EXPERIMENTAL STUDY OF PROJECTILE CLASSIFICATION USING RANDOM FOREST AND SVC

Abstract

The task of classifying projectiles is crucial for the safe and efficient disposal of ammunition and for planning military operations. This study compares the accuracy of two machine learning models—Random Forest and SVC —in classifying projectiles into six classes (A – (a-ammunition), A/M – (a/m-ammunition), A/P – (armor-piercing), A/RS – (armor-piercing-incendiary), M – (m-ammunition), R – (armor-piercing ammunition)) based on the components of the input feature vector: (x1 - position_x1, x2 - position_y1, x3 - position_h1, x4 – velocity, x5 - taget_class, x6 - explosion_x2, x7 - explosion_y2, x8 - explosion_h2, x9 – hour, x10 – minute, x11 – second, x12 - angle_big_tick, x13 - angle_small_tick, x14 - angle_degrees, x15 - angle_rotation_degrees, x16 - distance_2d, x17 - distance_3d, x18 - flight_time).

The classification using Random Forest was conducted with the following parameters: 100 trees in the ensemble (n_estimators), maximum tree depth (max_depth) of 4, minimum number of samples per leaf (min_samples_leaf) of 1, and minimum number of samples for node splitting (min_samples_split) of 5. The hyperparameters were optimized using GridSearchCV. The study revealed that increasing the number of trees improves model stability, but accuracy stabilizes beyond a certain point. Overly deep trees lead to overfitting, while shallow trees lack sufficient model complexity.

The classification using SVC was performed with an RBF kernel, a regularization coefficient (C) of 1.0, and a γ parameter of 0.1. Optimizing these parameters improved the accuracy of input space separation, ensuring a balance between overfitting and underfitting. The C parameter significantly influences model sensitivity to classification errors, while the optimal γ value enhances the model's ability to differentiate between classes.