[Paper Review] Automatic Roof Type Classification Through Machine Learning for Regional Wind Risk Assessment

Shuochuan Meng, Mohammad Hesam Soleimani‐Babakamali+1 more|arXiv (Cornell University)|May 27, 2023

Wind and Air Flow Studies9 citations

TL;DR

The paper develops an automatic roof type classification framework using CNNs to generate high-resolution roof data from satellite images, with imputation to fill missing data for regional wind-risk analyses in two U.S. study areas.

ABSTRACT

Roof type is one of the most critical building characteristics for wind vulnerability modeling. It is also the most frequently missing building feature from publicly available databases. An automatic roof classification framework is developed herein to generate high-resolution roof-type data using machine learning. A Convolutional Neural Network (CNN) was trained to classify roof types using building-level satellite images. The model achieved an F1 score of 0.96 on predicting roof types for 1,000 test buildings. The CNN model was then used to predict roof types for 161,772 single-family houses in New Hanover County, NC, and Miami-Dade County, FL. The distribution of roof type in city and census tract scales was presented. A high variance was observed in the dominant roof type among census tracts. To improve the completeness of the roof-type data, imputation algorithms were developed to populate missing roof data due to low-quality images, using critical building attributes and neighborhood-level roof characteristics.

Motivation & Objective

• Define a five-class roof type system to reflect wind resistance (five roof types including simple gable, simple cross-gable, complex cross-gable, simple hip, and cross-hip).
• Develop an automatic workflow to classify roof types from building-level satellite images using CNNs and remove low-quality images.
• Impute missing roof-type data using building attributes and neighborhood roof characteristics to enhance dataset completeness.
• Apply the workflow to two study areas to produce city- and census tract–level roof-type distributions for regional wind risk assessment.

Proposed method

• Create a benchmark satellite image dataset of 11,834 building-level roof images labeled into six classes (including an unknown class).
• Fine-tune a pre-trained VGG-19 CNN (ImageNet) for roof-type classification with data augmentation and train/validation split; achieve 97% validation accuracy.
• Evaluate model on three testing datasets (A, B, C) representing pure gable/hip, unknown, and mixed roofs; report per-class precision, recall, and F1 scores.
• Classify 161,772 single-family houses in New Hanover County, NC and Miami-Dade County, FL using Google Maps satellite images and the trained CNN.
• Identify and remove low-quality images (unknown class) to avoid erroneous roof data, and map remaining predictions to five roof types.
• Develop imputation models (Random Forest and SVM) using building attributes and neighboring roof information with a 80 m neighborhood radius to fill missing data.

Fig. 1: Proposed roof type classification workflow.

Experimental results

Research questions

• RQ1 Can a CNN trained on building-level satellite images accurately classify five roof types relevant to wind vulnerability?
• RQ2 How well does the CNN perform across different geographies and roof complexities (including mixed roofs)?
• RQ3 To what extent can imputation using building attributes and neighborhood context recover missing roof-type data for regional risk analyses?
• RQ4 What are the spatial distributions of roof types at city and census tract scales in the study areas, and how much variance exists across neighborhoods?

Key findings

• The CNN (VGG-19) achieves an overall validation accuracy of 97% on the roof-type classification task.
• Testing datasets A and B yield overall F1 scores of 0.96 and 0.95, respectively, with high per-class precision/recall and 1.00/1.00 for the unknown class.
• Nearly 80% of buildings in the study areas received valid roof-type data (unknowns 11–17%), enabling city- and census tract–level distributions.
• Roof-type distributions show gable roofs dominate in both areas, with hip roofs more prevalent in New Hanover County than in Miami-Dade County; 72% (NH) and 63% (MD) have complex roofs.
• Imputation models achieve validation accuracies of 0.89 (NH) and 0.72 (MD) for predicting roof type (gable/hip) and 0.64–0.66 for predicting roof complexity, with Random Forest generally outperforming SVM on testing data.
• Neighborhood context is the most important feature for predicting roof type, followed by year built and building area.

Fig. 2: Location of study areas