Domain Adaptation of Geospatial Vision Transformers (ViT) for Climate Resilience in the Gulf: A Self-Supervised CV Approach
2026 International Conference on Integrated Intelligence and Cognitive Engineering, ICIICE 2026, Dubai, Birleşik Arap Emirlikleri, 18 - 19 Nisan 2026, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/iciice69672.2026.11565097
- Basıldığı Şehir: Dubai
- Basıldığı Ülke: Birleşik Arap Emirlikleri
- Anahtar Kelimeler: Arabian Gulf, Climate Resilience, Domain Adaptation, Geospatial Foundation Models, Remote Sensing, Self-Supervised Learning, Vision Transformers
- Hatay Mustafa Kemal Üniversitesi Adresli: Evet
Özet
The Arabian Gulf region is experiencing rapid environmental stress driven by hyper-aridity, coastal salinization, accelerated urbanization, and extreme heat events. Accurate and continuous monitoring of these phenomena is critical for climate resilience and sustainable development. However, most existing Earth Observation (EO) deep learning systems rely on models trained on global datasets dominated by temperate and vegetated landscapes, resulting in a substantial domain gap when applied to the Gulf's unique spectral and spatial characteristics. In particular, conventional convolutional neural networks and even modern geospatial models often fail to robustly distinguish subtle land cover classes such as sabkha, sparse desert vegetation, coastal sediment variation, and heat-affected urban surfaces.This paper proposes a self-supervised domain adaptation framework for geospatial Vision Transformer (ViT) foundation models tailored to the Arabian Gulf region. The proposed approach leverages large-scale, unlabeled Sentinel-2 satellite imagery and employs masked autoencoding objectives to adapt global foundation models to hyper-arid environments without requiring extensive labeled data. Parameter-Efficient Fine-Tuning (PEFT), specifically Low-Rank Adaptation (LoRA), is utilized to achieve efficient adaptation while training less than 1% of the total model parameters. The framework benchmarks state-of-the-art geospatial foundation models, including Prithvi-100M and SatMAE, against traditional convolutional architectures on downstream climate-relevant tasks. Experimental results demonstrate that the adapted models significantly improve segmentation accuracy and representation robustness while maintaining low computational overhead. This work establishes a scalable, resource-efficient pathway for deploying foundation-model-based climate monitoring systems in under-represented arid regions.