A deep learning convolutional neural network algorithm for detecting saline flow sources and mapping the environmental impacts of the Urmia Lake drought in Iran

Abstract

Urmia Lake in Northern Iran is drying up, which is causing significant environmental problems in the region, including saline storms that devastate agricultural land. We developed a remote sensing-based monitoring application to detect and map the location of saline flow sources with a novel automated deep learning convolutional neural network (DL-CCN). In order to train the model, we derived a normalised difference dust index (NDDI) from MODIS satellite images and collected ground control points (GCPs). These GCPs were randomly split for training (70%) and accuracy assessment (30%). We identified the following seven predisposing factors for saline flow source modelling: normalised difference vegetation index (NDVI), humidity percentage, temperature, wind speed, geomorphology, soil and land use/cover. In order to train the DL-CNN, we used ReLu, the root mean square error function, and Stochastic Gradient Descent (SGD) for the activation, loss/cost function, and optimization, respectively. Finally, we used the frequency ratio (FR) method to identify the most effective variable for the prediction of saline storm occurrences. The results reveal a high confidence (91.86% overall accuracy and a Kappa of 90.26) for the detection of saline flow sources. According to the FR model, the NDVI (0.982), humidity percentage (0.963), and land use/cover (0.925) are the most relevant factors for detecting the occurrence of saline storms in the Urmia Lake basin. In addition, we carried out a spatial uncertainty analysis of the results based on the Dumpster Shafer Theory. The results will help the local stakeholders and decision-makers to better understating the saline flow sources and their respective environmental impacts.