New Machine Learning System Classifies Roof Types to Assess Wind Risk

Roof type is very important when looking at how buildings can be hurt by strong winds, especially in areas that often face hurricanes. Unfortunately, this information is usually missing in public databases, making it hard to assess risk accurately. To address this issue, a new system has been created that uses Machine Learning to automatically classify the types of roofs on buildings by analyzing satellite images. This helps to create detailed Data about roofs that can be used to better assess wind risk in different regions.

#Importance of Roof Type

In regions prone to hurricanes in the United States, houses typically have gable or hip roofs. Studies have shown that these types of roofs respond differently to high winds, which makes knowing the roof type critical for accurate risk assessments. However, as mentioned, roof type data is often missing from available building information databases.

#Using Machine Learning for Roof Classification

Recent efforts have utilized machine learning to fill in the gaps for missing roof data. One approach involves predicting Roof Types based on other building features like construction year and building value. Another advanced method utilizes satellite images to classify roofs using Convolutional Neural Networks (CNNs). CNNs automatically learn key features from images, which provides a more reliable way to classify roofs compared to traditional methods that depend on pre-set features.

#Roof Classification Framework

This framework involves several steps. First, satellite images of buildings are collected. The CNN model is then applied to classify these roofs, while also identifying low-quality images that cannot provide clear information about the roof types. For missing roof data, algorithms are used to predict roof types using important building characteristics and the roof types of nearby buildings.

#Case Study Areas

The new system was tested in two locations: New Hanover County in North Carolina and Miami-Dade County in Florida. Over 161,000 single-family houses were analyzed. The roof type data collected from these areas revealed significant differences in the distribution of roof types. For instance, the dominance of gable or hip roofs varied greatly from one census tract to another.

#Types of Roofs

The study defined five classes of roofs to recognize differences in both roof type and shape. This includes simple gable roofs, complex gable roofs, simple hip roofs, and so on. The aim here is to provide a more detailed understanding of the building stock, particularly in hurricane-prone areas. Traditional models often oversimplify roof types, making them less accurate when predicting wind vulnerability.

#Data Collection and Training

To train the CNN model, a large set of images was collected to create a benchmark database. Care was taken to ensure that images were of high quality and only contained one complete roof. Image labels were generated manually to build a reliable training set. Different techniques were then used to enhance the dataset and prevent overfitting.

#Model Performance

The performance of the CNN model was evaluated using various test datasets. These datasets included images of buildings with pure gable or hip roofs, as well as those not clearly categorized. The model achieved high accuracy rates, particularly in distinguishing between roof types. Misclassifications occurred primarily for roofs with similar shapes but different complexities. Overall, the model effectively generated reliable roof type data.

#Roof-Type Distribution

The trained model was used to classify roofs across all buildings in the two study areas. It was found that valid roof type data could be obtained for more than 80% of the buildings. In New Hanover County, for example, the dominant roof type was found to be gable roofs, while in Miami-Dade County, the distribution was more varied. The analysis revealed that many houses had unique roof configurations, supporting the need for detailed data.

#Missing Roof Data

Some images were classified as unknown due to poor quality, so algorithms were developed to fill in these gaps. This is important for ensuring comprehensive data, which is vital for regional assessments. Features such as the year a building was built or its size were found to influence roof type classification.

#Neighborhood Influence

The roof type of buildings can also depend on their neighbors. Houses in similar neighborhoods tended to have similar roof styles. To improve predictions for missing data, the roof types of nearby buildings were taken into account. By analyzing the roof types in the vicinity, the system was able to estimate the type of roof for buildings with incomplete data.

#Algorithm Performance

Different algorithms were used to predict missing roof data. The Random Forest and Support Vector Machine models showed promise, especially when predicting the roof types. These algorithms were evaluated through cross-validation and testing datasets. While achieving good accuracy rates, it was noted that the performance could be affected by the quality of nearby data.

#Potential Applications

The framework created can be adapted for other regions as well, wherever satellite imagery is available. Its simplicity and effectiveness can be beneficial for a range of projects, including assessing solar energy potential. The automatic classification of roof types improves the process of building inventories, which can contribute to more accurate risk assessments.

#Conclusion

Overall, the automatic roof classification workflow represents a significant advancement in building assessments for wind risk. The detailed roof type datasets created through this framework allow for better understanding and quantification of spatial distributions. With the ability to incorporate variations in roof types, the new methods can lead to more accurate predictions and better risk modeling in hurricane-prone areas. These improvements will support enhanced planning and safety measures for communities at risk.