Predicting impact of land cover change on flood peak using hybrid machine learning models

The present study evaluates the performance of hybrid machine learning models to predict flood peak due to land cover changes. Performance of feed forward neural network (FNN) and adaptive neuro-fuzzy inference system (ANFIS) was compared and analyzed to select the best model in which different conventional training algorithms and evolutionary algorithms were applied in the training process. The inputs consist of stream flow in previous time step, rainfall and area of each land use class, and output of the model is stream flow in the current time step. The models were trained and tested based on the available data in a river basin located in the Australian tropical region. Based on the results in the case study, invasive weed optimization is the best method to train the machine learning system for simulating flood peak. In contrast, some optimization algorithms such as harmony search algorithm are very weak to train the machine learning model. Furthermore, results corroborated that the performance of FNN and NFIS is the same in terms of generality. The FNN model is more reliable to predict the flood peak in the case study. Moreover, ANFIS-based model is more complex than FNN. However, ANFIS is advantageous in terms of interpretability. The main weakness of ANFIS-based model is underestimation of flood peak in the major and minor floods. Two scenarios of changing land cover were tested which demonstrated reducing natural cover might increase the flood peak more than twice.