Urban Land Use Change
My first remote sensing project taught me that a classification is only as good as its training data.
Urban growth leaves a record in satellite imagery. Every year of expansion shows up as pixels that used to read one way and now read another. But that record does not read itself. Getting from raw imagery to a clear picture of change takes deliberate classification work, and this project was my first real attempt at it.
The project came out of GEOS 373, Remote Sensing, in my Geography degree at UBC. Our goal was to measure how much Dubai had urbanized between 2002 and 2015, one of the fastest-growing cities on earth, by classifying satellite imagery from each year. We built the whole thing in ArcGIS Pro.
What classification actually involves
Classification means teaching the software to sort every pixel into a land cover category: urban, vegetation, water, and so on. You feed it training samples, areas where you already know what the ground is, and it extends those patterns across the full image. Do that for each year, then compare the classified rasters, and the differences between them become your change map.
That is the clean version. In practice, most of the work is judgment. A pixel at the edge of a parking lot and a pixel of dry grass can look almost identical to the classifier. Where you draw your training samples, and how honest they are, decides whether the output means anything.
That was the lesson that stuck. Classification accuracy depends on training data quality, full stop. Sloppy samples produce confident-looking maps that are quietly wrong. Careful samples take longer and produce maps you can stand behind. I did not fully appreciate the difference until I watched my own outputs shift based on nothing but the inputs I chose.
The result was stark. Dubai's developed land grew by nearly 50% between 2002 and 2015, and not only on land. The city expanded straight out into the Gulf, new islands and waterways where the classified 2002 image had shown open water. The change map read like time-lapse: a coastline redrawn in thirteen years. Fittingly, the classifier's hardest calls were exactly at the growing edge, where developed and undeveloped pixels blur, which is the whole lesson above made visible.
We documented the findings in academic poster format for course reporting. The poster is here if you want to see the maps.
This project carries forward more than any single map I made in it. Every data-heavy project I scope now starts with the same question it taught me to ask: how good are the inputs, really? The answer decides everything downstream.