Classification of Small Sample Nuclear Explosion Seismic Events based on MSSA–XGBoost

Abstract

The classification and distinction between nuclear explosions and natural earthquake events are essential to the Comprehensive Nuclear Test Ban Treaty. Nuclear explosion data are lacking; thus, classification problems must be studied in small sample scenarios. The classification problem of the eXtreme Gradient Boosting (XGBoost) model in one small sample scenario is examined using the sparrow search algorithm (SSA) algorithm to optimize the key hyperparameters of the model automatically. The shortcomings of SSA are addressed by using a Gaussian chaotic mapping method, introducing a population proportion dynamic adjustment strategy, and proposing a step-size adjustment factor for modification. The problem of the uneven initial population distribution is addressed by constructing the (modified SSA) MSSA–XGBoost classification model, thereby reducing population diversity and affecting the convergence speed of the algorithm. The fixed proportion problem of the sparrow population, which easily falls into the local optimal solution, is solved using the aforementioned approach. The fixed update step position of the discoverer is also resolved, thus limiting the global search capability and optimization efficiency of the algorithm and realizing the independent optimization of three important hyperparameters. Furthermore, artificial feature extraction can be avoided using this approach, and the number of iterations, maximum tree depth, and learning rate can be automatically optimized, achieving excellent results in small sample seismic event classification. Experimental results reveal that the classification accuracy of the MSSA–XGBoost model is 96.37%, demonstrating its superiority to the original model (93.47%) as well as to the support vector machine and convolutional neural network. Meanwhile, a nearly 30% improvement is observed in computational efficiency.