GeoBrain/blending-noise-suppression-avo

AVO Blending Noise Suppression Benchmark

This task performs supervised deblending on paired AVO common-receiver gathers.

Task

The model directly learns a mapping from pseudo-deblended input to the clean reference:

denoised = model(pseudo_deblended_input)

This is a paired regression task, not a synthetic noise-label regression task.

Dataset

Current paired SEG-Y inputs:

• input: /root/Desktop/data/T03_pseudo_deblended_seismic_common_receiver_mod.sgy
• target: /root/Desktop/data/T03_seismic_common_receiver.sgy

Logical volume shape:

• (120, 1001, 1500)
  • 120 common-receiver gathers
  • 1001 traces per common-receiver gather
  • 1500 time samples per trace

Split policy:

• sequential receiver-gather split
• train: 96
• val: 12
• test: 12

Models

• unet
• res_unet
• dncnn
• atten_unet

All model definitions are loaded from model/blending_noise_suppression/.

Preprocessing

• paired loading with shape consistency check
• shared normalization statistics for input and target
• max_abs normalization
• per-gather normalization through the shared shot normalization scope
• overlapping 2D patches of size 128 x 256
• overlap ratio 0.5
• FB-FRE uses dt: 0.004 and adaptive tapering via taper_width: null

Scripts

Training:

• bash scripts/blending_noise_suppression_avo/train_denoise_unet.sh
• bash scripts/blending_noise_suppression_avo/train_denoise_res_unet.sh
• bash scripts/blending_noise_suppression_avo/train_denoise_dncnn.sh
• bash scripts/blending_noise_suppression_avo/train_denoise_atten_unet.sh

Each training script runs a fixed three-seed sweep:

• seeds: 42, 43, 44
• run names: <experiment.name>_T03_avo_mod_seed<seed>_T03_pseudo_deblended_seismic_common_receiver_mod
• checkpoints: /root/Desktop/data/results/blending_noise_avo/<model>/<run_name>/checkpoints/

Inference:

• bash scripts/blending_noise_suppression_avo/inference_denoise_unet.sh
• bash scripts/blending_noise_suppression_avo/inference_denoise_res_unet.sh
• bash scripts/blending_noise_suppression_avo/inference_denoise_dncnn.sh
• bash scripts/blending_noise_suppression_avo/inference_denoise_atten_unet.sh

Each inference script matches the same three seeds, writes per-seed outputs to:

• /root/Desktop/data/results/blending_noise_avo/<model>/<run_name>/inference_integrated_metrics

and also saves a cross-seed aggregate summary to:

• /root/Desktop/data/results/blending_noise_avo/<model>/<experiment.name>_T03_avo_mod_seed_stats/metrics_summary_mean_std.json

Run all:

• bash scripts/blending_noise_suppression_avo/run_all_blending_models.sh

Upload

Dataset upload helper:

• bash scripts/blending_noise_suppression_avo/upload_dataset_to_hf.sh

This uploads:

• input/T03_pseudo_deblended_seismic_common_receiver_mod.sgy
• target/T03_seismic_common_receiver.sgy
• generated dataset card README.md

Model upload helper:

• bash scripts/blending_noise_suppression_avo/upload_blending_model_to_hf.sh

This uploads:

• models/<model>/T03_avo_mod_seed<seed>/best.pt
• models/<model>/T03_avo_mod_seed<seed>/config.yaml
• generated model card README.md

Outputs

Each inference directory contains:

• metrics_per_gather.csv
• metrics_per_sample.csv
• metrics_summary.json
• visualizations/
• optional npy/

Metrics are reported for:

• input vs target
• denoised vs target
• delta = denoised minus input

For naming consistency:

• metrics_summary.json uses input, denoised, and delta
• the file also keeps a backward-compatible restored alias
• metrics_per_sample.csv provides a generic sample-level view, while metrics_per_gather.csv keeps the gather-specific naming

Results

Mean +- std over available seeds, computed from *_seed_stats/metrics_summary_mean_std.json. Metrics are reported on common-receiver gathers in the normalized domain.

Dataset Variant T03_avo_mod

Method	Parameters (M)	SNR	PSNR	SSIM	MAE	MSE	RMSE	EB_WSE_MEDIUM_40_70_NE	EB_WSE_MEDIUM_40_70_SNR	EB_WSE_STRONG_70_100_NE	EB_WSE_STRONG_70_100_SNR	EB_WSE_VERY_WEAK_5_20_NE	EB_WSE_VERY_WEAK_5_20_SNR	EB_WSE_WEAK_20_40_NE	EB_WSE_WEAK_20_40_SNR	FB_FRE_HIGH_ENERGY_RATIO	FB_FRE_HIGH_FREQUENCY_RANGE_HZ	FB_FRE_HIGH_NE	FB_FRE_HIGH_SNR	FB_FRE_LOW_ENERGY_RATIO	FB_FRE_LOW_FREQUENCY_RANGE_HZ	FB_FRE_LOW_NE	FB_FRE_LOW_SNR	FB_FRE_MID_ENERGY_RATIO	FB_FRE_MID_FREQUENCY_RANGE_HZ	FB_FRE_MID_NE	FB_FRE_MID_SNR	FB_FRE_VERY_HIGH_ENERGY_RATIO	FB_FRE_VERY_HIGH_FREQUENCY_RANGE_HZ	FB_FRE_VERY_HIGH_NE	FB_FRE_VERY_HIGH_SNR
Input	-	-2.0292+-0.0000	23.1464+-0.0000	0.7028+-0.0000	0.025431+-0.000000	0.005029+-0.000000	0.070259+-0.000000	16.207542+-0.000000	-24.1817+-0.0000	0.211915+-0.000000	13.5116+-0.0000	131.715259+-0.000000	-42.3526+-0.0000	54.756158+-0.000000	-34.7438+-0.0000	0.140025+-0.000000	48.25-73.3	1.263001+-0.000000	-2.0278+-0.0000	0.125627+-0.000000	6.5-23.2	1.251054+-0.000000	-1.9452+-0.0000	0.637894+-0.000000	23.2-48.25	1.265097+-0.000000	-2.0422+-0.0000	0.001545+-0.000000	73.3-90	1.300019+-0.000000	-2.2787+-0.0000
UNet	7.76	19.8927+-0.6858	45.0683+-0.6858	0.9886+-0.0012	0.003580+-0.000311	0.000033+-0.000005	0.005644+-0.000456	0.924552+-0.029280	0.6885+-0.2748	0.081971+-0.006676	21.7567+-0.6961	4.448344+-0.984280	-12.7468+-1.8217	2.109300+-0.386105	-6.3407+-1.5137	0.140025+-0.000000	48.25-73.3	0.110440+-0.012523	19.2170+-0.9686	0.125627+-0.000000	6.5-23.2	0.128963+-0.003627	17.7975+-0.2436	0.637894+-0.000000	23.2-48.25	0.074045+-0.006893	22.6466+-0.7862	0.001545+-0.000000	73.3-90	0.672353+-0.045598	3.5418+-0.5725
DnCNN	0.56	21.1427+-0.0832	46.3184+-0.0832	0.9879+-0.0003	0.003212+-0.000035	0.000024+-0.000000	0.004882+-0.000046	1.041794+-0.009072	-0.3360+-0.0752	0.046714+-0.000249	26.6125+-0.0459	6.884248+-0.115049	-16.6983+-0.1451	3.056675+-0.050152	-9.6611+-0.1447	0.140025+-0.000000	48.25-73.3	0.084323+-0.001455	21.4974+-0.1497	0.125627+-0.000000	6.5-23.2	0.123886+-0.001142	18.1439+-0.0805	0.637894+-0.000000	23.2-48.25	0.067761+-0.000643	23.3821+-0.0822	0.001545+-0.000000	73.3-90	0.478678+-0.009963	6.4558+-0.1820
ResUNet	8.11	19.1792+-0.4159	44.3549+-0.4159	0.9863+-0.0006	0.004009+-0.000166	0.000038+-0.000004	0.006109+-0.000288	1.172883+-0.031293	-1.3759+-0.2336	0.073572+-0.004656	22.6840+-0.5632	7.385174+-0.356539	-17.3254+-0.4249	3.264928+-0.163859	-10.2494+-0.4437	0.140025+-0.000000	48.25-73.3	0.116582+-0.008488	18.7334+-0.6548	0.125627+-0.000000	6.5-23.2	0.137702+-0.003074	17.2280+-0.1959	0.637894+-0.000000	23.2-48.25	0.073755+-0.002954	22.6513+-0.3527	0.001545+-0.000000	73.3-90	0.796460+-0.049241	2.0905+-0.5491
Attention UNet	7.85	20.5846+-0.4090	45.7603+-0.4090	0.9900+-0.0004	0.003216+-0.000116	0.000028+-0.000003	0.005203+-0.000255	0.851327+-0.007454	1.3988+-0.0763	0.077606+-0.004972	22.2240+-0.5415	3.095065+-0.207962	-9.7852+-0.5872	1.551991+-0.078837	-3.8028+-0.4482	0.140025+-0.000000	48.25-73.3	0.098203+-0.006586	20.1950+-0.5752	0.125627+-0.000000	6.5-23.2	0.122955+-0.002681	18.2104+-0.1877	0.637894+-0.000000	23.2-48.25	0.070269+-0.003938	23.0786+-0.4780	0.001545+-0.000000	73.3-90	0.584715+-0.024681	4.7398+-0.3712