Machine Learning-Based Integration of Climate and Desert Landform for Analyzing Sand Dune Processes

This study aims to identify sand dunes and predict their movement rates using deep neural network-based automatic mapping in the Bafgh Desert, Iran. Additionally, spectrum behavior analysis, including albedo wave and infrared investigations, is conducted to assess dune movement from 1994 to 2024. A machine learning-based neural network is also employed to predict weather parameters and their impact on dune movement. The results indicate that the very deep super resolution (VDSR) deep neural network algorithm significantly enhances satellite image resolution and improves the extraction of morphometric features of sand dunes in desert regions. The U-Net model, with 92% accuracy, effectively classifies sand dunes, which reveals their progression toward residential and agricultural lands and poses significant risks to these areas in the study region. Additionally, spectral reflectance analysis reveals that increases in albedo reflection and infrared wave values indicate the expansion of sand dunes in the study area. Finally, the results of the long short-term memory (LSTM) neural network, with an accuracy of R = 0.90, predict that in the coming years, lower humidity, higher temperatures, and increased dryness will speed up dune movement, especially towards the east and south, creating risk for residential and agricultural areas in the study regions.