Inversion of large-scale time-domain transient electromagnetic (TEM) surveys is computationally expensive and time-consuming. The calculation of partial derivatives for the Jacobian matrix is by far the most computationally intensive task, as this requires calculation of a significant number of forward responses. We propose to accelerate the inversion process by predicting partial derivatives using an artificial neural network. Network training data for resistivity models for a broad range of geological settings are generated by computing partial derivatives as symmetric differences between two forward responses. Given that certain applications have larger tolerances for modeling inaccuracy and varying degrees of flexibility throughout the different phases of interpretation, we present four inversion schemes that provide a tunable balance between computational time and inversion accuracy when modeling TEM datasets. We improve speed and maintain accuracy with a hybrid framework, where the neural network derivatives are used initially and switched to full numerical derivatives in the final iterations. We also present a full neural network solution where neural network forward and derivatives are used throughout the inversion. In a least-squares inversion framework, a speedup factor exceeding 70 is obtained on the calculation of derivatives, and the inversion process is expedited ~36 times when the full neural network solution is used. Field examples show that the full nonlinear inversion and the hybrid approach gives identical results, whereas the full neural network inversion results in higher deviation but provides a reasonable indication about the overall subsurface geology.

Geophysical modelling and data inversion are important tools for interpreting the physical properties of Earth’s subsurface. Solving the inverse problem involves several computational steps and is generally a time consuming task. Artificial neural networks have the potential to speed up large computations. Such networks provide the means to model the relationship between the inputs and outputs without needing to know the physical model of the underlying problem. There are two main aspects that affect the performance of neural networks: optimization of network architecture and pre-processing of data. In this article, we investigate several traditional pre-processing techniques including the min-max scaling, z-score scaling, and the logarithmic transform scaling, and propose some novel data pre-processing approaches for the 1-D forward modelling of time-domain electromagnetic data based on signal characteristics. We evaluate the performance of the conventional and the proposed pre-processing methods against a 3% relative error metric, which corresponds to the typical data uncertainty, to show that forward data pre-processing has significant effect on the performance of neural networks. The proposed gate-wise min-max scaling achieves the best performance with 96% of gates within a 3% relative error, while the commonly used logarithmic transform results only in 75% of gates within a 3% relative error. We provide insights into how various pre-processing methods affect the performance of these networks and recommend optimal pre-processing strategies that may be used where similar data content is encountered to achieve superior performance. Finally, we show the effect of forward modelling accuracy in inverse modelling.

Airborne time-domain electromagnetic surveys produce extremely large data sets with thousands of line kilometers of data and millions of possible models to explain the data. Inversion of such data sets to obtain the resistivity structures of the subsurface is computationally intensive and involves calculation of a significant number of forward and derivative responses for solving the least-squares inverse problem. The flight altitude of the airborne system needs to be included in the modeling, which adds further complexity. We propose to integrate neural networks in a damped iterative least-squares inversion framework to expedite the inversion process. We train two separate neural networks to predict the forward responses and partial derivatives independently for a broad range of resistivity structures and flight altitudes. Data inversion is not only used for producing the final subsurface models but also used during data processing, or to produce intermediate results during a survey. With these purposes in mind, we provide three inversion schemes with a tunable balance between computational time and modeling accuracy: (1) numerical forward responses used initially in combination with neural network derivatives, and the derivatives switched to a numerical solution in final iterations, (2) numerical forward responses in combination with neural network derivatives used throughout the inversion, and (3) only neural network forward responses and derivatives used in inversion. Experiments on field data find that we improve inversion speed without any loss in modeling accuracy with our first approach, whereas the second scheme gives a significant speedup at the cost of minor and often acceptable deviations in the inversion results from the conventional nonlinear inversion. The last approach is the fastest and captures the overall resistivity structures quite well. Therefore, depending on the modeling accuracy, inversion speedup factors of up to 50 are realized by using the proposed schemes.

License plate recognition system (LPR) plays a vital role in intelligent transport systems to build up smart environments. Numerous country specific methods have been proposed successfully for an LPR system, but there is a need to find a generalized solution that is independent of license plate layout. The proposed architecture is comprised of two important LPR stages: (i) License plate character segmentation (LPCS) and (ii) License plate character recognition (LPCR). A foreground polarity detection model is proposed by using a Red-Green-Blue (RGB) channel-based color map in order to segment and recognize the LP characters effectively at both LPCS and LPCR stages respectively. Further, a multi-channel CNN framework with layer aggregation module is proposed to extract deep features, and support vector machine is used to produce target labels. Multi-channel processing with merged features from different-level convolutional layers makes output feature map more expressive. Experimental results show that the proposed method is capable of achieving high recognition rate for multinational vehicles license plates under various illumination conditions.