An integrated hybrid deep learning data driven approaches for spatiotemporal mapping of land susceptibility to salt/dust emissions

Abstract

One of the main problems in arid and semi-arid climates, which leads to environmental, economic, and health consequences, is dust storm. Dust causes significant environmental disasters worldwide and poses a threat to human life and property. Therefore, this study examined the effectiveness of deep learning algorithms for sensitivity to dust emissions analysis. To achieve this, two sets of datasets were utilised in the Urmia Lake basin: one containing detected dust points and the other comprising various environmental, physical, and climatic factors that influence dust distribution. Hybrid deep learning algorithms were applied using an innovative approach and integrated with a clustering technique. To achieve this goal, convolutional neural networks (CNN), deep belief networks (DBN), generative adversarial networks (GAN), long short-term memory (LSTM), restricted Boltzmann machines (RBM), and recurrent neural networks (RNN) were employed as effective deep learning algorithms to analyse the results. The optimisation and aggregation of different classifications for each factor were developed as the basis of spatiotemporal modelling. The inclusion of multiple factors was intended for a comprehensive investigation of dust triggers and intensifiers, and combined modelling with various models enhanced the validity and efficiency of the results. The models’ performance and prediction accuracy were assessed using statistical indices, including the mean squared error (MSE), root mean squared error (RMSE), and receiver operating characteristic (ROC).. The MSE and RMSE results for the models based on real and predicted data demonstrate the high accuracy of the DBN, RNN, RBM, GAN, CNN, and LSTM models. Moreover, during validation, results pointed out that the ROC values of the algorithms were relatively similar. The ROC values in the training phase of the six hybrid models ranged from 0.939 to 0.966, indicating the high accuracy of the models’ learning. The GAN model indicated the highest learning success with an ROC of 0.966 in the training set, whereas the DBN model showed the best performance in prediction with an ROC of 0.915. This study identified soil depth as the most influential factor in dust occurrence in the area, with temperature and humidity having the least impact. The results of this study are critical from the geoinformation perspective, as an interdisciplinary field of science by means of applying various methods and evaluating their efficiency in analysing environmental impacts and assessing salt/dust distribution.