Rooftop Polygon Extraction Under Tree Cover Occlusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models struggle to accurately extract building rooftop polygons from geospatial imagery due to partial occlusion by tree cover, leading to geometric distortions.
Innovation Solution
Implement data augmentation techniques by simulating tree cover occlusions in training data to enhance the model's ability to handle partial occlusions, using simulated tree cover representations to augment geospatial imagery and train the model to better recognize building rooftops despite occlusions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If standard machine learning models are used for building rooftop polygon extraction, then the extraction process is simple and fast, but the accuracy deteriorates when rooftops are partially occluded by tree cover
Solution Approach 1:
The patent applies preliminary action by pre-augmenting the training dataset with simulated tree cover occlusions before model training. This allows the model to learn occlusion handling in advance, improving its accuracy when extracting building rooftop polygons from imagery containing partial occlusions.
Solution Approach 2:
The patent changes the training data parameters by introducing synthetic occlusion layers with varying transparency, position, and coverage areas. This transforms the training process to expose the model to diverse occlusion scenarios, enhancing its adaptability to real-world occluded conditions.
2Reliability
If data augmentation with simulated tree cover is implemented, then the model's ability to handle occlusions improves, but the training data processing complexity increases
Solution Approach 1:
The patent uses copying by creating synthetic copies of original imagery with superimposed tree cover masks. This approach generates diverse training samples without requiring additional real-world occluded imagery, improving model reliability while keeping the augmentation process manageable through algorithmic replication.
Solution Approach 2:
The patent introduces an intermediary occlusion layer that acts as a mediator between the original imagery and the training process. This layer simulates tree cover with adjustable parameters, allowing controlled augmentation that improves robustness without excessive complexity.
3Measurement precision
If occlusion handling is prioritized in model training, then extraction accuracy under occlusion improves, but processing time for training increases
Solution Approach 1:
The patent applies partial action by selectively augmenting only the portions of imagery that contain or are near building rooftops with tree cover simulations. This focused approach maintains extraction accuracy while reducing the overall processing time compared to augmenting entire images or datasets.
Data Source
AI summary
Example systems and methods for data augmentation for occlusion handling in feature extraction are provided. An example method involves accessing a tree cover source image that depicts one or more areas of tree cover, extracting a tree cover map from the tree cover source image that labels areas of the tree cover source image as either areas covered by tree cover or not covered by tree cover, selecting a contiguous area of tree cover from the tree cover map to use as simulated tree cover, accessing a geospatial image depicting a building, placing the simulated tree cover in the geospatial image to generate a tree cover-augmented geospatial image, wherein the simulated tree cover partially occludes a rooftop of the building as depicted in the geospatial image, and training a machine learning model, using the tree cover-augmented geospatial image, to extract building rooftop polygons outlining buildings as depicted in geospatial imagery.