Migration prediction

rivabar.prediction implements the Howard & Knutson (1984) convolution model for meander migration, following Sylvester et al. (2019, Geology): the predicted migration rate is

R1(s) = Omega * R0(s) + Gamma * conv(R0, G) / integral(G)

where R0 = kl * W * C is the nominal (curvature-driven) migration rate and G(xi) = exp(-2 * Cf * xi / D) weights the upstream curvature history. Depth D is estimated from width via the Konsoer et al. (2013) regression.

Calibration

Starting from a pairwise analysis results dictionary and its classification DataFrame:

import rivabar as rb

calibration = rb.calibrate_segment(results, df, pair_class_filter='good')
print(calibration['Cf_median'], calibration['kl_median'])

Each 'good' pair is calibrated by optimizing Cf (the phase of the migration signal) and amplitude-matching kl; the segment-level result aggregates medians and quartiles. High-variance (cutoff-affected) reaches are excluded from the statistics, while the convolution itself always runs on the full contiguous centerline.

Spatially-varying erodibility

kl often varies along a reach with bank material and vegetation while Cf stays roughly uniform. calibrate_local_kl estimates a kl(s) profile with a sliding window (auto-sized to ~2 meander wavelengths), and calibrate_segment_local aggregates profiles across pairs by DTW-aligning them to a reference centerline:

local_cal = rb.calibrate_segment_local(results, df, pair_class_filter='good')

Forward prediction and validation

# forward prediction with uncertainty envelope
prediction = rb.predict_forward(rivers[-1], calibration, prediction_years=20)

# spatially-varying version
prediction = rb.predict_forward_local(rivers[-1], local_cal, prediction_years=20)

# hold out the most recent scenes and test the model on them
cv = rb.temporal_cross_validate(results, df, n_holdout=3, use_forward_model=True)

# neck cutoff risk along the most recent centerline
risks = rb.detect_cutoff_risk(rivers[-1], threshold=2.0)