Application of Remote Sensing and Machine Learning for the Identification of Dust Sources in Khuzestan Province

Khuzestan Province, due to its arid climate and unique geographical location, is one of the main regions affected by dust storms in Iran. This study utilized satellite data and the Random Forest machine learning algorithm to identify dust source areas in the province during the period from January 26 to April 26, 2025. Sentinel-2, Sentinel-5P, and SMAP satellite data were used to calculate the Normalized Difference Vegetation Index (NDVI), Absorbing Aerosol Index (AAI), and surface soil moisture respectively. These indices monitored vegetation cover, the presence of dust-related aerosols, and soil dryness.
Google Earth Engine (GEE) was employed for data processing. The steps included defining the study area (Khuzestan, with a surface area of 64,057 km², based on FAO/GAUL/2015 data), filtering satellite images with cloud cover less than 10%, calculating the average values of NDVI, AAI, and soil moisture, and integrating these layers into a composite image. A Random Forest classifier with 50 decision trees was used to categorize regions into source and non-source classes. Sample points included 200 source points (e.g., Hoor-Al-Azim wetland) and 200 non-source points (urban and agricultural areas). Of these points, 70% were used for model training, 15% for validation, and 15% for final testing and evaluation. NDVI, Soil Moisture and AAI time series were also analyzed to assess aerosol variation trends, and model accuracy was evaluated using a confusion matrix. The results obtained from the validation data (overall accuracy = 0.95, F1-score = 0.95, Cohen’s Kappa = 0.90) and the final test (overall accuracy = 0.967, F1-score = 0.967, Cohen’s Kappa = 0.933) indicate a very strong and stable performance of the mode. The results indicated that the southern and western parts of Khuzestan—especially areas around Hoor-Al-Azim, southeastern Ahvaz, and the regions of Mahshahr, Hendijan, Omidiyeh, Susangerd, and Hoveyzeh are major dust sources. The classification map (Figure 8) highlighted these areas in red and non-source areas in green. The NDVI map (Figure 2) showed low vegetation coverage (values less than 0) in the southwest, particularly near Hoor-Al-Azim, indicating dry, bare-soil conditions. The AAI index (Figure 3) confirmed high aerosol values (0.2 to 0.5) in the central and southern regions, especially around Ahvaz and Hoor-Al-Azim. Soil moisture values (Figure 4) were also low in the south and central areas (below 0.1 cm³/cm³), indicating high potential for wind erosion. Key contributing factors to dust source activation included drought, reduced rainfall, the drying of the Jarrahi and Karun rivers, and desiccation of the southern parts of Hoor-Al-Azim.
This study successfully identified dust source areas in Khuzestan and is consistent with previous research, such as that by Rangzan et al. (1393) The inclusion of soil moisture as a new variable improved identification accuracy. The Random Forest algorithm provided reliable performance, due to its ability to model nonlinear relationships and handle heterogeneous data. However, limitations included the lack of field data for validation and the influence of atmospheric conditions on satellite observations.
Future studies are encouraged to incorporate field observations for validation and increase the number of sample points. Moreover, more advanced methods such as deep learning and the integration of satellite data with meteorological models (e.g., wind speed and direction), could improve dust storm forecasting accuracy. This approach especially with the flexible coding environment of Google Earth Engine offers an effective tool for monitoring and managing dust sources and can be applied in wetland restoration and soil stabilization efforts in desert regions.