to produce
$$
\mathbf{Z}_{geo} \in \mathbb{R}^{16 \times H \times W}.
$$

The **semantic encoder** ingests the land-cover raster. After an analogous stem convolution to three channels, an `hrnet_w18` backbone extracts multi-resolution semantic features with repeated cross-scale exchange. The resulting feature hierarchy is upsampled to a common grid, concatenated, reduced through a $1 \times 1$ convolution, and upsampled to yield
$$
\mathbf{Z}_{sem} \in \mathbb{R}^{16 \times H \times W}.
$$
This branch is intended to encode fine-scale surface distinctions such as trees, grass, water, roads, buildings, and non-road impervious surfaces.

The **meteorological encoder** is a bidirectional LSTM operating on the hourly meteorological sequence:
$$
\mathbf{Z}_{met} = f_{BiLSTM}(\mathbf{X}_{met}) \in \mathbb{R}^{T \times d}, \quad d=64.
$$
The text characterizes $\mathbf{Z}_{met}$ as a temporal feature matrix summarizing meteorological dynamics. Dense layers then generate the FiLM parameters, although their exact internal specification is not given.

The key fusion mechanism is **Feature-wise Linear Modulation (FiLM)**. Meteorological context is converted into channelwise scales and shifts for both spatial branches:
- $\gamma_{geo}, \gamma_{sem} \in \mathbb{R}^{C \times 1 \times 1}$,
- $\beta_{geo}, \beta_{sem} \in \mathbb{R}^{C \times 1 \times 1}$.

For a spatial feature map $\mathbf{Z}$, FiLM applies
$$
\tilde{\mathbf{Z}}_{c,h,w} = \gamma_c(\mathbf{Z}_{met}) \cdot \mathbf{Z}_{c,h,w} + \beta_c(\mathbf{Z}_{met}).
$$
The modulated geometric and semantic features are then concatenated and convolved:
$$
\tilde{\mathbf{Z}}_{geo} = \gamma_{geo} \odot \mathbf{Z}_{geo} + \beta_{geo},
$$
$$
\tilde{\mathbf{Z}}_{sem} = \gamma_{sem} \odot \mathbf{Z}_{sem} + \beta_{sem},
$$
$$
\mathbf{Z}_{fused} = Conv\Big(Concat(\tilde{\mathbf{Z}}_{geo}, \tilde{\mathbf{Z}}_{sem})\Big).
$$
A small convolutional head then outputs the 1-channel UTCI prediction:
$$
\hat{\mathbf{Y}}_{UTCI} \in \mathbb{R}^{1 \times H \times W}.
$$

Training minimizes mean squared error,
$$
\mathcal{L}_{MSE} = \frac{1}{n} \sum_{i=1}^n (y_i - \hat{y}_i)^2,
$$
using AdamW with learning rate $1 \times 10^{-3}$, weight decay $1 \times 10^{-4}$, batch size $24$, and $1000$ epochs. The ViT and HRNet backbones are initialized with ImageNet weights. No explicit dropout is reported.

## 4. Evaluation protocol, accuracy, and computational performance

The experimental setup uses the $12{,}642$ Philadelphia tiles split randomly across the city into **70\% training** and **30\% testing**; the text notes that the held-out portion is also referred to as “validation” in one place, but conceptually serves as the evaluation set [2507.23000]. Implementation is in PyTorch on hardware consisting of **2 × NVIDIA RTX A6000 GPUs (48 GB each)** and **Dual Intel Xeon Gold 6258R CPUs, 112 logical cores**.

Performance is reported with MAE, MSE, MAPE, and $R^2$:
$$
\text{MAE} = \frac{1}{n}\sum |y_i - \hat{y}_i|,
$$
$$
\text{MAPE} = \frac{100\%}{n} \sum \left|\frac{y_i - \hat{y}_i}{y_i}\right|,
$$
$$
R^2 = 1 - \frac{\sum (y_i - \hat{y}_i)^2}{\sum (y_i - \bar{y})^2}.
$$

For the full GSM-UTCI model on the 12,642 validation tiles, the reported values are:
- parameters: **16,798,103**,
- **MAE: 0.4130 °C**,
- **MSE: 0.4477 °C²**,
- **MAPE: 1.3750\%**,
- **$R^2$: 0.9151**.

The study explicitly characterizes this performance as **near-physical** relative to SOLWEIG. Since $\sqrt{0.4477} \approx 0.67^\circ$C, the reported MSE also implies an RMSE of approximately $0.67^\circ$C.

Ablation and comparison models clarify the relative importance of the modalities and the fusion strategy:

| Model | Key composition | Key result |
|---|---|---|
| A1 | ViT + BiLSTM (FiLM) | MAE 0.7394 °C, $R^2 = 0.7690$ |
| A2 | HRNet + BiLSTM (FiLM) | MAE 0.4406 °C, $R^2 = 0.8992$ |
| A3 | ViT + HRNet + BiLSTM (Concat, no FiLM) | MAE 0.4435 °C, $R^2 = 0.9046$ |
| GSM-UTCI | ViT + HRNet + BiLSTM with FiLM fusion | MAE 0.4130 °C, $R^2 = 0.9151$ |

These results support three specific interpretations already stated in the source summary. First, land cover is more informative than morphology alone, as the HRNet-only spatial branch markedly outperforms the ViT-only branch. Second, adding morphology to land cover and meteorology helps even without FiLM. Third, FiLM-based conditioning produces the best result, improving $R^2$ from $0.9046$ to $0.9151$ and reducing MAE from approximately $0.44^\circ$C to $0.41^\circ$C.

Qualitatively, the model preserves sharp gradients between tree-shaded and sun-exposed pavements, building shadows, and small parks relative to adjacent streets. Lower residuals are reported where tree canopy and vegetated areas are well captured in land cover and height, and in built-up areas with straightforward morphology. Higher residuals can occur at complex intersections, mixed land covers, or sharp height transitions, although no class-stratified error statistics are provided.

Runtime is the principal operational advantage. The GPU-accelerated SOLWEIG + UTCI pipeline still requires **tens of minutes up to hours** for hourly 1 m UTCI over Philadelphia for August, whereas GSM-UTCI produces a **city-wide 1 m UTCI map in under five minutes** on the A6000 GPUs. This reduction by orders of magnitude is the basis for its use as a scenario engine.

## 5. Scenario simulation and urban climate planning use

After training, GSM-UTCI is used to simulate systematic landscape transformation scenarios under fixed meteorological conditions, so that the effect of land cover and morphology change can be isolated [2507.23000]. The workflow is explicit. Starting from baseline land cover and baseline nDSM, the model selects target class pixels, converts those pixels’ land-cover class to tree canopy, adjusts nDSM height at those locations either to the mean tree canopy height in the tile or, where no local trees exist, to the city-wide mean tree height, keeps meteorology fixed, and predicts the new UTCI field. The difference from baseline gives $\Delta UTCI$.

Three city-wide greening scenarios are examined:
1. **Bare earth $\rightarrow$ tree canopy**
2. **Grass $\rightarrow$ tree canopy**
3. **Impervious surfaces (roads & other impervious) $\rightarrow$ tree canopy**

The summary results are as follows:

| Scenario | Area (km²) | Avg $\Delta$UTCI (°C) |
|---|---:|---:|
| Bare Earth $\rightarrow$ Tree | 23.15 | -4.87 |
| Grass $\rightarrow$ Tree | 281.09 | -2.90 |
| Impervious $\rightarrow$ Tree | 270.66 | -4.18 |

The source additionally reports standard deviation, post-UTCI, and total $\Delta$UTCI in K·m²:
- Bare Earth $\rightarrow$ Tree: SD $1.32$, Post-UTCI $27.41^\circ$C, Total $\Delta$UTCI $112.83$M K·m²
- Grass $\rightarrow$ Tree: SD $1.58$, Post-UTCI $27.52^\circ$C, Total $\Delta$UTCI $815.95$M K·m²
- Impervious $\rightarrow$ Tree: SD $1.89$, Post-UTCI $27.75^\circ$C, Total $\Delta$UTCI $1{,}132.21$M K·m²

The quantitative interpretation given in the source is precise. **Bare earth to tree** produces the strongest local cooling per pixel, at **$-4.87^\circ$C**, but over a relatively small area. **Impervious to tree** produces the largest city-wide aggregate benefit, with mean cooling of **$-4.18^\circ$C across 270.66 km²** and a total cooling potential of **1,132.21 million K·m²**. **Grass to tree** covers the largest area, **281.09 km²**, but produces a smaller marginal cooling effect, **$-2.90^\circ$C**, because grass is already somewhat cooling relative to hardscape.

At census-tract level, the maps show that cooling potential is spatially heterogeneous and strongly aligned with land-cover composition. Tracts with higher impervious fractions show larger cooling under impervious-to-tree conversion; bare-earth conversion benefits are more scattered but strong where bare soil is concentrated, such as industrial or vacant areas; and peripheral or park-dominated tracts show smaller marginal effects. The framework is therefore positioned as a tract- and neighborhood-scale decision support tool for prioritizing greening interventions.

The planning applications described in the source extend from baseline heat-exposure mapping to scenario evaluation, prioritization, and GIS integration. Outputs can be aggregated to census tracts, blocks, pedestrian networks, or transit stops, and combined with socioeconomic vulnerability, health outcomes, ridership, zoning, parcel data, transportation networks, urban forest master plans, heat emergency response planning, and climate adaptation plans. A plausible implication is that the model’s practical significance lies less in single-run prediction than in repeated, comparative assessment of alternative intervention portfolios.

## 6. Relation to prior methods, limitations, and prospective generalization

Relative to physical urban climate models, GSM-UTCI is presented as a surrogate rather than a replacement. It matches SOLWEIG closely in the reported Philadelphia setting, but its predictive target is explicitly derived from SOLWEIG, and any biases or simplifications in SOLWEIG propagate into the learned model [2507.23000]. This is the central caveat against interpreting the network as an independently validated physical simulator.

The paper situates GSM-UTCI against several existing machine-learning lines of work: CNNs for UTCI microclimate, multilayer neural networks for $T_{mrt}$ around buildings, the global 1 km monthly UTCI dataset GloUTCI-M, WebMRT for rapid $T_{mrt}$ estimation, and broader ML literature on land surface temperature. The limitations attributed to prior work are that they often predict **LST rather than UTCI or $T_{mrt}$**, frequently use **single-modality** input without explicit 3D morphology, focus on **instantaneous** conditions rather than **daytime-average** thermal indices, and rarely support **forward landscape change simulations** at hyperlocal scales. On that basis, the paper identifies four novel contributions: multimodal FiLM fusion of 3D morphology, semantic land cover, and meteorological time series; hyperlocal 1 m UTCI prediction at full-city scale; a scenario-ready framework based on modified land-cover and nDSM inputs; and a scalable planning tool.

The principal limitations are explicit. The representation of urban form is static, so the model does not dynamically account for seasonal foliage changes, diurnal movement of shadows, or transient anthropogenic heat sources. The network does not encode radiative physics explicitly; rather, the ViT and CNN components learn correlations that may stand in for shadow geometry, multiple reflections, or spectral properties only insofar as these are implicitly represented in land-cover categories and the SOLWEIG training labels. Training is restricted to a **single city**, Philadelphia, and to **August 2020** meteorological conditions. Transfer to other climates, sun angles, or urban forms is therefore **not validated**. Likewise, robustness to extreme or future heat conditions outside the August 2020 meteorological domain remains an open question, even if FiLM theoretically provides a mechanism for adapting spatial feature weighting to unusual atmospheric states.

The authors nevertheless argue that the architecture itself is general because its required inputs—nDSM, land cover, and hourly meteorology—are standard in many urban contexts. The suggested adaptation pathway for another city is to produce a local 1 m nDSM and land-cover dataset, acquire representative meteorological forcing, run SOLWEIG or an equivalent high-fidelity physical model for a subset of conditions, and train GSM-UTCI from scratch or via transfer learning, initializing the ViT and HRNet with Philadelphia-trained weights before fine-tuning. This suggests that the framework is intended as a transferable template for **multi-city and multi-climate** expansion, but that such transfer remains an empirical task rather than an established result.

In practical terms, GSM-UTCI denotes a city-scale, hyperlocal, scenario-capable surrogate for computationally intensive microclimate modeling. Its defining combination is the use of nDSM for geometry, land cover for semantics, hourly meteorology for atmospheric forcing, and FiLM to condition spatial features on diurnal weather context. Its reported significance lies in making full-city 1 m UTCI mapping and iterative greening scenario analysis feasible in minutes rather than hours while preserving near-physical agreement with the underlying SOLWEIG simulation target.

Source: https://www.emergentmind.com/topics/gsm-utci