The problem
Analysing satellite imagery for urban planning involved repeated manual steps. The team needed both more accurate classification and a faster way to produce usable reports.
What I did
I compared three machine-learning approaches, built an analysis pipeline, and automated recurring reporting. I also connected the outputs to flood-risk and land-use questions.
The result
Segmentation accuracy rose from 87.8% to 92.8%, while recurring workflow time fell by roughly 65%.
What I learned
The useful result was not a model score alone; it was a more dependable path from imagery to a planning decision.