SAR Building Damage Classification via Environmental Mediators
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge in using Synthetic Aperture Radar (SAR) for damage assessment during natural disasters is the lack of pre-event data, leading to delayed and inaccurate detection of building damage due to the reliance on post-event data comparison.
Innovation Solution
A method using machine learning to classify buildings as damaged or undamaged in SAR images based on statistical characteristics of signal values, such as amplitude, phase, and environmental factors, without requiring pre-event data, by training a model on post-event SAR data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If pre-event data is used for damage detection, then measurement precision is improved, but loss of time increases due to delayed detection
Solution Approach 1:
The system performs preliminary actions by collecting and storing SAR image data and environmental information before disasters occur. The machine learning model is pre-trained on this prepared data, enabling rapid damage assessment after events without requiring pre-event baseline comparisons, thus resolving the time-delay issue while maintaining accuracy
Solution Approach 2:
The patent introduces environmental information (terrain, climate, geographic data) as an intermediary element that enables the machine learning model to understand and account for regional variations. This mediator allows accurate damage detection without pre-event data by providing context about what normal conditions look like for each location
2Loss of time
If pre-event data is not used, then loss of time is reduced for timely detection, but measurement precision deteriorates due to lack of comparison baseline
Solution Approach 1:
Environmental information acts as an intermediary that compensates for the absence of pre-event data. The machine learning model uses terrain, climate, and geographic data to establish what normal conditions should look like, enabling accurate damage assessment of post-event images without temporal comparisons
Solution Approach 2:
The system creates a digital copy or representation of normal conditions through environmental information and pre-collected data. This copied information serves as a reference baseline that the machine learning model compares against post-event images to detect damage, eliminating the need for actual pre-event satellite images
3Device complexity
If only post-event data is used, then device complexity is reduced by eliminating pre-event data requirements, but reliability worsens due to insufficient data for accurate classification
Solution Approach 1:
The machine learning model is designed with multi-functionality to process and integrate multiple data types: SAR image data, environmental information, and geographic data. This universal model can reliably classify buildings as damaged or undamaged using only post-event SAR images combined with environmental context, eliminating the need for separate pre-event data acquisition systems
Solution Approach 2:
Environmental information and geographic data are collected and prepared in advance as preliminary actions. This pre-prepared contextual data is stored and integrated with post-event SAR images during processing, providing the reliability needed for accurate classification without requiring pre-event satellite imagery
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
Enables timely and accurate classification of damaged buildings using only post-event SAR data, improving the reliability and efficiency of damage assessment.
Implementation Method 1
The movement of the antenna during imaging forms a 'synthetic' aperture that can provide similar resolution to much larger antennas, by exploiting the change in frequency of received signals changes due to the Doppler effect
Data Source
Figure 1
Figure 2~3
Figure 4
AI summary
A method of classifying a building as a damaged building or an undamaged building in a synthetic aperture radar "SAR" image using machine learning, the method comprising: receiving SAR image data relating to an area of interest including a building to be classified; determining physical information relating to one or both of the area of interest and the SAR image data acquisition environment; and classifying, using a trained machine learning model, whether the building is a damaged building or an undamaged building, wherein: the trained machine learning model has been trained on a training dataset comprising training SAR image data relating to damaged buildings; and input features of the training SAR image data comprise physical information relating to one or both of areas including the damaged buildings and the data acquisition environment of the training SAR image data.