Abstract: Objectives: The Wide-swath Thermal Infrared Imager (WTI) of the Gaofen-5A satellite (GF5A) can obtain ultra-width images with the width of 1500km and the spatial resolution of 100m by performing whiskbroom imaging. During the imaging process, it undergoes "large-angle side-view imaging, imaging-vertical ground imaging, and largeangle side-view imaging" in sequence, resulting in different observation angles and imaging object distances during different scanning rows, with a large range of variation, making the original image characterized by variable resolution and large distortion. This makes traditional correction models and image stitching methods difficult to apply to GF5AWTI processing. Methods: To address these issues, this paper proposes an object-space geometric stitching method for ultra-width whiskbroom sequence images. First, for the complex light transmission during the imaging process, an accurate mapping from image-plane coordinates to object-space coordinates is achieved by constructing a physical strict imaging model(PSM) for the line-array wiskbroom system. Then, an navigation lookup table (NLT) is constructed to determine the accurate correspondence between corrected image points and original image points. Finally, for the stitching of sequence images, an object-space stitching virtual plane model is established, where the spatial range, spatial resolution of the sub-satellite point, and starting geographic spatial coordinates of the original images are calculated frame by frame. Multiple frames of original images are then directly projected into the virtual stitching plane through grayscale resampling. Results: After the object-space geometric stitching process, the effective-area pixels of the output image are all covered by grayscale values, eliminating the issues of variable resolution and large distortion in the images. Simultaneously, seamless stitching is achieved in the complex overlapping areas between frames. Conclusions: The stitching accuracy of GF5A-WTI image products using the method in this paper is better than 0.6 pixels, and the internal geometric accuracy is better than 1.5 pixels, meeting the needs of high-quality large-area mapping and successfully applied to the GF5A-WTI operational processing system.