~/aixsci
weekly updates

fields/climate-weather/model families

Gradient-boosted trees in weather, climate and oceans

How gradient-boosted trees, a method that builds many small decision rules one after another, are used to study the atmosphere, climate and the sea, and why researchers choose them.

How it is used

Weather, climate and ocean science works with measurements: temperatures, rainfall, wind, chemical concentrations, readings from satellites and instruments on the ground. Much of the work is turning one set of measurements into another, such as estimating rainfall where no rain gauge stands, or predicting a chemical outcome from the properties of a sample. Gradient-boosted trees do this by stacking many simple yes-or-no rules. Each rule splits the data on one variable, a threshold on humidity or on an optical index, and gives a crude answer. The next set of rules is fitted to what the previous ones got wrong. Added together, hundreds of such crude answers make a detailed one.

In the examples here the inputs are tables. Columns hold whatever the study measured: cloud and rain-forming variables gathered from several satellite and model sources onto a shared grid, optical and fluorescence indices of atmospheric organic matter, instrument and environmental settings alongside a laboratory signal, or an airport's traffic, fuel use and weather averages. The target is a single quantity in another column, often log-transformed first to tame a wide range of values. Related tasks appear too: deciding whether it rained at all before estimating how hard, or ranking which columns matter. Boosted trees usually sit beside other methods and are compared against them.

Why researchers use it

These tables are awkward for most methods. The columns are measured in different units and on different scales, some are strongly related to each other, some barely matter, and the relationship to the target is rarely a straight line. Trees split on one column at a time, so units and scales do not need matching, and a column that carries no information is simply never used. Boosting then captures bends and thresholds, the point where cloud-top temperature starts to mean rain, without anyone writing those bends down in advance. The physics is often known only in part, and a fitted model stands in for the missing part.

In the record

Limits

Boosted trees learn the dataset they are given, and these datasets are small, uneven and local. A model fitted to one coastline, one instrument or one airport need not hold elsewhere, and performance in a held-out test split can flatter it when neighbouring times and places are closely related. Trees predict by averaging what they have seen, so they handle unusual conditions, the heaviest rain or the rarest chemistry, poorly, which is exactly where the science often matters. And they carry no physics: nothing forces the answer to respect conservation of mass or energy.

Written by AIxSci: the background from general knowledge, the examples from checked records.

Related

Every article