Birdi’s AI tools (powered by Segment Anything / SAM-based models) are designed to help you quickly identify and extract features from your geospatial data. While powerful, they’re not always perfect—and understanding how they work can help set the right expectations and improve results.
Why AI Detections May Be Delayed
You may notice that AI detections (e.g. lakes, vegetation, stockpiles) don’t appear instantly after running a task.
Processing Time
AI detections require:
Analysing high-resolution imagery
Running segmentation models across large areas
→ Larger datasets or higher resolution imagery will take longer to process.
Layer Generation
Similar to other outputs, AI results need to be:
Generated
Converted into vector layers
Loaded into your workspace
→ This can create a short delay before results appear.
Why Results Aren’t Always Perfect
AI models like SAM are general-purpose segmentation models, not trained for every specific use case.
Not Object-Specific by Default
The model identifies shapes and boundaries—not always semantic meaning
It may detect something that looks like a lake, but isn’t one
→ Accuracy depends heavily on context, imagery quality, and how the tool is used.
Prompting Matters
How you guide the AI impacts results:
Different click points or prompts can produce very different outputs
Small changes in input can significantly improve accuracy
Some Use Cases Require Training
Highly specific objects (e.g. certain infrastructure, materials, or edge cases) may require:
Custom model training
Additional refinement workflows
→ Out-of-the-box AI won’t always meet specialised needs, you can contact us to discuss your specific use case
Common Issues You Might See
AI Detections Delayed in Appearing
You may run a detection (e.g. lakes) and not see results immediately.
Why:
Processing is still running in the background
Large datasets take longer to analyse
Artifacts After AI Processing
You might notice unwanted outputs like:
“Squiggly” lines
Irregular or noisy boundaries
Why:
The model is trying to interpret complex textures or edges
Image noise or low contrast can affect results
Over-segmentation in detailed areas
Best Practices for Better Results
Always Review AI Outputs
AI is a starting point—not a final answer.
Validate detections before using them in reports or decisions
Clean up or refine where needed
Guide the Model Carefully
Use clear, intentional prompts (clicks, selections)
Try multiple prompt variations if results aren’t accurate
Optimise Your Inputs
Use clear, high-quality imagery
Avoid overly noisy or low-contrast areas where possible
Start Simple
Focus on a smaller area first
Validate results before scaling to larger datasets
Need Better Results? We Can Help
If you’re working with a specific object or use case:
Share examples of:
What you’re trying to detect
The results you’re getting
We can:
Review your workflow
Suggest improved prompting approaches
Advise on whether model improvements or custom solutions are possible
In some cases, we can also partner with you to explore model enhancements tailored to your needs.