Convolutional neural networks optimized for land cover mapping in the Iraqi province of Al-Muthanna’s Al-Khidhir region

In this study, we analyse the performance of CNNs and six conventional ML models, namely decision tree (DT), random forest (RF), Extremely Randomized Trees (Etree), k-nearest neighbours (KNN), multilayer perceptron (MLP) and support vector machine (SVM), for LULC classification in the Al-Khidhir region of Al-Muthanna governorate, Iraq. Based on Landsat imagery from 1998, 2008, and 2018, we optimized model hyperparameters using randomized search to improve classification performance. Reference data for supervised classification were digitally derived manually from very-high resolution Google Earth images. The outcomes indicate that CNN performed well in comparison to conventional ML models in terms of overall accuracy and F1-score, especially concerning intricate land cover types like wet soil and agricultural land. CNN also generated maps which were spatially more coherent and noise-free. During the 20-year period, significant land cover changes occurred, with a sharp decrease in farmland and an increase in bare areas—suggestive of ongoing desertification. Results show that optimized deep learning models are a promising approach for reliable and interpretable land monitoring in dry and post-conflict areas. The findings are useful for land use planning, environmental policy and Sustainable Development Goals (SDGs), in particular SDG 15 (Life on Land) and SDG 2 (Zero Hunger).

Text 124645.pdf - Published Version Restricted to Repository staff only Download (3MB) Official URL or Download Paper: https://link.springer.com/10.1007/s12524-025-02318...

Actions (login required)

View Item