Results and Discussions The classification model based on the first-order derivative combined with the AlexNet network can realize the fast and accurate classification and identification of this study s four aquatic plants. Compared with the VGG16 and CNN3 networks, this study s model has the highest test accuracy of 99.50%. The model s training and testing speeds are 13.56 s/epoch and 0.032 frame/s, respectively, which are 30.12 s/epoch and 0.016 frame/s lower than those of VGG16. Although the model s training speed is 8 s/epoch higher than that of CNN3 and the testing speed is 0.002 frame/s higher, the classification accuracy is 14.44 percentage points higher than that of the CNN3 model. To verify the model s classification accuracy under small samples, 40%, 60%, and 80% of the sample dataset were randomly selected as the training set. The lowest classification accuracy of the model was 99.15%, higher than the classification accuracy of the CNN3 and VGG16 models. The influences of spectral overlapping bands and background interference on the classification results were reduced using four spectral preprocessing methods to process the sample data, and the classification accuracy of the three models before and after preprocessing was compared. The first-order derivative method improved the classification accuracy. The first-order derivative combined with the AlexNet network has the highest classification accuracy of 99.50%. The Grad-CAM algorithm was used to visualize the established aquatic plant identification model, and the classification-sensitive bands of four aquatic plants were analyzed, including seven classification sensitive bands for Typha angustifolia L., two classification sensitive bands for Pontederia cordata L., eight classification sensitive bands for Hydrocotyle vulgaris, and five classification sensitive bands for Thalia dealbata.

Aquatic plants can purify pollutants and inhibit algae growth. Therefore, obtaining accurate information on the number and growth status of aquatic plant species helps monitor the aquatic ecological environment. Spectral analysis, as a vital method for aquatic plant identification, has the characteristics of noncontact, fast, and pollution-free. However, because they are affected by the surrounding water environment, the characteristic spectral peaks of green aquatic plants are more challenging to distinguish than terrestrial plants. The ground spectral data have high dimensions and numerous overlapping bands and background interferences, and the characteristic spectrum is not obvious. The data are more challenging, and a few ground spectral datasets are suitable for deep learning. Currently, conventional machine learning classification methods cannot accurately and comprehensively extract deep features on small samples, resulting in unsatisfactory final classification results. Therefore, the deep learning algorithm and hyperspectral data are used to classify aquatic plants for the problems of many overlapping spectral bands, background interference, inconspicuous characteristic peaks, and less self-built aquatic plant spectral sample data.

This study uses the first-order derivative method combined with the AlexNet network to classify and identify four nearshore aquatic plants. The classification accuracy and training speed of three convolutional neural networks (AlexNet, CNN3, and VGG16) were compared to verify the classification effect of our model on the nearshore aquatic plant spectrum and the AlexNet network was determined as the optimal network structure. Furthermore, the influence of the number of samples on different classification models was studied, and classification effect of three models under small samples was explored. The influence of spectral preprocessing on the model s classification effect was studied, and the sample data before and after preprocessing using four spectral preprocessing methods were compared. Finally, the Grad-CAM algorithm was used to study the classification model visually to extract the characteristic bands of four aquatic plants. The sensitive spectrum bands of nearshore aquatic plants were analyzed, extracted, and compared with the existing aquatic plant datasets. The results are compared to verify the effectiveness of the feature extraction of this study s model.

Typha angustifolia L., Pontederia cordata L., Hydrocotyle vulgaris, and Thalia dealbata. It provides an essential reference for classifying and identifying these four aquatic plants under hyperspectral remote sensing.