Deep Learning-Based Prediction of Land Use and Land Cover Dynamics in the Urmia Lake Basin

Document Type : Original Article

Authors

1 School of Surveying and Geospatial Engineering, College of Engineering, University of Tehran

2 School of Civil Engineering, Shahrood University of Technology, Shahrood, Iran

3 University of Natural Resources and Life Sciences, Vienna

10.22059/eoge.2026.409793.1205

Abstract

Monitoring and analyzing land use and land cover (LULC) is crucial for understanding environmental transformation, urban planning, deforestation, and water resource management. The Urmia Lake basin-a critical water body in northwestern Iran- has vital LULC changes over the past two decades.
The study aims to predict future LULC in the Urmia Lake basin and generate LULC maps using deep learning models applied to satellite imagery. by comparing the performance of diverse deep learning models, the study seeks to identify most accurate approach for modeling spatio-temporal LULC changes. To achieve this, four models are implemented: a Multilayer Perceptron (MLP) which captures complex nonlinear relationships in data, two-dimensional and three-dimensional Convolutional Neural Network (2D CNN and 3D CNN) for effective extraction of spatial features from satellite imagery, and a Recurrent Neural Network (RNN) to model temporal changes over time. The input data consist of 64 features derived from six optical Moderate Resolution Imaging Spectroradiometer (MODIS) bands, Land Surface Temperature (LST), vegetation and surface indices (Normalized Difference Vegetation Index (NDVI), Normalized Difference Snow Index (NDSI), and Normalized Difference Water Index (NDWI)), topographical layers (Digital Elevation Model (DEM), slope, and aspect), 50 ERA5-Land monthly climate variables, and precipitation data from the PERSIANN Climate Data Record (PERSIANN-CDR). The dataset spans 22 years and includes eight land cover classes as output labels. Among the evaluated models, the 3D CNN reached the highest performance with a test accuracy of 92.82%, mean Intersection over Union (IoU) of 0.6351, and the lowest number of mismatched pixels. Test accuracy of 3D CNN indicates that 92.82% of the pixels in the independent test dataset for the year 2022 were correctly classified into their respective land use/land cover categories. These results confirm its superior ability to capture complex spatio-temporal patterns for accurate LULC prediction in the Urmia Lake basin.

Keywords

Main Subjects

Earth Observation and Geomatics Engineering
Volume 9, Issue 2
December 2025

Files

Share

How to cite

Statistics