Impervious Surface Extraction via Deep Learning and CRF
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolutional neural networks struggle to effectively establish spatial relationships between terrains in remote sensing images with lower spatial resolutions, leading to fragmented or irrational classification results in impervious surface extraction.
Innovation Solution
A method and system utilizing joint deep learning and semantic probability, involving a deep convolutional network for feature extraction, pixel-by-pixel category prediction, and a conditional random field model for global optimization, to accurately extract impervious surfaces from high-resolution remote sensing images by normalizing image data and constructing a network with convolution, pooling, unpooling, and deconvolution layers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If a convolutional neural network is used for feature extraction and classification, then automated processing is achieved, but spatial relationship establishment between terrains deteriorates, leading to fragmented or irrational classification results
Solution Approach 1:
A conditional random field model is introduced as an intermediary between the deep convolutional network's pixel-level predictions and the final classification results. This mediator incorporates spatial relationship constraints and semantic probability information to refine the automated predictions, ensuring that spatial relationships between different terrains are properly established while maintaining automation.
2Measurement precision
If deep learning is applied to extract features from remote sensing images, then classification accuracy improves, but computational complexity increases
Solution Approach 1:
The deep learning pipeline is segmented into distinct functional components: a deep convolutional network for feature extraction, a conditional random field model for spatial relationship modeling, and a semantic probability framework for classification. This segmentation allows each component to be optimized independently and facilitates better computational resource allocation, managing overall complexity while maintaining high classification accuracy.
3Productivity
If image data is processed without normalization, then processing speed is maintained, but model learning effectiveness deteriorates due to singularity data
Solution Approach 1:
Image data normalization is performed as a preliminary action before feeding data into the deep convolutional network. This pre-processing step transforms the image data to have zero mean and unit variance, eliminating singularity data issues and improving model learning effectiveness. The normalization is computationally efficient and does not significantly impact processing speed.
Data Source
AI summary
A method of extraction of an impervious surface of a remote sensing image. The method includes: 1) obtaining a remote sensing image of a target region, performing normalization for image data, and dividing the normalized target region image into a sample image and a test image; 2) extracting an image feature of each sample image by constructing a deep convolutional network for feature extraction of the remote sensing image; 3) performing pixel-by-pixel category prediction for each sample image respectively; 4) constructing a loss function by using an error between a prediction value and a true value of the sample image and performing update training for network parameters of the deep convolutional network and network parameters relating to the category prediction; and 5) extracting an image feature from the test image through the deep convolutional network based on the training result obtained in 4).