From Rules to Learning

Your regression model from Mission 6 predicts streamflow from rainfall reasonably well, but your supervisor poses a harder problem: predicting which of Bluewater Basin's 847 land parcels will flood during the next major storm event. There are dozens of variables, slope, upstream drainage area, soil type, land cover, distance to river, previous soil moisture, and no simple equation that combines them in the right way.

She slides a folder across the table: ten years of post-flood field surveys, where crews drove the basin after every major event and recorded which parcels flooded and which didn't. 'We already have the answer for the past,' she says. 'Can we teach a computer to learn the pattern and predict the future?'

Traditional modelling writes the rules explicitly: you decide which variables matter and how to combine them. Machine learning inverts this, it finds the rules itself by looking at historical examples where the answer was known, and extracts patterns the human analyst might miss.

This only works if the historical examples are representative of the future. The algorithm is only as smart as the data you feed it. Garbage in, garbage out applies to machine learning harder than anywhere else, because the algorithm can't tell you when the data is wrong.

It is like the difference between programming a robot with a rigid rulebook for chess versus letting it play a thousand games and figure out good moves on its own. Machine learning is that second approach: instead of writing the rules yourself, you show the computer a pile of past examples and let it work out the pattern.

The shift from rule-based to machine-learning models transformed environmental science over the 2010s. Random forests, the algorithm you'll build in this mission, became the workhorse model for flood risk, species distribution, soil contamination, and wildfire spread, largely because they're accurate, relatively easy to interpret, and handle the messy, mixed-type data that real fieldwork produces.

Explain what 'learning from labelled examples' means, using the flood survey data as your example.