Object-level Hyperspectral Target Detection Based on Spectral-Spatial Features Integrated YOLOv4-Tiny Network

The spectral resolution and spatial resolution of hyperspectral remote sensing images are mutually limited. To keep the same signal-to-noise ratio, the spatial resolution will decrease when the spectral resolution improves. The targets in low-resolution hyperspectral image, such as airplanes, cars and ships, appear as several pixels or sub-pixels. Current hyperspectral target detection methods mainly focus on pixel-level targets, which process spectral information and simple neighbourhood-pixel-related information in a pixel-by-pixel detection strategy. The contribution of spatial features is limited, and it takes a long time to train and detect pixel-by-pixel. Inspired by the deep learning-based object detection technologies for RGB images, we designed a hyperspectral image target detection method based on spectral-spatial features integrated YOLOv4-tiny network (SS-YOLONet). The 3D hyperspectral images were directly sent to the detection network, their spectral information and complex spatial features were extracted by channel attention module, spatial attention module and 3D convolution. Considering the small size of targets such as airplanes, we extracted two shallow features for small-scale objects. In the experiment, we used the pansharpened EO-1 hyperspectral images to verify the effectiveness of the proposed algorithm.