Dynamic Material Classification for Remotely Sensed Imagery
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing automated systems for identifying materials in geospatial imagery face challenges due to variations in material composition by region, time, atmospheric conditions, and image acquisition characteristics, making it difficult to develop reliable and accurate material detection.
Innovation Solution
A dynamic classifier is trained on the same image it is applied to, using machine learning to account for regional and temporal variations, and a system leveraging sensor movement to identify materials by detecting moving vehicles to determine roadways made of concrete and asphalt.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a static classification model trained on reference spectral data is used for material identification, then the system structure is simple and easy to implement, but the classification accuracy deteriorates due to material composition variations by region and time
Solution Approach 1:
The patent transforms the static classification model into a dynamic one by automatically training a classification model on each input image using machine learning. This dynamic approach allows the system to adapt to regional and temporal variations in material composition, thereby improving classification accuracy while maintaining automated operation.
Solution Approach 2:
The system changes the parameters of the classification model by retraining it on each new image's spectral data. This parameter adaptation enables the model to account for variations in material composition across different regions and times, resolving the contradiction between simple structure and accurate identification.
2Productivity
If automated machine learning classification is used to identify materials in images, then manual review is eliminated and processing efficiency increases, but reliability deteriorates due to variances in material composition and atmospheric conditions
Solution Approach 1:
The system dynamically adjusts the classification model parameters by training on each image's actual spectral data, allowing it to adapt to atmospheric conditions and material variations specific to that image, thereby maintaining reliability while preserving automated processing efficiency.
Solution Approach 2:
The classification model performs self-training on each input image, automatically adapting to the specific conditions of that image without requiring manual intervention or external calibration, thus maintaining both automation and reliability.
3Measurement precision
If a dynamic classifier trained on each image is used to account for regional and temporal variations, then material identification accuracy is improved, but computational complexity and processing time increase
Solution Approach 1:
The system implements a dynamic classification approach where the model is automatically trained on each image, accepting the increased processing time as necessary to achieve accurate adaptation to regional and temporal variations in material composition.
Solution Approach 2:
The system performs preliminary automated training of the classification model on each image before actual material identification, preparing the model in advance to handle the specific conditions of that image, which justifies the additional processing time投入.
4Ease of operation
If automated systems use consistent reference spectral data for classification, then the classification process is simple and fast, but accuracy deteriorates due to atmospheric conditions and viewing geometry variations
Solution Approach 1:
The system replaces the static reference spectral data approach with a dynamic model training approach, where the classification model is retrained on each image to account for atmospheric conditions and viewing geometry variations, improving accuracy while maintaining automated operation.
Solution Approach 2:
Instead of using fixed reference spectral parameters, the system dynamically changes the classification model parameters by training on each image's actual spectral characteristics, allowing adaptation to varying atmospheric and geometric conditions.
Data Source
Figure 1
Figure 2
Figure 2A
AI summary
An automated system is provided for classifying materials in remotely-sensed imagery based on automated construction of a dynamic classifier - namely, a classifier that is automatically trained on the same image to which it is then subsequently applied. A first automated process identifies high confidence exemplars of each class using tailored classification techniques. This data is then used to train a supervised classification model (e.g., discriminant analysis), and the resultant classifier is applied to other pixels in the image that are unclassified or uncertain. Dynamic classification is automatically customized to the current image and can yield a more accurate and efficient material classification versus a static (image-independent) or manually trained classifier. It can overcome various confounding factors including inconsistencies in radiometric calibration, atmospheric conditions, and atmospheric distortions of ground spectra; different viewing and illumination geometries; and regional variations in the composition of certain materials like asphalt and concrete.