GeoBrain/random-noise-attenuation-avo

Random Noise Suppression AVO Benchmark

Deep-learning-based random-noise attenuation on pre-stack seismic shot gathers, using the clean seismic.segy AVO dataset and synthetic Gaussian / Poisson noise injection.

Task

Given a clean shot gather, the benchmark first injects synthetic random noise at a specified SNR, then trains a model to directly reconstruct the clean signal:

denoised = model(noisy_input)

This is a paired regression task with clean-target supervision. Most models directly predict the clean gather. The DDPM variant predicts diffusion noise during training and reconstructs the clean gather through reverse sampling at evaluation / inference time.

Dataset

• Source: Clean AVO seismic.segy data
• Current training volume: seismic.segy
• Geometry: 201 traces per shot, time sampling interval dt = 2 ms
• Split: Shot-level sequential 7:1:1
  • 801 training shots
  • 100 validation shots
  • 100 held-out test shots

Synthetic Noise Settings

• Noise kinds: gaussian, poisson
• Default sweep in training / inference scripts: SNR = -5, 0, 5 dB
• Noise injection: per-shot variance-controlled synthetic corruption
• Reproducibility: noise generation is seeded from experiment.seed

Model Architectures

• UNet (unet): classic encoder-decoder with skip connections. Base channels: 32, depth: 4.
• ResUNet (res_unet): U-Net with residual blocks. Base channels: 32, depth: 4.
• DnCNN (dncnn): residual denoising CNN with 17 layers and 64 feature channels.
• Attention UNet (atten_unet): U-Net with attention gates. Base channels: 32, depth: 4.
• DDPM (ddpm): standard conditional denoising diffusion probabilistic model with reverse sampling from Gaussian noise to the clean shot gather.
• SCRN (SCRN): Swin Transformer convolutional residual network adapted to the same random-noise benchmark pipeline.

Preprocessing

• Amplitude correction: spherical divergence correction is skipped by default
• Normalization: max_abs, per-shot
• Patching: overlapping 2D patches of size 128 x 256 (trace x time)
• Patch overlap: 50%

Training uses patched shot gathers. Inference reloads the raw volume, applies inference.shot_split, injects synthetic noise, runs patch-based reconstruction, and inverse-normalizes outputs for visualization.

Repository Structure

scripts/random_noise_suppression_avo/
|- train_denoise_unet.sh
|- train_denoise_res_unet.sh
|- train_denoise_dncnn.sh
|- train_denoise_atten_unet.sh
|- train_denoise_ddpm.sh
|- train_denoise_SCRN.sh
|- inference_denoise_unet.sh
|- inference_denoise_res_unet.sh
|- inference_denoise_dncnn.sh
|- inference_denoise_atten_unet.sh
|- inference_denoise_ddpm.sh
|- inference_denoise_SCRN.sh
`- run_all_random_noise_models.sh

configs/random_noise_suppression_avo/
|- denoise_unet.yaml
|- denoise_res_unet.yaml
|- denoise_dncnn.yaml
|- denoise_atten_unet.yaml
|- denoise_ddpm.yaml
`- denoise_SCRN.yaml

Each experiment directory is named by model, noise kind, SNR, and seed, for example:

random_noise_avo_unet_base_gaussian_snr5_seed42/
random_noise_avo_dncnn_base_poisson_snr0_seed43/
random_noise_avo_ddpm_base_gaussian_snrneg5_seed44/

Training Details

Shared benchmark defaults:

Hyperparameter	Value
Loss	MSE
DDPM note	trains on diffusion-noise prediction and validates clean reconstruction after reverse sampling
Optimizer	AdamW (lr=1e-4, weight_decay=1e-5)
Scheduler	Cosine annealing (min_lr=1e-6)
Epochs	200
Gradient clipping	1.0
Seeds	42, 43, 44 by default in shell sweeps
Batch size	192 in current YAML defaults

Usage

Train One Model Family

bash scripts/random_noise_suppression_avo/train_denoise_unet.sh

or

bash scripts/random_noise_suppression_avo/train_denoise_ddpm.sh

Each training shell script sweeps:

• noise kind
• SNR
• seed

by rewriting a temporary YAML config before calling torchrun.

Run Inference

bash scripts/random_noise_suppression_avo/inference_denoise_unet.sh

Inference outputs:

• per-shot metrics CSV
• summary JSON
• visualizations
• optional .npy files
• multi-seed mean/std aggregation JSON

Run All Model Families

bash scripts/random_noise_suppression_avo/run_all_random_noise_models.sh

Current total-run script executes:

1. unet
2. dncnn
3. res_unet
4. atten_unet
5. ddpm

Each model is trained first, then its inference sweep is launched immediately after training finishes.

Inference Outputs

For each experiment, the inference directory typically contains:

inference/
|- inference.log
|- metrics_per_shot.csv
|- metrics_summary.json
|- visualizations/
`- npy/                    # only when save_npy=true

Metrics

The benchmark reports:

• snr
• psnr
• ssim
• mae
• mse
• rmse

for three groups:

• noisy: noisy input vs clean target
• denoised: model prediction vs clean target
• delta: denoised - noisy

Metrics are computed in the normalized domain. Saved visualization outputs are inverse-normalized back to the original amplitude domain.

Notes

• This benchmark injects synthetic random noise once per experiment before patch extraction; it is not an epoch-wise dynamic noise augmentation setup.
• Shot-level inference uses the held-out test shot defined by inference.shot_split.
• Batch size and patch size may need adjustment for memory-heavy models such as DnCNN and DDPM.

References

• Ronneberger et al., U-Net: Convolutional Networks for Biomedical Image Segmentation, MICCAI 2015
• He et al., Deep Residual Learning for Image Recognition, CVPR 2016
• Zhang et al., Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising, IEEE TIP 2017
• Ho et al., Denoising Diffusion Probabilistic Models, NeurIPS 2020
• Song et al., Denoising Diffusion Implicit Models, ICLR 2021
• Oktay et al., Attention U-Net: Learning Where to Look for the Pancreas, MIDL 2018
• Gao et al., Swin Transformer for simultaneous denoising and interpolation of seismic data, Computers and Geosciences 2024
• SEG C3 Velocity Model: https://wiki.seg.org/wiki/C3

Results

Mean +- std over available seeds, computed from *_seed_stats/metrics_summary_mean_std.json. Metrics are reported in the normalized domain. Raw (noisy) is the synthetic noisy input before denoising.

Gaussian Noise

SNR -5 dB

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
Raw (noisy)	-	-5.0000+-0.0015	28.4184+-0.0015	0.6663+-0.0002	0.030315+-0.000008	0.001448+-0.000001	0.037994+-0.000007	30.056404+-0.003195	-29.5408+-0.0009	0.974550+-0.000441	0.2240+-0.0039	217.468024+-0.075775	-46.7401+-0.0030	100.131537+-0.019746	-39.9965+-0.0018	0.168114+-0.000000	48.3333-73.5333	1.841567+-0.000122	-5.2710+-0.0006	0.283209+-0.000000	6.3333-23.1333	1.150310+-0.000255	-1.2055+-0.0019	0.425271+-0.000000	23.1333-48.3333	1.147630+-0.000330	-1.1892+-0.0025	0.001317+-0.000000	73.5333-90.3333	18.962028+-0.014011	-25.2465+-0.0056
UNet	7.76	8.9440+-0.0034	42.3624+-0.0034	0.9807+-0.0000	0.004273+-0.000002	0.000058+-0.000000	0.007623+-0.000003	2.416080+-0.000817	-7.6466+-0.0029	0.338792+-0.000149	9.4048+-0.0037	15.867472+-0.007823	-23.9986+-0.0043	7.376587+-0.006652	-17.3406+-0.0071	0.168114+-0.000000	48.3333-73.5333	0.371155+-0.000550	8.6215+-0.0129	0.283209+-0.000000	6.3333-23.1333	0.350323+-0.000279	9.1190+-0.0065	0.425271+-0.000000	23.1333-48.3333	0.308464+-0.000132	10.2247+-0.0039	0.001317+-0.000000	73.5333-90.3333	1.857709+-0.001298	-5.2139+-0.0053

SNR 0 dB

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
Raw (noisy)	-	0.0000+-0.0015	33.4184+-0.0015	0.8644+-0.0001	0.017047+-0.000004	0.000458+-0.000000	0.021365+-0.000004	16.901958+-0.001796	-24.5408+-0.0009	0.548030+-0.000248	5.2240+-0.0039	122.291257+-0.042611	-41.7401+-0.0030	56.308101+-0.011104	-34.9965+-0.0018	0.168114+-0.000000	48.3333-73.5333	1.035589+-0.000068	-0.2710+-0.0006	0.283209+-0.000000	6.3333-23.1333	0.646867+-0.000144	3.7945+-0.0019	0.425271+-0.000000	23.1333-48.3333	0.645360+-0.000186	3.8108+-0.0025	0.001317+-0.000000	73.5333-90.3333	10.663132+-0.007879	-20.2465+-0.0056
UNet	7.76	10.5129+-0.0048	43.9313+-0.0048	0.9847+-0.0000	0.003534+-0.000002	0.000040+-0.000000	0.006361+-0.000003	1.862257+-0.001138	-5.3904+-0.0051	0.285767+-0.000168	10.8867+-0.0049	11.745703+-0.009011	-21.3908+-0.0068	5.497729+-0.002016	-14.7914+-0.0029	0.168114+-0.000000	48.3333-73.5333	0.295392+-0.000340	10.6085+-0.0097	0.283209+-0.000000	6.3333-23.1333	0.298379+-0.000381	10.5175+-0.0108	0.425271+-0.000000	23.1333-48.3333	0.250014+-0.000159	12.0534+-0.0055	0.001317+-0.000000	73.5333-90.3333	1.658566+-0.001730	-4.2297+-0.0092

SNR 5 dB

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
Raw (noisy)	-	5.0000+-0.0015	38.4184+-0.0015	0.9528+-0.0000	0.009587+-0.000002	0.000145+-0.000000	0.012015+-0.000002	9.504669+-0.001010	-19.5408+-0.0009	0.308180+-0.000139	10.2240+-0.0039	68.769427+-0.023962	-36.7401+-0.0030	31.664372+-0.006244	-29.9965+-0.0018	0.168114+-0.000000	48.3333-73.5333	0.582355+-0.000039	4.7290+-0.0006	0.283209+-0.000000	6.3333-23.1333	0.363760+-0.000081	8.7945+-0.0019	0.425271+-0.000000	23.1333-48.3333	0.362912+-0.000104	8.8108+-0.0025	0.001317+-0.000000	73.5333-90.3333	5.996320+-0.004431	-15.2465+-0.0056
UNet	7.76	12.7151+-0.0046	46.1335+-0.0046	0.9911+-0.0000	0.002776+-0.000001	0.000024+-0.000000	0.004937+-0.000003	1.481680+-0.000344	-3.4066+-0.0018	0.221581+-0.000125	13.0955+-0.0048	9.064477+-0.003318	-19.1393+-0.0034	4.239337+-0.001748	-12.5320+-0.0035	0.168114+-0.000000	48.3333-73.5333	0.239064+-0.000116	12.4555+-0.0039	0.283209+-0.000000	6.3333-23.1333	0.215258+-0.000339	13.3532+-0.0133	0.425271+-0.000000	23.1333-48.3333	0.191701+-0.000072	14.3603+-0.0033	0.001317+-0.000000	73.5333-90.3333	1.533126+-0.002486	-3.5443+-0.0129

Poisson Noise

SNR -5 dB

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
Raw (noisy)	-	-4.9986+-0.0015	28.4198+-0.0015	0.6663+-0.0001	0.030308+-0.000004	0.001447+-0.000000	0.037988+-0.000007	30.055742+-0.008101	-29.5406+-0.0024	0.974447+-0.000121	0.2249+-0.0011	217.357200+-0.102979	-46.7357+-0.0040	100.096922+-0.017429	-39.9936+-0.0012	0.168114+-0.000000	48.3333-73.5333	1.841246+-0.000500	-5.2697+-0.0024	0.283209+-0.000000	6.3333-23.1333	1.149769+-0.000100	-1.2014+-0.0009	0.425271+-0.000000	23.1333-48.3333	1.147469+-0.000565	-1.1882+-0.0043	0.001317+-0.000000	73.5333-90.3333	18.965560+-0.004406	-25.2484+-0.0018
UNet	7.76	8.9857+-0.0031	42.4041+-0.0031	0.9808+-0.0000	0.004256+-0.000000	0.000058+-0.000000	0.007587+-0.000003	2.415076+-0.002099	-7.6446+-0.0074	0.337022+-0.000162	9.4499+-0.0042	15.838296+-0.020752	-23.9837+-0.0112	7.364007+-0.012769	-17.3268+-0.0150	0.168114+-0.000000	48.3333-73.5333	0.369695+-0.000207	8.6562+-0.0048	0.283209+-0.000000	6.3333-23.1333	0.346966+-0.000291	9.2024+-0.0073	0.425271+-0.000000	23.1333-48.3333	0.307719+-0.000289	10.2453+-0.0085	0.001317+-0.000000	73.5333-90.3333	1.854940+-0.000681	-5.2019+-0.0030

SNR 0 dB

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
Raw (noisy)	-	0.0023+-0.0012	33.4206+-0.0012	0.8645+-0.0000	0.017041+-0.000002	0.000458+-0.000000	0.021360+-0.000003	16.897970+-0.004420	-24.5386+-0.0022	0.548122+-0.000078	5.2225+-0.0012	122.192401+-0.050008	-41.7331+-0.0035	56.272479+-0.003774	-34.9910+-0.0005	0.168114+-0.000000	48.3333-73.5333	1.035247+-0.000049	-0.2683+-0.0004	0.283209+-0.000000	6.3333-23.1333	0.646679+-0.000100	3.7971+-0.0014	0.425271+-0.000000	23.1333-48.3333	0.645245+-0.000461	3.8120+-0.0062	0.001317+-0.000000	73.5333-90.3333	10.666173+-0.001591	-20.2493+-0.0019
UNet	7.76	10.2685+-0.0045	43.6869+-0.0045	0.9836+-0.0000	0.003591+-0.000001	0.000043+-0.000000	0.006543+-0.000003	1.889969+-0.000705	-5.5192+-0.0033	0.294167+-0.000159	10.6331+-0.0049	11.981634+-0.005965	-21.5637+-0.0042	5.589865+-0.005678	-14.9357+-0.0088	0.168114+-0.000000	48.3333-73.5333	0.302042+-0.000172	10.4159+-0.0051	0.283209+-0.000000	6.3333-23.1333	0.306526+-0.000292	10.2817+-0.0086	0.425271+-0.000000	23.1333-48.3333	0.251918+-0.000122	11.9870+-0.0046	0.001317+-0.000000	73.5333-90.3333	1.700620+-0.001753	-4.4417+-0.0097

SNR 5 dB

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
Raw (noisy)	-	5.0013+-0.0005	38.4197+-0.0005	0.9528+-0.0000	0.009585+-0.000001	0.000145+-0.000000	0.012013+-0.000001	9.502945+-0.000967	-19.5391+-0.0007	0.308110+-0.000111	10.2259+-0.0031	68.744539+-0.029754	-36.7371+-0.0038	31.665301+-0.014444	-29.9967+-0.0040	0.168114+-0.000000	48.3333-73.5333	0.582273+-0.000219	4.7300+-0.0031	0.283209+-0.000000	6.3333-23.1333	0.363669+-0.000022	8.7967+-0.0007	0.425271+-0.000000	23.1333-48.3333	0.362916+-0.000190	8.8104+-0.0045	0.001317+-0.000000	73.5333-90.3333	5.996074+-0.000881	-15.2461+-0.0005
UNet	7.76	12.6735+-0.0009	46.0919+-0.0009	0.9909+-0.0000	0.002779+-0.000000	0.000025+-0.000000	0.004961+-0.000001	1.481825+-0.000701	-3.4074+-0.0040	0.222778+-0.000029	13.0481+-0.0011	9.008784+-0.003577	-19.0862+-0.0034	4.219501+-0.002957	-12.4921+-0.0066	0.168114+-0.000000	48.3333-73.5333	0.239523+-0.000217	12.4398+-0.0074	0.283209+-0.000000	6.3333-23.1333	0.216420+-0.000178	13.3066+-0.0068	0.425271+-0.000000	23.1333-48.3333	0.191949+-0.000036	14.3481+-0.0016	0.001317+-0.000000	73.5333-90.3333	1.547839+-0.002906	-3.6236+-0.0174