Low Pixel Resolution Hyperspectral Image Mosaics Generation Using Learning-Based Feature Matching

Remote sensing has become a key component of precision agriculture in the recent decade. Hyperspectral imaging is one key technology that is predicted to be one of the primary decision-making tools in remote sensing-based precision agriculture. Unmanned Aerial Vehicle-based hyperspectral image acquisition is becoming viable due to the reduced cost and the form factor of recently introduced hyperspectral cameras. However, these advantages come at the cost of pixel resolution. Due to factors such as uniform textured surfaces in farmlands, low-pixel resolution, and repeated patterns, traditional stitching methods are unsuccessful at identifying matched features that are needed in generating mosaics from these images. Generating mosaics is a key step in the decision-making process since it opens the ability to interpret information field-wide instead of per-image information interpretations. This paper proposes an image mosaic generation pipeline based on LoFTR - a local feature matching method using transformers as a feature matcher for the low-pixel resolution hyperspectral images. Furthermore, the GPS point-based optimization method is also presented in order to minimize the computational cost and allow the multicore processing capability. The proposed method was evaluated using several field datasets obtained using a low-resolution hyperspectral camera and an unmanned aerial platform. Results present successfully stitched image cubes that could be used in future analysis tasks in agriculture-related decision-making processes.