Machine-Learning-Enhanced Procedural Modeling for 4D Historical Cities Reconstruction

Question

1. What strategies are employed to overcome data incompleteness?

2. How do we create a diachronic vector dataset from historical maps?

3. How is polygon orientation determined?

4. What parameters are used for procedural modeling of buildings in LOD1?

Accepted Answer

In Section 2.3, the researchers describe their strategy to overcome data incompleteness. This step is crucial to ensure that all necessary parameters for procedural modeling are available and accurately documented. Among the strategies proposed, a machine learning approach is used to fill gaps in partially present fields in the dataset. This approach allows the user to choose target fields and their related predictors, enabling greater customization in the prediction process. The techniques employed to infer missing information are discussed, along with the significance and performance of the statistical data completion process. The researchers emphasize the dependency of quantitative results on the available data and compare the performances of algorithms on their dataset, avoiding comparisons with different case studies. This approach ensures that the generated data is accurate and reliable for further analysis and modeling.

Accepted Answer

To create a diachronic vector dataset from historical maps, we employ semantic segmentation to automatically extract building footprints from 10 historical maps published between 1838 and 1947. We vectorize the footprints as high-quality vector data by designing an algorithm specifically adapted for 3D modeling. The 10 maps are combined in a diachronic vector dataset by detecting the first and last appearance of the polygons. Additional details, such as the number of floors, construction materials, and roof types, are incorporated from a secondary source. The maps are first georeferenced manually using GIS software, and then semantically segmented to obtain a diachronic dataset that captures the temporal information regarding the initial and final occurrences of building footprints. The building footprints are obtained from two specific maps: one depicting the Old City in detail (scale 1:2500, from 1947) and one depicting its surroundings (scale 1:5000, from 1938). The semantic segmentation aims to extract the footprints or built-up areas by recognizing the contours on the historical maps. A total of 252 patches are manually annotated from the 10 maps, and 132 patches are selected from the Historical City Maps Semantic Segmentation Dataset. A convolutional neural network is trained for each task, using the dhSegment framework and a simple UNet architecture with a ResNet101 encoder. This step creates two binary masks: the first corresponding to building contours, and the second to built-up areas. The vectorization of 2D geometries is a demanding part of the 3D model creation process, requiring the simplification of geometries to avoid aliasing and the ability to parameterize the level of simplification without affecting local coherence. The vectorization algorithm comprises several steps, including georeferencing, semantic segmentation, and vectorization, which are illustrated in Figure 1 and accompanied by Python code.

Accepted Answer

Polygon orientation is determined using Equation (2), which calculates the orientation of each polygon. The cycles can be reoriented clockwise or counterclockwise based on this calculation. The orientation convention follows the shapefile format, where 'donut holes' are oriented counterclockwise, and the outer cycle is oriented clockwise. This ensures consistency and accuracy in the vector data representation.

Accepted Answer

For LOD1, a set of three parameters is used to define the height of the extrusion: Height (H), Floor Height (H f ), and Number of Floors (N f ). The relationship between these parameters ensures that the extrusion process accurately represents the specified number of floors, maintaining the proportional scaling of the building's vertical dimensions. This parameter interdependence becomes particularly valuable in situations where explicit height information is missing from the dataset, but the number of floors is known. In situations where the height or number of floors is unknown, a predictive method is employed to estimate the number of floors and/or floor height, thereby enabling the calculation of the total height. This approach generates credible height values, even when the dataset lacks specific height information.

Accepted Answer

To address missing numerical parameters, we sample the value from a statistical distribution, either informed or based on educated guesses. The choice of the range depends on the case study and available information. By default, we assume a normal distribution centered around a parameterizable mean value. However, the specific distribution pattern may vary depending on the context. Future statistical tests on the normality of parameter distributions should be conducted. When culturally related datasets are available, we provide the ability to sample values from the real distribution. For example, the distribution of building height in Hamburg is bimodal, so a normal approximation may not be suitable. In this case, we provide the ability to fill missing values by performing a random choice with replacement from any provided distribution. When the missing parameter is categorical, the random choice is performed using uniform probability between all categories, but users can provide custom probabilities. When a parameter is partially present, we propose leveraging knowledge available for other buildings in the city to infer the missing values using machine learning models like Random Forest. This requires selecting relevant explanatory variables for prediction. For instance, in the Jerusalem case study, we focused on the number of floors and roof type, which were only partially available. Four different machine learning classifiers were implemented and compared in Section 2.4.

Accepted Answer

Decision Tree, Random Forest, and Adaptive Boosting models offer several advantages for filling gaps in data. Firstly, these models can handle missing values in the input, which is crucial when dealing with incomplete historical data. Instead of simply excluding incomplete rows or filling missing values with mean values, these algorithms natively manage missing values, providing more accurate predictions. Secondly, these models are interpretable, allowing visualization of the decision-making process and understanding how the model reaches its predictions. This transparency helps validate the model's predictions and provides insights into the dataset. Additionally, these models are trained relatively quickly and energy-efficient, making them suitable for dynamically retraining the model when new information is added to the dataset. Overall, Decision Tree, Random Forest, and Adaptive Boosting models are effective in handling missing values, providing interpretability, and ensuring efficient training for filling gaps in data.

Accepted Answer

Random Forest appears to be the most adequate method for predicting roof type and number of floors. It outperforms the naive solution and shows improvement in accuracy. The approach is robust to small datasets and can handle rarefied data effectively. By progressively decreasing the number of available data points, the model's effectiveness is assessed, and it rapidly exceeds the baseline accuracy even with limited data. The results stabilize when reaching 1% of the dataset, or around 50 training examples for the number of floors. A few more samples are necessary for predicting the type of roof.

Accepted Answer

The proposed model handles missing data by transforming a 2D dataset into a 3D CityJSON representation, making the pipeline robust to missing data. This approach ensures the production of a high-LOD model without being hindered by common incompleteness in the dataset. The resulting model is not limited to geometric information but also includes informative attributes, allowing for tracking of parameters, their origin, and assumptions made. The CityJSON extension enables the maintenance of initial attributes mapped from the original geodata, providing a comprehensive and informative model.

Accepted Answer

Random Forest offers several advantages for data completion. Firstly, it can handle missing input values, allowing for training the model even with incomplete data. Secondly, it provides visualization of decision trees, aiding in understanding the model's decision-making process. Additionally, Random Forest can automatically fill up datasets by identifying hidden patterns between attributes, especially in sparse datasets. Users can also refine results by adding new information or confirming inferences. The generic nature of Random Forest enables users to select target features and predictors, making it adaptable to dataset growth. Compared to traditional methods, Random Forest offers flexibility in incorporating new information, making it suitable for 3D reconstruction and incremental model growth without data loss. Versioning in CityJSON files further enhances collaborative usage of 3D models.

Accepted Answer

The methodology addressed challenges related to data incompleteness, cultural specificity, iterative nature of scientific projects, and subjectivity of reconstruction and interpretation. It introduced a comprehensive framework with open-source tools for transforming 2D GIS datasets into 3D CityJSON representations. Machine learning techniques were integrated to complete missing values, and the CityJSON extension ensured transparency and traceability. The approach allowed for dynamic and iterative model adaptation, addressing evolving data and maintaining accuracy and relevance.

Machine-Learning-Enhanced Procedural Modeling for 4D Historical Cities Reconstruction

Chat with Paper

AI Agents for this Paper

Most frequently asked questions

1. What strategies are employed to overcome data incompleteness?

2. How do we create a diachronic vector dataset from historical maps?

3. How is polygon orientation determined?

4. What parameters are used for procedural modeling of buildings in LOD1?

5. How to address missing numerical parameters?

6. What are the advantages of using Decision Tree, Random Forest, or Adaptive Boosting models for filling gaps in data?

7. How does Random Forest perform in predicting roof type and number of floors?

8. How does the proposed model handle missing data?

9. What are the advantages of using Random Forest for data completion?

10. What challenges did the methodology address?

Citations

Artificial Intelligence for Digital Heritage Innovation: Setting up a R&D Agenda for Europe

Digitizing Karachi's Decades-Old Cadastral Maps: Leveraging Unsupervised Machine Learning and GEOBIA for Digitization

Integration Method for Generating Complete Hierarchical Layouts From Incomplete Virtual Scene Trees

A Framework for Optimizing Open Spatial Data in Urban Planning and Policy Applications

PyTorchGeoNodes: Enabling Differentiable Shape Programs for 3D Shape Reconstruction

References

Deep Residual Learning for Image Recognition

Random Forests

U-Net: Convolutional Networks for Biomedical Image Segmentation

Scikit-learn: Machine Learning in Python

Greedy function approximation: A gradient boosting machine.

Related Papers (5)

PYTHON 2.6 Reference Manual

Usability of Software On-line Documentation: A User Study

Current issues in assessing and improving documentation usability

Usability Aspects of MP3 Player Documentation. Documentation Usability Heuristics Revised – a Case Study

Ease of use and the richness of documentation adequacy