Geospatial Label Transfer for Reliable Remote Sensing Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The inefficiency and resource-intensiveness of acquiring large amounts of training data for machine learning models, particularly in remote sensing imaging, due to the labor-intensive process of labeling features and the limitations of conventional data augmentation techniques.
Innovation Solution
A method of natural augmentation that aligns and associates images with shared geographic location information to create training datasets, utilizing coregistered imagery to enhance label accuracy and efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If conventional data augmentation techniques (flips, rotations) are applied to training images, then additional training data can be generated, but the augmented data may not resemble real variations and can be unsuitable for training
Solution Approach 1:
The patent uses copying by acquiring multiple images of the same geographic location from different times or sensors and using these copies as augmented training data. This replaces conventional synthetic augmentations (flips, rotations) with real observational copies that naturally capture real variations in the data, thereby improving both quantity and reliability of training data simultaneously
2Reliability
If large amounts of training data are acquired to improve model performance, then model accuracy improves, but the process becomes time-consuming and resource-intensive
Solution Approach 1:
The patent performs preliminary action by pre-acquiring and storing multiple images of the same geographic location from different times or sensors before the training process begins. These pre-acquired images serve as ready-to-use augmented training data, eliminating the need for time-consuming conventional data acquisition and labeling processes during model development
Solution Approach 2:
The patent creates training data copies by acquiring multiple images of the same geographic location and using these copies as augmented training data. This copying approach generates additional training data without requiring new manual labeling, significantly reducing the time and resources needed while maintaining high model performance
3Measurement precision
If manual labeling of training data is performed to ensure accuracy, then label quality improves, but the process becomes labor-intensive and expensive
Solution Approach 1:
The patent uses copying by acquiring multiple images of the same geographic location and using these copies as augmented training data with transferred labels. This eliminates the need for manual relabeling of augmented data, maintaining label accuracy while dramatically improving labeling efficiency by reusing existing labels across multiple images
Solution Approach 2:
The patent recovers and reuses existing labels by transferring them from source images to augmented images. Instead of discarding labels after use or requiring manual relabeling, the system recovers and reapplies existing labels to multiple images, maintaining high label accuracy while reducing labor requirements
Data Source
Figure 1A
Figure 1B
Figure 2
AI summary
In some embodiments, a method for training a machine learning device includes: ascertaining geographic location information of at least one portion of a first image associated with a label; associating with the label a second image including at least a portion having substantially the same geographic location information as the at least one portion of the first image; optional alignment or coregistration of the first and second image to maximize mutual information overlap; forming a training dataset comprising the first and second images as input images and the label that the first and second images are associated with as outputs; optional binary categorization and curation of the resulting training dataset to ensure accuracy; and training the machine learning model using the augmented dataset.