Topographic Feature Extraction Preprocessing via Simulated-Annealing Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for training machine learning systems to extract topographic features from 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, particularly for rare features.
Innovation Solution
A method involving simulated annealing to iteratively select and balance the distribution of image patches, ensuring that rare classes are prioritized and the training data is optimized to include a diverse range of geographic features, using data augmentation techniques to enhance the training dataset.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual inspection methods are used to create training data, then labeling accuracy can be ensured, but the process becomes time-consuming and expensive
Solution Approach 1:
The system uses automated algorithms to generate training data labels without human intervention. The machine learning model processes imagery and automatically assigns class labels to pixels, making the system self-sufficient in creating training datasets while maintaining consistency and scalability
Solution Approach 2:
Manual inspection processes are replaced with automated computational algorithms. The mechanical human labeling process is substituted with electronic image processing and automated classification systems that can handle large volumes of data rapidly
2Ease of manufacture
If arbitrary subsections of the area are selected for manual mapping, then data collection is simplified, but the distribution of feature classes becomes unbalanced with insufficient representation of rare features
Solution Approach 1:
The system dynamically adjusts the selection parameters of training patches based on class distribution requirements. By changing the selection criteria from arbitrary to purpose-driven (prioritizing rare classes), the system achieves balanced representation while maintaining automated selection processes
Solution Approach 2:
The system incorporates feedback loops that monitor the distribution of feature classes in the training dataset. Based on this feedback, the patch selection process is adjusted to ensure adequate representation of underrepresented classes, creating a self-correcting mechanism for balanced data collection
3Productivity
If the model is trained with unbalanced class distribution, then training speed is maintained, but the model's ability to recognize rare features deteriorates
Solution Approach 1:
The training data is structured with local quality variations where different regions (patches) are selectively chosen to provide appropriate class representation. Rare classes are intentionally oversampled in specific patches while common classes are balanced, creating localized quality adjustments that improve overall model performance without sacrificing training efficiency
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. Training data can then be selected from these labelled samples using a method based on simulated annealing, which iteratively searches through subsets of the available patches to identify a solution that is as close to the most optimal solution as possible.