Detection Of Cotton Waterlogging Stress Based On Hyperspectral Images And Convolutional Neural Network

Waterlogging in the early stage of cotton will reduce the number of bolls and do harm to yield. Early detection of waterlogging will help farmers to adjust cotton management and save the loss. To evaluate the application of deep learning for the detection of early waterlogging, this study applied a convolutional neural network (CNN) to classify different durations of waterlogging stress (0, 2, 4, 6, 8, 10 d) based on hyperspectral images (HSIs) of cotton leaves. An experiment was designed to simulate the situation of cotton under waterlogging stress and collect HSIs of visible and near-infrared (VNIR 450-950 nm) spectra with 126 bands 66 d after cotton sowing (66 DAS). It was found the spectral curve reflectance of waterlogging cotton was higher than that of non-waterlogging cotton. Especially near 550 nm and 750 nm, and the spectral curve increased with durations of waterlogging stress and there were 'blue shift' phenomena for the position of the red edge of the spectra. The first principal components of HSIs after band randomly discarding and principal component analysis (PCA) were used to build a dataset. GoogLeNet Inception-v3 (GLNI-v3) and VGG-16 models were selected to detect cotton waterlogging stress with the dataset. The results showed that the average time for a round training for GLNI-v3 was 13.337 s, with a classification accuracy of 96.95% and a loss value of 0.09431. The average time for a round training for VGG-16 was 21.470 s, with a classification accuracy of 97.00% and a loss value of 0.17912. Though these two models had similar classification accuracy and loss value, GLNI- v3 achieved a high accuracy with fewer training iterations. The durations of waterlogging stress of cotton in a short-term can be detected by HSIs of cotton leaves and CNN models are suitable for the classification of HSIs, and this method can provide support for cotton yield estimation and loss assessment after waterlogging.