AI Height Map Generation from Electro-Optic Imagery
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current geospatial modeling systems face challenges in accurately estimating height values from single images and efficiently generating detailed 3D maps, particularly with large-scale satellite data of inhomogeneous quality and resolution, often requiring expert-driven manual processes and struggling with image registry and processing complexity.
Innovation Solution
An AI system utilizing a processor and memory to store labeled electro-optic image classified objects in a semantic label database, training models with stochastic gradient descent and game theory optimization, and generating estimated height maps from new imagery, which includes convolutional neural networks for semantic segmentation and ensemble model optimization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional geospatial modeling systems use manual expert-driven processes for height estimation from single images, then measurement precision may be maintained through expert judgment, but productivity is significantly reduced due to time-consuming manual operations
Solution Approach 1:
The patent replaces manual expert-driven mechanical processes with an automated AI system comprising convolutional neural networks and game theory optimization algorithms. The system automatically estimates pixel heights from single electro-optic images through trained models, eliminating the need for manual expert intervention while maintaining accuracy through sophisticated computational methods including stochastic gradient descent and ensemble model optimization.
2Area of stationary object
If traditional systems process large-scale satellite data with inhomogeneous quality and resolution, then comprehensive coverage is achieved, but device complexity increases due to image registry and processing challenges
Solution Approach 1:
The patent transforms the processing approach by changing parameters from traditional image registry methods to AI-based direct height estimation from single images. The system accepts electro-optic images with varying qualities and resolutions as input and uses trained neural network models to directly predict height values, bypassing complex image matching and registration procedures while maintaining comprehensive geographic coverage.
3Extent of automation
If automated workflows are implemented for land-use land-cover classification, then extent of automation is improved, but measurement precision may deteriorate without expert validation
Solution Approach 1:
The patent implements feedback mechanisms through game theory optimization that evaluates multiple AI model predictions and selects the optimal result. The system uses reward matrices that incorporate classification confidence scores and cross-validates predictions across different neural network models, providing automated quality control that maintains measurement precision while achieving full workflow automation for land-use land-cover classification.
Data Source
AI summary
An artificial intelligence (AI) system for geospatial height estimation may include a memory and a processor cooperating therewith to store a plurality of labeled predicted electro-optic (EO) image classified objects having respective elevation values associated therewith in a semantic label database, and train a model using trained EO imagery and the semantic label database. The processor may further estimate height values within new EO imagery for a geographic area based upon the trained model, and generate an estimated height map for the geographic area from the estimated height values and output the estimated height map on a display.