Geo-AI

The primary dust sources on the Earth are found in desert regions and drylands such as those of the southern Sahara, the Middle East, central and northeastern Asia, and parts of the Australian subcontinent. Nevertheless, dust storms appear to be an inescapable phenomenon throughout all dry and semi-arid regions of the globe. Originating in several arid regions on earth (Safriel et al., 2005), dust storms accompany several risks for human communities due to the long-term transportability of aeolian deposits to regions far away from their origin (Goudie and Middleton, 2006, Middleton, 2017). Considerable attention has also been given throughout the past three decades—coinciding with intense changes in climate, global warming, and drought runs throughout the globe (Meehl et al., 2007)—to the study of the dust storm phenomenon and several of its effects on both land and marine ecosystems (Prospero et al., 2002, Goudie, 2009).

Assessing the environmental impacts of dust requires knowledge of dust sources, as the prime origin for atmospheric dust cycles (Schepanski et al., 2007). Accurate identification of dust sources, however, is a challenging and complex task, demanding knowledge and understanding of surface characteristics such as land cover, land formations, and physicochemical characteristics of underlying soil (Parajuli et al., 2014). Large-scale studies on dust, in cases where accurate data on land surface characteristics is unavailable, can benefit from the incorporation of dust source susceptibility mapping (DSSM) using remote sensing data as an efficient tool for the management of dust-environmental crisis on local to global scales. Despite the significance and applicability of DSSM, little attention has been devoted to the concept of the required techniques and the accompanying challenges. Recently, some investigations have incorporated multi-criteria decision-making (Cao et al., 2015, Darvishi Boloorani et al., 2020) and mostly data mining techniques (Boroughani et al., 2020, Rahmati et al., 2020, Gholami et al., 2020a, Gholami et al., 2020b) in attempts to identify and map susceptible dust sources (SDS) in the Middle East. A thorough analysis of these studies revealed certain shortcomings that limit the accuracy and precision of the output maps and the overall generalizability of the developed models. These shortcomings can be assessed from three main perspectives: type and logic of the proposed model for DSSM; type and number of data used for modeling the environmental drivers of dust events; and type of training and data used for the machine learning-based approaches in DSSM.

From the first perspective, it can be concluded that employing decision-making models based on human knowledge such as Analytical Hierarchy Process (AHP) (Darvishi Boloorani et al., 2020) has increased the level of uncertainty and decreased the generalizability of the results. More precisely, the process of comparison, prioritization, and determination of the importance/weight of input parameters on the output map was highly dependent on expert opinion. A more reliable approach is to incorporate machine learning models based on artificial intelligence, which help reduce the role of human factors (opinion) and, therefore, provide a more thorough understanding of the complexities and relations among input parameters (Olden et al., 2008). However, a more detailed look at the results from machine learning models employed for DSSM in studies by Boroughani et al., 2020, Rahmati et al., 2020 clearly shows that the low overall performance and low prediction accuracy of a classifier model is not enough to justify dismissing the model as ineffective. In other words, different classifiers show different prediction accuracies for each class, which determines the final performance of the model. Therefore, given the unique capabilities of machine learning models in predicting SDS, it is necessary to incorporate the models via ensemble to maximize their performance. The topic of ensemble approaches developed for DSSM has been investigated by Gholami et al., 2020a, Gholami et al., 2020b using data mining models, which produced favorable outcomes. They used an optimal combination of classifiers based on the overall prediction accuracy of models for all classes. However, it remains an issue as to “how can different classifiers be combined with respect to their prediction accuracies for each class”. Moreover, the proposed modeling frameworks for DSSM using machine learning models and data mining techniques have been primarily implemented and assessed on a local scale. Considering the wide diversity of environmental parameters affecting dust events and the type of dust sources throughout the globe, the proposed models in previous studies likely did not possess the required generalizability to be employed for different environmental conditions and, therefore, failed to achieve acceptable outcomes. This highlights the importance of selecting an appropriate case study that covers a wide range of environmental parameters for feasibility testing of new models and helps identify and eliminate potential flaws as well as increase overall generalizability.