Hyperspectral Training Data for AI Remote Sensing Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current AI algorithms for remote sensing data analysis face challenges in accuracy due to insufficient high-quality and quantity training data, particularly when transferring models across geographical regions or time, as a result of temporal and spatial de-correlation in land surface properties.
Innovation Solution
The method involves using unsupervised hyperspectral data analyses to derive training data for AI algorithms, which are then applied to multispectral or microwave images, improving the quality and quantity of training data and enabling more accurate classification and quantification of earth land surface properties.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If ground truth data is collected for AI training, then training data quality is improved, but temporal and spatial de-correlation reduces accuracy when transferring models to different regions or times
Solution Approach 1:
The patent introduces an intermediary approach by using physically-based retrieval algorithms as a mediator between ground truth measurements and AI training. Instead of directly using ground truth data, the system retrieves biophysical parameters through physical models that capture the underlying mechanisms, creating a bridge that improves both training quality and transferability to different regions and times.
Solution Approach 2:
The patent transforms the training data from direct ground truth measurements to physically-derived biophysical parameters. This parameter transformation allows the training data to capture fundamental physical relationships that remain consistent across different spatial and temporal contexts, thereby improving model transferability while maintaining training quality.
2Measurement precision
If more training data is collected to cover all land surface heterogeneities, then AI classification accuracy is improved, but data quantity and coverage requirements become unmanageable
Solution Approach 1:
The patent changes the nature of training data from extensive ground truth measurements to physically-derived biophysical parameters. This transformation reduces the quantity of training data needed because physical models generalize across different conditions, capturing fundamental relationships rather than requiring separate training for each specific land surface heterogeneity.
Solution Approach 2:
The physically-based retrieval algorithms serve multiple functions: they generate training data, validate AI predictions, and provide physically-consistent parameter estimates across diverse land surface conditions. This multi-functionality reduces the overall data quantity requirements while maintaining comprehensive coverage of land surface heterogeneities.
3Measurement precision
If ground truth measurements are taken frequently to capture temporal dynamics, then temporal accuracy is improved, but measurement cost and complexity increase significantly
Solution Approach 1:
The patent uses physically-based retrieval algorithms as intermediaries that can process remote sensing data without requiring frequent ground truth measurements. These physical models act as mediators that maintain temporal accuracy by enforcing physical consistency across time, reducing the need for complex and frequent field measurement campaigns.
Solution Approach 2:
The patent replaces the mechanical field measurement system with a physics-based computational system. Instead of physically measuring parameters at multiple time points, the system uses remote sensing data processed through physical models, substituting complex field measurements with computationally-efficient physical retrievals that maintain temporal accuracy.
Data Source
AI summary
The invention relates to a method of training Artificial Intelligence (AI) algorithms for remote sensing image analyses by hyperspectral earth land surface property data analyses, said method comprising the steps of a. selecting an earth land surface property to be analysed; b. subjecting the selected earth land property to unsupervised measurements of said property by using a hyperspectral image obtained from a defined area X of the earth land surface at a defined time Y by a remote sensing hyperspectral radiance measurement instrument; c. converting the obtained hyperspectral radiance measurement data into at least one set of data of said selected property using the appropriate unsupervised retrieval algorithm using physical based retrieval methods, wherein each property data set i is further assigned a geographic location Xi and time of observation Yi ; d. using the at least one property data set obtained in step (c) as training data for training the AI algorithm for multispectral images or microwave images of the defined area X; and e. applying the thus trained AI algorithm for the property to multispectral or microwave remote sensing data sets of a geographical area equal to, or larger than, the defined area X and for a time span equal to or larger than the defined time span Y. The invention also relates to a use of hyperspectral earth land surface property data for training Artificial Intelligence (AI) algorithms to be applied to data obtained from remote sensing instruments in order to improve the instruments' performance in retrieving said property data.