Unsupervised Deep Seismic Pursuit: Recovering the Compressed Seismic Data Without Ground Truth

Future oil and gas exploration demands large-scale seismic data acquisition with increased versatility, flexibility, automation, and scalability. The challenge lies in transmitting hundreds of recordings per geophone through narrowband channels without overloading the data center or the geophones themselves. This is where lightweight and stand-alone compressive sensing (CS) methods become crucial. This work proposes an efficient in-field seismic data compression approach that uses a sparsity-aware geophone and combines it with an unsupervised deep neural network-based method for recovery. Two knowledge-guided algorithms are used and compared in this regard. These model-based signal reconstruction frameworks with a convolutional neural network (CNN)-based regularization prior offer a systematic approach for optimizing deep architectures for reconstruction without ground truth labels. The formulation reduces the demand for training data and training time while providing improved performance over traditional approaches. During the learning process, alternating minimization is employed in an unrolled iterative framework to optimize both the neural network weights and the signal reconstruction. A key strength of this approach is that it is general within the CS framework and works with any underlying seismic data statistics, allowing it to adapt to a wide range of exploration scenarios. Our results on real seismic datasets demonstrate that the proposed unsupervised method achieves a good balance between data compression ( 16×) and signal quality [ ≈ -22 dB normalized mean square error (NMSE)]. This innovation holds promise not only for oil and gas exploration but also for applications, such as geothermal energy exploration and hydroelectric power, where efficient and reliable identification is essential for optimizing resource exploration, development, and safer operations.