Development and Validation of a Predictive Model for Herbaceous Plant Growth Based on Water-Sediment Stress

Sedi ment accumulation and alternately submerged caused by water conservancy engineering pose challenges to ecological restoration in riparian zones. In this study, a hybrid model coupling convolutional neural network (CNN), long short-term memory network (LSTM), and batch normalization (BN) techniques was employed to predict the concentrations of physiological indicators (soluble sugars (SS), soluble proteins (SP), malondialdehyde (MDA), and chlorophyll-a (Chla) content) in five herbaceous plants under varying water-sediment conditions. The findings indicated that the CNN-LSTM model outperformed the LSTM model in forecasting the overall trend and fluctuations of the data, particularly with reduced bias in predicting extreme values. Additionally, the CNN-LSTM-BN model improved predictive capabilities for the four physiological indicators. The mean absolute error (MAE) for the true and predicted values in the verification set were 0.040, 0.039, 0.017, and 0.020, respectively. The integration of the BN layer facilitated gradient propagation of the loss function across each parameter of the model, thereby accelerating learning efficiency and enhancing training process stability. Furthermore, the CNN-LSTM-BN model successfully predicted the concentrations of SS, SP, MDA, and Chla in Paspalum wettsteinii under a 3 cm water-sediment condition, yielding MAE values of 0.308, 0.249, 0.365, and 0.169, respectively. These findings highlight the strong predictive capabilities of the CNN-LSTM-BN model. Overall, this study offers valuable insights for plant selection and maintenance strategies in riparian ecological restoration efforts within the context of water conservancy engineering.