This study aims to identify underwater benthic animals in situ, use random forest algorithm to achieve recognition classification detection, classify and identify target organisms for analysis, dig deeper into the data, and improve efficiency and reliability of decision making. The hyperspectral data of five common economic animals (scallop, ctenophore, veined red snail, wrinkled disc abalone, and imitation spiny ginseng) in different underwater environments were acquired, normalized and processed using random forest (Random Forest, RF) in machine learning, random forest based on principal component analysis method (Principal Component Analysis-Random Forest, PCA-RF), and random forest based on recursive feature elimination method (Recursive feature elimination- Random Forest, RFE-RF). Three random forest algorithms were used to classify five benthic species and for comparative analysis. By ranking the importance of the variables of RF, the reflection spectrum intensity data corresponding to the bands with higher ranking and higher contribution to the model were filtered. Then the top-ranked feature band data were input into the classifier, and the classification accuracy was obtained by optimizing the parameters. The classification results of the data were output to the confusion matrix, and the identification of the five samples could be seen. The lowest recognition accuracy of 64% was obtained for the veined red snail sample; the highest recognition accuracy of 100% was obtained for imitation spiny ginseng and ctenophore scallops; the recognition accuracies of 91% and 96% were obtained for the scallop and wrinkled disc abalone, respectively. The final classification accuracies of the three methods were 90.13% for RF, 95.20% for PCA-RF, and 98.74% for RFE-RF, which showed the feasibility of using the random forest algorithm in the classification of underwater hyperspectral data.