Urban Heat Island Intensity Prediction with Multi-Stream CNN and Attention U-Net

Overview

This project develops a deep learning framework to predict Urban Heat Island Intensity (UHII) in Pune, India, using a multi-stream convolutional neural network (CNN) with attention-based fusion. By integrating three geospatial data modalities—Land Surface Temperature (LST), PM₂.₅ particulate matter, and high-resolution land-cover maps—our model achieves a Mean Squared Error (MSE) of 0.1104 on held-out test data, outperforming baseline models (MSE=0.185). The framework leverages Google Earth Engine (GEE) for scalable data processing and provides interpretable attention maps to guide urban planning and climate resilience strategies.

Key Features

• Multi-Modal Inputs: Processes Terra MODIS LST (MOD11A2), GLC_FCS30D land-cover, and satellite-derived PM₂.₅ datasets.
• Model Architecture: Coordinated encoders with CoordConv and CBAM attention, fused via an Attention U-Net for end-to-end UHII prediction.
• Humanitarian Impact: Identifies heat risk drivers (e.g., impervious surfaces, pollution) to inform urban greening and emission control policies.
• Interactive Pipeline: Jupyter notebook with widgets for temporal (2003–2020) and seasonal data exploration.

Dataset

• Sources:
  • MOD11A2 LST: 8-day, 1 km resolution [NASA LP DAAC, 2025].
  • GLC_FCS30D Land Cover: Annual, 30 m resolution, 2000–2022 [GEE Community Catalog, 2025].
  • PM₂.₅: Monthly, 0.01° resolution [GEE Community Catalog, 2025].
  • UHII Labels: Derived from urban-rural LST differences [GEE Community Catalog, 2025].
• Preprocessing:
  • Clipped to Pune’s administrative boundary.
  • Resampled to 1 km grid for alignment.
  • Split: 80% train, 10% validation, 10% test.

Model Architecture

• Coordinated Encoders:

  • LST/PM₂.₅: Three Conv–BN–ReLU stages with stride-2 downsampling, Kaiming initialization.

    

  • Land Cover: Embedding layer (37 classes → 8-dim), CoordConv, residual dilated block, CBAM attention.

    

• Attention U-Net as Feature Fusion Mechanism:

  • Concatenates encoder outputs, uses attention gates to reweight spatial features, and outputs full-resolution UHII via 1×1 convolution.

    

• Training Details: MSE loss, Adam optimizer, CosineAnnealingLR scheduler, trained over 2003–2020 data.

Results

• Quantitative:

  • Test MSE: 0.1104 (vs. baseline UNet: 0.185).

• Qualitative:

  • Attention maps highlight urban districts with high impervious surfaces and PM₂.₅ hotspots, aligning with known UHI patterns.

    

• Interpretability:

  • CBAM attention emphasizes pollution-heat synergies, with blue streaks in attention masks indicating high-influence regions (e.g., urban centers).

Humanitarian Implications

• Identifies UHI drivers (impervious surfaces, PM₂.₅) to guide urban greening, zoning, and emission policies.
• Supports equitable interventions in Pune, where 2024 data show heat stress in low-income areas, potentially reducing ~15% of heat-related vulnerabilities.

Citation

If you use this work, please cite:

Patil, A., Kalani, P., Lolage, O., Lanjewar, N. (2025). Learning Urban Heat Signatures via Coordinated Multi-Stream Convolution and Attention-Based Fusion. Vishwakarma Institute of Technology.