Look for tiny rosettes from drone imagery

In this paper, we present our recent research work on applying the Generative adversarial networks(GAN)-based super-resolution method to upscale low-resolution airborne RGB images for precise weed detection. Often better detection accuracy requires high-resolution(HR) imagery for remote weed detection using drones. However, HR imagery means narrow ground coverage per survey flight. So this paper is to find the balance between the image resolution and the remote detection accuracy for better cost-efficiency. Our proposed super-resolution model is well trained by selected training samples of low-resolution(LR) and HR image pairs. Our model shows better performance in improving the quality of high-resolution remote sensing images and achieving accurate detection. In addition, we reviewed several super-resolution techniques, including interpolation, neighbor embedding, sparse coding, random forest, and CNN-based methods. Hence, this paper presents a cost-effective use of low-resolution airborne color images paired with a super-resolution approach to detect weeds in complex landscapes.