Research Gap

Reviewed Paper

Category Research Gap Specific Example from Paper Severity Affected Paper/Model Potential Impact
Regression Tasks Text-based tokenization leads to precision loss in continuous value prediction. Tokenization of "100" into ["1", "00"] causes errors in AGB (Above-Ground Biomass) estimation tasks. High REO-VLM Improved environmental monitoring and carbon stock calculations.
Multispectral/SAR Adaptation RGB-centric architectures underperform on non-RGB data (e.g., SAR, hyperspectral). SpectralGPT struggles with SAR data due to inherited RGB-focused pretraining frameworks. High SpectralGPT Enhanced utility in flood mapping, mineral exploration, and military surveillance.
Multimodal Output Limitation Text-only outputs limit utility in dense predictions (e.g., segmentation, 3D). VLMs cannot generate segmentation masks or 3D models, restricting flood mapping applications. Moderate-High CPSeg Enable real-time hazard mapping and infrastructure planning.
Temporal Data Handling Static image analysis neglects temporal dynamics critical for trends. Changen2’s synthetic temporal data lacks real-world complexity for deforestation monitoring. Moderate Changen2 Improved climate change prediction and disaster response.
Benchmarking No unified standards for cross-task evaluation (e.g., VQA, RSICC). RSVQA-HR and LEVIR-CC use incompatible metrics, complicating cross-model comparisons. Moderate RSVQA, LEVIR-CC Accelerate model innovation via standardized benchmarks.
Ethical AI & Bias Auto-annotated datasets inherit biases from foundation models. RS5M’s captions generated by BLIP2/GPT-4 reflect urban/rural bias, skewing agricultural monitoring insights. Moderate RS5M Fairer AI applications in policy-making and environmental justice.
Sustainability High computational costs hinder accessibility and scalability. SkyEyeGPT’s training consumes excessive GPU resources, limiting adoption in resource-constrained regions. High SkyEyeGPT Democratize access to VLM tools for global environmental monitoring.
Priority Category Issues Notes
High Regression Tasks Text-based tokenization causes precision loss in continuous value prediction AGB estimation errors in REO-VLM due to splitting "100" into ["1", "00"]
High Multispectral/SAR Adaptation RGB architectures ill-suited for SAR/hyperspectral data SpectralGPT struggles with SAR due to RGB-focused pretraining
High Sustainability Excessive GPU requirements limit accessibility SkyEyeGPT requires High-end GPUs unavailable in developing regions
Moderate-High Multimodal Output Text-only outputs restrict segmentation/3D tasks CPSeg cannot generate flood masks directly via VLMs
Moderate Temporal Analysis Static models ignore temporal dynamics Changen2's synthetic data lacks real-world temporal patterns
Moderate Benchmarking Lack unified evaluation standards RSVQA vs LEVIR-CC use incompatible metrics
Moderate Ethical AI & Bias Auto-annotated data inherits LLM biases RS5M captions show urban/rural bias affecting agriculture insights