Input data semantics and dimensions for seismic data

[jayantb1019]

Hey Liu

I really like your work. It's a great contribution to improving AI capabilities in the seismology domain.

I'm trying to reproduce this work in the petroleum seismic acquisition domain where a single component trace amplitude is recorded.
I've attached a picture of embeddings I obtained using the following pipeline.

flowchart LR
    trace --> bandpass_filtering
    trace --> resampling 

    resampling & bandpass_filtering --> detrending --> amplitude_to_velocity_conversion

    amplitude_to_velocity_conversion --> hilbert_space_estimation --> n_component 
    hilbert_space_estimation --> e_component 

    
    n_component & e_component & trace --> trace_3c
    trace_3c --> standardisation

    m1[("`wav2vec 
    random_init`")]

    m2[("`wav2vec 
    pretrained`")]

    standardisation --> m1 
    standardisation --> m2 

    t1((t-sne))

    m1 -->  t1 --> embeddings_wav2vec_random_init
    m2 --> t1 --> embeddings_wav2vec_pretrained

    detrending --> standardisation2 -->|detrended trace|t1 --> embeddings_trace

    embeddings_wav2vec_random_init & embeddings_wav2vec_pretrained & embeddings_trace --> v1["`
    **Visual Comparison**
    good vs dead
    good vs reverse
    good vs noise
    good vs powerline
    all classes
    `"]

    class_labels --> v1

Loading

. The colors indicate different classes of seismic trace contaminations.

Could you tell me if I need to convert the amplitude information per trace to velocity / acceleration before I extract features using the pre-trained SeisLM ?

Also, could you confirm if your input to conv encoder during pre-training & downstream tasks, had 12000 samples per trace ( 120s signal sampled at 100 hz ) ? Or was it randomly sampled 3001 samples per trace as shown in the pretraining example notebook?

Have you analysed the effects of change in embedding distribution with input length L?