Improving rice nitrogen stress diagnosis by denoising strips in hyperspectral images via deep learning

Abstract

It is critical to assess crop nitrogen status for precision fertilization and crop management. Hyperspectral imaging (HSI) has been proved as one of the valuable sensing technologies for plant nitrogen stress detection, while exploring efficient methods for hyperspectral image analysis remains challenging due to the high dimensionality, information redundancy, and noise interference. This study aims to develop a de-striping convolution neural network (DS-CNN) to remove strip noise in hyperspectral images and design a nitrogen diagnosis CNN (ND-CNN) for rice leaves. The DS-CNN included an image encoder, a strip dropout bottleneck, and a decoder to remove the strip noise in hyperspectral images. Activation function Leaky-ReLU was introduced to activate the potential neurons to keep the real text features of the hyperspectral image. Element-wise feature addition mechanism was also applied to guarantee the full range of band images that reached the best structural and textural similarities. We constructed six datasets with different noise scales to explore the potential of DS-CNN. The best performance of DS-CNN was on the lowest strip noise dataset ( σ = 0.02), with the mean squared error (MSE) lower than 2 × 10 −4 , highest structure similarity index metric (SSIM) of 0.99, and peak signal-to-noise ratio (PSNR) of around 36 dB on the validation dataset. Moreover, we compared the nitrogen diagnosis performance before and after denoising. The ND-CNN developed from the denoised dataset not only avoided the overfitting but also improved the accuracy of nitrogen stress diagnosis of rice leaves. • DS-CNN works well for hyperspectral image strip noise removal. • DS-CNN can be applied in real strip noise removal scenario. • DS-CNN can improve the performance of ND-CNN. • DS-CNN and ND-CNN can work as workflow for hyperspectral phenotyping.