Convolution Feature Mode

Set task_mode = "convolution". This is the recommended feature regime as of 1.2.0. Instead of a hand-tuned predictor file, features are derived by convolving a kernel bank over a set of radar variables.

Kernel bank

A ConvolutionKernel is a named Matrix{Float32} of weights. build_kernel_bank(kernel_sizes) builds the bank: for each size k it adds a mean_<k>x<k> kernel; it always adds laplacian_3x3, sobel_range_3x3, and sobel_azi_3x3; and for each k ≥ 5 it adds a gaussian_<k>x<k>.

With the default conv_kernel_sizes = [3, 5, 7] this yields 8 kernels: mean_3, mean_5, mean_7, laplacian_3x3, sobel_range_3x3, sobel_azi_3x3, gaussian_5x5, gaussian_7x7.

Feature layout

compute_convolution_features produces, per conv-variable × kernel, two columns:

• <var>_<kern> — the convolved value.
• <var>_<kern>_vfrac — the valid (non-missing) fraction within the kernel footprint.

If a met-prob mask is supplied, the experimental masked-convolution columns are appended next (see below). Finally, four scalar features are appended, in this order: AHT, ELV, RNG, NRG.

The total feature count is given by get_convolution_feature_count:

n_vars * n_kernels * 2  +  n_masked_vars * n_masked_kernels * 2  +  4

conv_variables

config.conv_variables lists the inputs the kernel bank operates on. The resolver (load_conv_variable) accepts:

• Raw CfRadial fields (e.g. "DBZ", "VEL", "WIDTH").
• Derived variables PGG (probability of ground gate) and SIG (signal quality).
• Spatial reducers AVG(var), ISO(var), STD(var) — resolved recursively, so ISO(PGG) works.
• Prior-pass output met_prob_pass_<N> (used by cascade pass N+1).

conv_variables can differ per pass via PASS_CONFIG — a common pattern is to append "met_prob_pass_1" to Pass 2's variables.

Computing features

calculate_features_conv(config, output_file; …) is the convolution-mode replacement for calculate_features. In the workflow this is driven by 02_train.jl / run_workflow.jl; you rarely call it directly.

Feature selection / importance

Training on all features then computing permutation importance (compute_importance) writes recommended_features (and importances) into the model's JLD2. Inspect with inspect_model_configuration, set config.selected_features to the recommended indices, and retrain.

Note: you must retrain on the selected feature subset — you cannot train on N features and prune to K at inference, because the Random Forest's tree splits reference specific column indices. The model's selected_features is reloaded at inference so the feature matrix matches what the trees expect.

Met-prob-masked convolutions

When enabled, an additional set of convolution columns is computed where the kernel only accumulates contributions from neighbours that the previous pass considered meteorological. masked_convolve distinguishes the valid set (which gates get an output) from the neighbor mask (valid ∧ met-prob mask — which gates contribute to the sum).

Configuration fields:

• masked_conv_variables — variables to compute masked-conv columns for.
• masked_conv_kernel_types — subset of {mean, gaussian, laplacian, sobel_range, sobel_azi}.
• masked_conv_kernel_sizes — kernel sizes.
• masked_conv_threshold — met-prob cutoff (default 0.1f0).
• masked_conv_met_prob_field — which met_prob_pass_<N> field to mask with.

The filtered kernel bank is the Cartesian product of types × sizes, with the same structural constraints as the main bank (laplacian/sobel only at k == 3, gaussian only at k ≥ 5); other combinations are silently skipped and the skipped set is reported once at training time.