Improved helicopter classification via deep learning and overlapped range-doppler maps

Abstract

Radar target classification is the process of identifying and categorizing objects based on their radar echoes. This is an important task in a variety of fields, including air defense, air traffic control. One of the main reasons why radar target classification is important is because it allows for the effective detection and tracking of objects. This is particularly important in military defense, where it is necessary to be able to identify and track enemy targets. By accurately classifying radar targets, it is possible to determine the type and number of objects that are present, which can inform decision making and allow for the development of effective countermeasures. There are several different methods that have been developed for radar target classification, which can be broadly grouped into five categories based on the input used by the classification algorithms. These categories include micro-doppler signatures, High Resolution Range Profiling, image-based approaches, kinematic feature-based approaches (such as RCS, velocity, and position), and RD maps (Range-Doppler maps). Additionally, the outputs of these approaches can also be combined using probabilistic methods. In this thesis, we focus on RD maps and propose a method for classifying helicopter targets using radar data and a 2D overlapping range-Doppler map approach with a GRU network. Range-Doppler (RD) maps are a commonly used tool in radar target classification, as they provide information about the range and velocity of a target. RD maps are created by taking the Fourier transform of the radar echo signal and plotting the resulting data in a two-dimensional grid. The range of the target is represented on the x-axis and the velocity of the target is represented on the y-axis, and the amplitude can be seen in the z-axis. There are several advantages to using RD maps for radar target classification. One of the main benefits is that they provide detailed information about the movement and characteristics of a target. This can be particularly useful when trying to distinguish between different types of targets, as different targets will have unique RD map signatures. For example, a helicopter will have a different RD map signature than a car or a plane, which allows for accurate classification of the target. Recurrent neural networks (RNNs) are a type of neural network that are particularly well-suited for processing sequential data. One type of RNN, known as a gated recurrent unit (GRU), is particularly effective at handling smaller datasets. GRUs have a simpler structure compared to other types of RNNs, such as long short-term memory (LSTM) networks.They are able to effectively capture long-term dependencies in the data, which is often a challenge when working with smaller datasets. This is because smaller datasets often have less data available to learn from, which can make it difficult for the model to learn complex patterns in the data. By allowing the model to capture long-term dependencies, they are able to better capture the underlying structure of the data, which can help to improve the performance of the model. Moreover, like LSTMs, they GRUs are able to handle variable-length input sequences which is suitable for radar classification since RD map sizes are not the same. Moreover, the overlapped window is a method used in data processing and analysis, particularly in the field of deep learning. It involves dividing a dataset into smaller segments or "windows," and then overlaying these segments on top of one another, with some overlap between them. This overlap allows for the incorporation of additional context and information in the data, which can be useful for tasks such as classification and prediction. One of the main benefits of using overlapped windows is that it allows for the inclusion of more temporal or spatial information in the data. For example, in the field of natural language processing, an overlapped window approach can be used to process longer texts or sequences of words, allowing for the inclusion of more context and meaning in the analysis. In image processing, overlapped windows can be used to analyze images at different scales, incorporating information about the overall structure as well as fine details. Another advantage of overlapped windows is that they can improve the robustness and generalizability of the model being used. By incorporating more information and context in the analysis, the model is able to better handle variations and deviations in the data, leading to better performance on unseen or out-of-sample data. This can be particularly useful in cases where the dataset is small or imbalanced, as is often the case in real-world applications. Furthermore, overlapped windows are often used in time series analysis and other tasks that involve sequential data. The idea is to divide the data into overlapping segments or "windows" of a fixed size, and then use these windows as input to a deep learning model. Thus, it can improve the accuracy of the classification, as more of the relevant information about the target is retained. Additionally, the use of overlapping windows allows for the classification of RD maps regardless of the range resolution, as the window size can be adjusted to fit the size of the targets. This makes the method highly flexible and adaptable to different scenarios. We demonstrate that this approach is effective in accurately identifying different types of helicopters using various radar systems. The use of an overlapped window model allows for the incorporation of spatial information in the data, while the GRU network is able to handle smaller datasets and preserve more of the relevant information about the target. Our results indicate that the proposed method outperforms traditional approaches in terms of accuracy and robustness. This method is highly flexible and adaptable to different scenarios, as it is able to classify radar data of varying sizes. The overlapped multicell GRU method for radar target classification was found to have the highest helicopter classification accuracy on data from radars A and B, respectively, in the first experiment. In the second experiment, it also achieved a better helicopter classification accuracy. Moreover, evaluation metrics such as the F1 score and Mathew Correlation Coefficient (MCC) were also calculated, and the results are in the same direction as accuracies. Overall, the overlapped multicell GRU method was found to perform better compared to the other methods tested.