[Reprint] Application of Parallel Computing in Remote Sensing Images

Reprinted from: http://tech.it168.com/a2010/0722/1080/000001080895.sht

Among the many natural disasters, the earthquake caused the largest number of casualties and economic losses. In the past two years, Wenchuan, China. From the destruction of a small number of buildings to the destruction of urban centers or large national socio-economic infrastructure. However, at the current technical level, effective earthquake disaster prediction and early warning cannot be achieved, and only post-disaster rescue, resettlement of victims, post-disaster reconstruction and other means can be used to reduce the loss of life and property caused by disasters. The collapse of houses is an important indicator of the damage degree in urban areas. Its quantity and distribution can directly reflect the loss of life and property in the disaster area. Therefore, it is the most urgent need to quickly obtain information about collapse of houses in the affected area, so that emergency rescue can be delivered to critical locations within 72 hours after the earthquake.

1. project application background

Optical Images with a spatial resolution of less than 1 MB can accurately reflect the detailed information on the surface of collapsed houses. These images are the main data sources for obtaining information on collapsed houses after disasters and play an important role in disaster reduction and relief. Manual visual interpretation and change detection are the main methods for extracting collapsed houses through high-score images. However, the implementation efficiency of this method is difficult to meet the rapid information acquisition needs in the disaster emergency phase. At the same time, due to individual operator differences and other factors, it is easy to cause inconsistency in the interpretation results of many people, it brings difficulties to subsequent information analysis and decision-making. In recent years, through the combination of image processing and pattern recognition algorithm change detection methods, gradually applied to achieve high-precision and Rapid Extraction of collapsed housing information after disaster. However, as the spatial resolution of the remote sensing image data increases, the data volume also doubles, generally, high-resolution remote sensing data of a time phase that can reflect the details of city building damage must be above g level, or even reach T level. This puts forward higher requirements on the computer's ability to automatically process the data. 72-hour emergency response time after disaster is the need for rapid extraction of life information. It is vital to further improve the efficiency of computer processing of high-resolution remote sensing images.

This project uses the high-resolution aerial remote sensing images of the region obtained during the Yushu earthquake to develop a "rapid recognition system for remote sensing image house collapse" based on these images ", this will provide advantageous support for similar disaster recovery decisions in the future.

2. Application business bottlenecks and Scale

High-resolution remote sensing images generally refer to remote sensing images with a spatial resolution of less than 10 MB. Compared with traditional low-spatial-resolution satellite images, high-resolution satellite images significantly increase the data volume of a single image, narrow imaging spectral bands, and more obvious geometric structure and texture information. The emergence of high-resolution satellite remote sensing images makes it possible to observe the detailed changes of the earth surface at a small spatial scale, perform large-scale remote sensing drafting, and monitor the environmental impact of human activity, it has broad application prospects and is currently the main data for obtaining spatial information for earthquake disasters. Because the damage caused by earthquake disasters is mainly caused by urban buildings, and the goal of disaster relief is to help residents' personal safety, the resolution requirements for remote sensing data are high, which leads to a doubling of the data volume of a single image, the computing workload increases exponentially with the geometric base. This increases the workload of computer processing of image data, and the "Remote Sensing Image house collapse fast identification system" requires a large amount of computing time, or even dozens of hours of processing at a time. This is contrary to the goal of quickly extracting collapse information from data, which affects the role of the system in the second-to-second disaster emergency decision-making, it is easy to cause the software development project to fail without being accepted by users.

3. parallelization Method

Remote Sensing Image Data is a type of raster data, and its processing algorithm is based on repeated operations in the row and column units, system R & D uses NVIDIA Cuda architecture to realize GPU-CPU coupling parallel design to improve the budget efficiency of the system. In the GPU acceleration Algorithm Design for remote sensing image processing, this architecture is used to separate pixels or image blocks to map to parallel threads for processing, then, the processed pixels or image blocks are combined into the result image through multi-level storage.

The Cuda architecture regards GPU as a compute device that can execute many threads in parallel ). It is used as the coprocessor of the main CPU or host, that is, the computing intensive part of the data in the applications running on the host is detached to this device in parallel. Cuda maps computing tasks to a large number of parallel threads, including hardware Dynamic Scheduling and execution. More accurately, applications that run independently on different data can be stored on this device as functions executed by many different threads. To achieve this, You can compile such a function in the device's Instruction Set and load the obtained Program (called kernel, kernel) to the device. The kernel is organized in the form of a thread grid. Each thread grid consists of several thread blocks, and each thread block consists of several threads. Based on the raster feature of remote sensing images, its processing algorithm is very suitable for establishing parallel computing. The Calculation of one grid in the Data raster space can be mapped to one thread, and multiple threads can be considered as parallel computing of The raster data.

4. Implementation and results

The CPU (host thread) is used to read the computing parameters and image raster data before computing starts. The data is mapped to corresponding threads and copied to the video memory for calling during GPU computing, other parallel operations are completed on the GPU.

In addition to the gray value of pixels, texture of high-resolution remote sensing images is an important feature for describing images. texture segmentation based on wavelet transform is an ideal image segmentation algorithm. Daubechies1 wavelet is used in the system's Object Type Recognition operations. The number of layers of wavelet packet decomposition is set to m, and the two-dimensional wavelet packet decomposition of the m layer is performed respectively. For each sample image sub-block, 4 m sub-band images are obtained. For convenience, we decided to perform single-scale wavelet transformation. m was 1, and a total of four decomposed sub-band images were obtained. The mean absolute value of the four sub-bands of each image sub-block is used as a feature vector. Then, perform wavelet transformation on the classification image to extract the feature vectors of each pixel. The K-mean algorithm is used to classify remote sensing images, calculate the distance between each point and the initial center of the classification, determine the category, adjust the cluster center, and iterate. After the iteration, set the color of the pixel Based on the classification result, and finally calculate the ratio of collapsed houses to collapsed houses.

In GPU computing, the feature vectors of each pixel of the image are calculated for each thread in the thread block, and the pixels of the image and the values related to the wavelet transform are used by the thread, read from the global memory to the shared memory. After data is loaded by all threads in the same thread block (the data is synchronized within the thread block ), the average absolute value of the four sub-bands after the pixel wavelet transform calculated by the thread. In this way, the synchronous control completes the computing of each pixel wavelet transform in parallel, and finally writes the pixel feature vector information in the updated shared memory back to the corresponding global memory.

The system experiment uses the high-resolution aerial remote sensing image of the region obtained during the Yushu earthquake, 1. After quick processing by the system, 2 results are obtained.



Figure 1. Original Image (image from the Chinese Emy of sciences Earth observation center)



Figure 2. Result Image

When processing large-sized images, the total computing efficiency of the GPU can reach 19.2 times that of the CPU, and the core computing efficiency of the GPU can reach 38.3 times that of the CPU. Using the Cuda programming interface, it is worthwhile to re-design and compile the original image processing algorithm for the GPU to achieve dozens of times the efficiency improvement. In addition, considering the cost of GPU computing relative to CPU computing, high-performance computing capabilities for fast processing of high-resolution remote sensing images can be obtained with a relatively low GPU.