Table 1 Configuration parameters for Spatiotemporal feature extraction components.

From: Spatiotemporal graph neural networks for analyzing the influence mechanisms of river hydrodynamics on microplastic transport processes

Feature type

Extraction method

Parameter settings

Performance metrics

Short-term velocity

1D Temporal Conv

Kernel = 3, Stride = 1, Channels = 64

R² = 0.892, RMSE = 0.045

Long-term flow

Dilated Conv

Dilation = 4, Kernel = 5, Channels = 32

R² = 0.847, RMSE = 0.067

Spatial connectivity

Graph Conv

Neighbors = 8, Hidden = 128

Accuracy = 0.913, F1 = 0.856

Turbulence features

Multi-head Attention

Heads = 8, Dim = 256

Correlation = 0.789

Particle concentration

LSTM-GCN

Hidden = 64, Layers = 2

MAE = 0.023, MAPE = 12.4%

Transport velocity

Residual Block

Filters = 96, Dropout = 0.2

R² = 0.871, Bias = 0.012

Settling dynamics

Attention-Conv

Window = 12, Alpha = 0.3

Precision = 0.834

Resuspension events

Threshold detection

Tau = 0.15, Beta = 2.1

Recall = 0.902, AUC = 0.917

  1. The attention mechanism implementation is theoretically grounded in the principle of scale-dependent transport processes34, where the attention weight \(\:{\alpha\:}_{\left\{t,s\right\}}\) for time step t and scale s is computed as:
Back to article page