Convolutional Neural Network Rasterization for Point Data Interpolation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for interpolating point observations into raster data, such as Kriging, often introduce linear dependencies and redundancies, limiting their accuracy and effectiveness in machine learning applications.
Innovation Solution
The use of convolutional neural networks to dilate point data and combine it with organic raster data, avoiding linear dependencies and redundancies by training convolutional filters on historical data to extrapolate raster values, thereby improving the interpolation process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If Kriging is used for interpolation, then the interpolated values are linearly unbiased and close to the most likely values, but linear dependencies and redundancies are introduced into the interpolated data
Solution Approach 1:
The patent replaces the traditional Kriging mechanical interpolation system with a neural network-based system. The neural network learns complex non-linear relationships from training data and generates interpolated values without the linear dependency constraints of Kriging, thereby eliminating redundancies while maintaining or improving interpolation accuracy.
Solution Approach 2:
The patent transforms the interpolation approach by changing from linear parameter-based interpolation (Kriging) to a non-linear parameter learning approach using neural networks. The network learns optimal parameters from data rather than relying on predefined linear covariance models, thus avoiding the introduction of linear dependencies and redundancies.
2Ease of manufacture
If simple interpolation methods (nearest neighbor, bilinear) are used, then the processing is computationally simple, but the interpolation accuracy is limited
Solution Approach 1:
The patent replaces simple mechanical interpolation algorithms with a neural network system that learns complex patterns from data. The neural network substitutes basic arithmetic operations with learned non-linear transformations, achieving high accuracy without the computational complexity of traditional advanced interpolation methods.
Solution Approach 2:
The patent performs preliminary training of the neural network on available data before actual interpolation. This preliminary learning phase enables the network to capture complex relationships in advance, allowing for accurate interpolation without requiring complex computational procedures during the actual interpolation process.
Data Source
AI summary
A computer produces predictions throughout a raster field in response to point data, by obtaining a partially empty matrix of point data, filling a matrix of extrapolated raster data by dilating the point data in a first convolutional neural network, and generating a matrix of aggregate raster data by combining the extrapolated raster data with organic raster data in a second convolutional neural network.