Neural Network Results

90.4%

Cover type accuracy

On held-out test data

48K / 12K

Train / test plots

Sampled from 581K total

3 layers

Network depth

128 → 64 → 32 neurons

60 epochs

Training iterations

Covertype · ReLU · Adam

Cover Type Classifier 90.4% accurate

✅ Excellent result. The network learned to predict forest cover type from just 10 terrain measurements (elevation, slope, distances) with 90.4% accuracy. Lodgepole Pine is the easiest to identify (95.7%) because it occupies a very distinct elevation band. Ponderosa Pine is hardest (61.1%) — it overlaps with Douglas-fir in mid-elevation terrain.

Training progress — accuracy over 60 epochs

Network architecture

Input → 128 neurons → 64 → 32 → 7 outputs · ReLU activation · Adam optimiser

Per-class accuracy — how well the model identifies each species

Feature importance — which inputs matter most

Confusion matrix — rows = actual species · columns = predicted · numbers are % of actual class

Decision space — what species does the model predict at each elevation & fire-distance combination? (all other features fixed at median values)

0 m1,000 m2,000 m3,000 m4,000 m

← Distance from nearest fire ignition point (m)

Fire Size Predictor A hard problem

⚠️ Fire size is notoriously difficult to predict from weather alone. The model's R² is negative — meaning it's worse than simply predicting the average. This isn't a modelling failure — it's a well-known result. Over 47% of fires in the dataset burned zero hectares, and a handful burned >100 ha. The distribution is so extreme that no simple model can reliably separate "will this be catastrophic?" from weather data alone. What the model does show is which features matter most: temperature, drought code (DC), and wind speed.

Training loss — model learns but overfits quickly

Feature importance — what the model leaned on

Actual vs predicted burn area (ha) — each dot is one fire in the test set

A perfect model would have all points on the diagonal line. The spread shows how unpredictable individual fire sizes are. The model predicts moderate values (2–15 ha) for almost everything, while actual sizes range from 0 to 1,091 ha.