Aerial Imagery Preprocessing for Balanced Topographic Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for training machine learning systems to extract topographic features from aerial imagery are inefficient and costly due to the need for laborious manual inspection and the challenge of achieving a balanced distribution of training data across different classes, leading to biased and imbalanced models.
Innovation Solution
A method for pre-processing image data to generate balanced training data by measuring the distribution of classes within each image, normalizing these measurements, and using simulated annealing to select patches that prioritize less common features, ensuring an optimal distribution of training samples.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual inspection is used to create training data, then data accuracy is improved, but time consumption and cost increase significantly
Solution Approach 1:
The system performs preliminary automated processing to generate initial training data and labels before human annotation is applied. The automated system pre-processes imagery to identify potential features and create candidate labels, which then serve as a foundation for human annotators to refine and verify, significantly reducing the total time required while maintaining accuracy.
Solution Approach 2:
The system uses automated machine learning algorithms to perform self-labeling of training data. The model makes initial classifications of topographic features based on learned patterns from existing data, generating its own training labels without requiring manual annotation for every feature, thereby reducing time consumption while maintaining data quality.
2Loss of time
If arbitrary subsection of area is selected for manual mapping, then time consumption is reduced, but class distribution balance deteriorates
Solution Approach 1:
The system applies different sampling strategies to different geographic areas and feature types. Instead of uniform random sampling, the system identifies regions with specific topographic features and prioritizes their inclusion in the training data. This ensures that rare but important features are adequately represented while efficiently processing the majority areas, achieving both time efficiency and class distribution balance.
Solution Approach 2:
The system dynamically adjusts sampling parameters based on the distribution of topographic features in the imagery. It calculates feature frequency metrics and modifies the sampling probability for different classes accordingly. This parameter adaptation ensures that the training data achieves a balanced representation of all topographic features, including rare classes, while minimizing the total time required for data collection.
3Reliability
If more training data is collected to improve model accuracy, then model robustness is improved, but production cost increases
Solution Approach 1:
The system generates synthetic training data by copying and transforming existing annotated imagery. It creates augmented versions of training samples through geometric transformations, color adjustments, and composite generation, effectively multiplying the available training data without requiring additional manual annotation. This approach significantly increases the quantity of training data while controlling production costs.
Solution Approach 2:
The system applies parameter transformations to existing training data to create varied training examples. By adjusting parameters such as scale, rotation, color, and illumination, the system generates diverse training samples from a limited set of original images. This parameter-based data generation strategy increases model robustness by exposing the model to varied conditions while avoiding the high cost of collecting entirely new training data.
Data Source
AI summary
Computer-implemented methods and systems are provided for pre-processing image data to generate improved training data for training a machine learning system to automatically classify imagery. One or more images of a geographic region are captured and processed to obtain a plurality of labelled samples (also referred to as patches) for training the machine learning system. To do this, the amount that each class is contained within each area image relative to the whole dataset is measured, with the distribution of classes being weighted according to this measurement and normalised to determine how many patches per class and per area should be taken.