Parameter-Efficient Adaptation of Generative-Foundation (Flux, Qwen) vs. Zero-Shot (Gemini, SAM3) Models for Aerial Image Segmentation

Dataset Overview:

This dataset and the associated LoRA-adapted model weights (FLUX.1 Kontext, FLUX.2, and Qwen) provide the experimental framework and results for the study: 'Parameter-Efficient Adaptation of Generative-Foundation (Flux, Qwen) vs. Zero-Shot (Gemini, SAM3) Models for Aerial Image Segmentation', published in Buildings (MDPI), 2026, 16(7), 1369 https://doi.org/10.3390/buildings16071369 . The records include the Brisbane metropolitan training/testing tiles and the fine-tuned adapters that achieved a peak mean Intersection over Union (IoU) of 89% and 96% pixel accuracy, demonstrating the efficacy of generative foundation models for high-precision spatial analysis.

Technical Specifications:

·         Geographic Focus: Brisbane, Australia (Metropolitan regions).

·         Task: Binary Semantic Segmentation (Rooftop vs. Non-rooftop).

·         Data Split: 121 total tiles (80/20 split)- Training: 97 tiles.

·         Testing: 24 tiles.Format: High-resolution aerial imagery paired with dense binary annotations.

Benchmarked Models & Performance:

This dataset serves as the benchmark for several state-of-the-art architectures evaluated under a unified protocol:

Key Research Insights:

·         Diffusion as Segmentors: The dataset demonstrates that with only 250–5000 steps of LoRA fine-tuning, generative models like FLUX.1-Kontext can outperform dedicated segmentation models (SAM3), reaching a mean IoU of 89%.

·         Generalization: While Gemini 3 Pro provides a powerful zero-shot baseline, fine-tuned diffusion models offer superior boundary fidelity in heterogeneous urban morphologies.

·         Efficiency: The results highlight that parameter-efficient tuning is a viable path for transforming general-purpose AI into scalable spatial analysis tools for urban monitoring and solar potential assessment.

Citation

• Dataset could be cited from:   https://explore.openaire.eu/search/result?pid=10.5281%2Fzenodo.18571111
• APA style ciataion: "Shata, D., Denman, S., Omrani, S., Drogemuller, R., Ali, H., & Wagdy, A. (2026). Parameter-Efficient Adaptation of Generative-Foundation (Flux, Qwen) vs. Zero-Shot (Gemini, SAM3) Models for Aerial Image Segmentation. https://doi.org/10.5281/zenodo.18571111"
• APA style Publication ciataion: "Shata, D., Denman, S., Omrani, S., Drogemuller, R., Ali, H., & Wagdy, A. (2026). Parameter-Efficient Adaptation of Generative-Foundation (Flux, Qwen) vs. Zero-Shot (Gemini, SAM3) Models for Aerial Image Segmentation. Buildings, 16(7), 1369. https://doi.org/10.3390/buildings16071369"