Prediction of NDVI dynamics under different ecological water supplementation scenarios based on a long short-term memory network in the Zhalong Wetland, China

Abstract

Wetland plants are a key factor in ecosystems but are threatened by water extraction and water resource exploitation. Ecological water supplementation is a common solution to the water scarcity problem in wetlands. In this study, the Zhalong Wetland with a complex water regime was divided into several subareas with relatively strong hydrologic connectivity based on a hydrodynamic model and cluster analysis. The normalized difference vegetation index (NDVI) dynamics were simulated and verified with long short-term memory (LSTM) neural network models established for the various wetland subareas based on the subarea water levels and temperature and sunshine duration data. To evaluate the effects of different ecological water supplementation scenarios on the spatiotemporal NDVI variation across the wetland, the water levels of subareas under different scenarios were used to drive the LSTM model. The results indicate that (1) the spatiotemporal variation in the NDVI at most nodes within the wetland was accurately simulated, but large errors generally occurred in regions with small lakes. (2) Ecological water supplementation with continuous and low-flow discharge imposed highly positive influences on the annual maximum NDVI in the wetland against the background of limited water resources. (3) The lower reach in the largest subarea was less positively affected by wetland ecological water supplementation. (4) Water supplementation with discharge in April and September was recommended because of the benefits for the nidification and breeding of water fowl species and the limited conflict between agricultural and ecological water. This study provides important references for both wetland and water resource managers.