Bathymetry Feature Segmentation for Accurate Underwater Mapping
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing geospatial data processing techniques, particularly for underwater topography, face challenges in accurately identifying and mapping features of interest due to insufficient or biased training datasets, leading to faulty or insufficient automated mapping using machine learning systems.
Innovation Solution
A segmentation machine learning network is employed to generate segmentation masks for bathymetry data, utilizing convolutional neural networks and deep learning architectures to improve feature detection in bathymetry waveforms, enabling high-resolution mapping and classification of underwater features.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If machine learning systems are used for automated mapping of bathymetry data, then processing efficiency is improved, but accuracy deteriorates due to insufficient or biased training datasets
Solution Approach 1:
The patent applies preliminary action by performing manual annotation and expert review before final automated mapping. Trained operators manually label features of interest in bathymetry data, creating high-quality training datasets that improve subsequent machine learning accuracy. This preliminary manual processing ensures accurate feature identification before automation takes over.
Solution Approach 2:
The patent implements feedback mechanisms where machine learning system outputs are reviewed and corrected by human operators. The system continuously learns from these corrections, with annotated results fed back into the training process to improve future predictions. This closed-loop feedback resolves the accuracy issue while maintaining automated processing efficiency.
2Measurement precision
If manual feature identification is used, then accuracy is improved, but processing time increases
Solution Approach 1:
The patent applies segmentation by dividing the processing workflow into distinct phases: initial automated processing for broad coverage, followed by targeted manual review of specific regions or uncertain detections. This segmented approach allows most data to be processed quickly by machines while human expertise is applied only where needed, balancing speed and accuracy.
Solution Approach 2:
The patent uses partial action by applying manual review selectively rather than to all data. Human operators focus their effort on specific portions of the dataset where machine learning confidence is low or where features are particularly critical, rather than manually processing everything. This partial manual intervention maintains accuracy for critical features while preserving overall processing efficiency.
3Productivity
If automated mapping is implemented, then productivity is improved, but reliability deteriorates due to insufficient training data
Solution Approach 1:
The patent ensures reliable training data through preliminary manual annotation efforts before deploying automated mapping. Expert operators create comprehensive labeled datasets that cover various seabed conditions and feature types, establishing a solid foundation for reliable machine learning operation at scale.
Solution Approach 2:
The patent builds reliability through continuous feedback loops where automated mapping results are validated and corrected by human experts. These corrections are fed back into the training system, progressively improving model reliability while maintaining high productivity. The system becomes increasingly trustworthy through iterative learning from real-world results.
Data Source
AI summary
The present disclosure is directed to systems and techniques for processing frames of data. For example, a method can include obtaining a plurality of geospatial data inputs, each geospatial data input of the plurality of geospatial data inputs associated with a sample time and a surveyed area; generating a plurality of features corresponding to each geospatial data input of the plurality of geospatial data inputs; and generating, using a segmentation machine learning network, one or more segmentation masks for the plurality of geospatial data inputs.