Automatic Processing of Chinese GF-1 Wide Field of View Images

Abstract. The wide field of view (WFV) imaging instrument carried on the Chinese GF-1 satellite includes four cameras. Each camera has 200km swath-width that can acquire earth image at the same time and the observation can be repeated within only 4 days. This enables the applications of remote sensing imagery to advance from non-scheduled land-observation to periodically land-monitoring in the areas that use the images in such resolutions. This paper introduces an automatic data analysing and processing technique for the wide-swath images acquired by GF-1 satellite. Firstly, the images are validated by a self-adaptive Gaussian mixture model based cloud detection method to confirm whether they are qualified and suitable to be involved into the automatic processing workflow. Then the ground control points (GCPs) are quickly and automatically matched from the public geo-information products such as the rectified panchromatic images of Landsat-8. Before the geometric correction, the cloud detection results are also used to eliminate the invalid GCPs distributed in the cloud covered areas, which obviously reduces the ratio of blunders of GCPs. The geometric correction module not only rectifies the rational function models (RFMs), but also provides the self-calibration model and parameters for the non-linear distortion, and it is iteratively processed to detect blunders. The maximum geometric distortion in WFV image decreases from about 10-15 pixels to 1-2 pixels when compensated by self-calibration model. The processing experiments involve hundreds of WFV images of GF-1 satellite acquired from June to September 2013, which covers the whole mainland of China. All the processing work can be finished by one operator within 2 days on a desktop computer made up by a second-generation Intel Core-i7 CPU and a 4-solid-State-Disk array. The digital ortho maps (DOM) are automatically generated with 3 arc second Shuttle Radar Topography Mission (SRTM). The geometric accuracies of the generated DOM are 20m for camera-2 and 3, and 30m accuracy for camera-1 and 4. These products are now widely used in the fields of land and resource investigation, environment protection, and agricultural research.