Concepts

This page is the shared mental model behind Ronin. It applies to both feature modes (convolution and legacy hand-tuned).

Gates and the classification problem

A radar sweep is a grid of gates (range × time/azimuth). Each gate is either meteorological (real weather echo we want to keep) or non-meteorological (ground clutter, noise, sidelobes — we want to remove it). Ronin trains a Random Forest to make that binary call per gate, then blanks the non-meteorological gates in the raw fields.

MD / NMD label convention

• Meteorological data (MD) is 1 / true.
• Non-meteorological data (NMD) is 0 / false.

This convention holds everywhere: training targets, predictions, masks.

ModelConfig is the single source of truth

Ronin is built around one configuration struct, ModelConfig. Everything needed to define, train, evaluate, and apply a QC model lives in it: data paths, the number of cascade passes, feature settings, RF hyperparameters, per-pass thresholds, and QC output variables. Construct it with make_config (which auto-derives paths and mask_names), persist it with save_config, and reload it with load_config.

config.task_mode selects the feature regime — "convolution" for the recommended kernel-bank mode, "" for the legacy hand-tuned predictors. The behavior is not agnostic to this value; it branches across training, prediction, QC, and mask generation.

The pipeline: features → RF → met-prob → mask

For each pass:

1. Features are computed per gate (a kernel bank in convolution mode, or hand-tuned spatial predictors in legacy mode).
2. A Random Forest is trained / applied to those features.
3. The RF emits a meteorological probability per gate — the model's confidence that the gate is weather. In training this is written back into the CfRadial files as met_prob_pass_<N>.
4. A mask is derived by keeping gates whose met-prob falls in the pass's (low, high) threshold window.

The multi-pass cascade

num_models sets the number of passes. With num_models > 1, each pass after the first only sees the gates that survived the previous pass's mask. Pass i (for i < num_models) writes met_prob_pass_<i> and the mask mask_names[i+1]; the next pass trains/predicts only within that mask. This lets later passes specialize on the harder, ambiguous gates.

• met_probs::Vector{Tuple{Float32,Float32}} — one (low, high) window per pass.
• mask_names — default ["mask_pass_0", "mask_pass_1", ...] from make_config; mask_pass_0 is the (empty) starting mask.

Two passes is the recommended starting point.

FILL_VAL and missing data

Missing / blanked gates are represented by Ronin.FILL_VAL (typemin(Int16) = -32768). Windowed spatial reducers exclude missing gates from their neighbourhoods; a fully-missing window collapses to FILL_VAL.

Variable name glossary

Ronin works with multiple radar conventions. The QC'd (interactively edited) fields serve as ground truth for training.

ELDORA

NOAA TDR