Digital Surface Model Resolution Enhancement via Image Regression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Radar-derived digital surface models (DSMs) suffer from noise introduction during processing, leading to reduced vertical accuracy and obscured spatial features, with conventional filtering methods lowering resolution and obscuring details.
Innovation Solution
The method enhances DSM resolution by mapping grayscale changes from images to elevation changes using a regression over a local neighborhood of pixels, independent of radar illumination geometry, and applicable to any sensor technology, allowing for the restoration of high-resolution details without reconstructing surface normal vectors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If filtering is applied to reduce noise in DSM, then noise level is reduced, but DSM resolution is reduced
Solution Approach 1:
The method segments the DSM enhancement process into multiple processing stages: initial filtering at native resolution, super-resolution upsampling to intermediate resolutions, and iterative detail transfer from source images. This segmentation allows noise reduction to be applied selectively at different resolution levels without permanently losing fine details.
Solution Approach 2:
The method transitions from 2D image space to 3D elevation space by generating multiple DSMs at different resolutions and iteratively transferring details. The super-resolution process creates intermediate resolution layers that bridge the gap between filtered low-resolution DSMs and high-resolution source images, enabling detail recovery without direct 2D filtering.
2Measurement precision
If filtering is applied to reduce noise in DSM, then noise level is reduced, but spatial features are obscured
Solution Approach 1:
The method performs preliminary filtering at the native resolution level before upsampling, establishing a clean base DSM. Then, during the iterative super-resolution process, spatial features are progressively recovered by transferring details from the original high-resolution images, ensuring that noise reduction does not permanently eliminate important spatial information.
Solution Approach 2:
Multiple intermediate-resolution DSMs serve as mediators between the filtered low-resolution DSM and the high-resolution source image. These intermediate layers allow gradual detail transfer and feature recovery, preventing direct loss of spatial features that would occur with aggressive filtering at native resolution.
3Measurement precision
If DSM is generated at lower resolution to reduce noise, then noise level is reduced, but resolution is reduced by 4-8 times
Solution Approach 1:
The method employs dynamic resolution adjustment through iterative super-resolution processing. Instead of generating a single static low-resolution DSM, the system dynamically creates multiple DSMs at progressively higher resolutions (e.g., 2x, 4x, 8x upsampling factors), adapting the resolution level to balance noise reduction and detail preservation based on the specific characteristics of the input data.
Solution Approach 2:
The method changes key processing parameters including upsampling factor, filtering strength, and iteration count during the super-resolution process. By dynamically adjusting these parameters across multiple processing stages, the system recovers resolution lost during initial filtering while maintaining vertical accuracy, achieving final DSMs with resolution comparable to or exceeding the original source images.
Data Source
AI summary
Systems and methods of enhancing the resolution or restoring details associated with high resolution images into a filtered digital surface model (DSM) for location-based applications and analyses. The disclosed methods include mapping the changing gray scale values (intensity) from the images to changes in elevation in the DSM using a regression over a local neighborhood of pixels. Further, the disclosed methods do not rely on information about the sensor illumination geometry, and are extendable to be able to utilize any types of images. Additionally, the disclosed embodiments are sensor agnostic. That is, the disclosed methods can be applied on any type of images collected by any type of sensor.