Land cover recognition from few samples of radar high-resolution range profile

Abstract

Radar echo data, a refined representation of detected targets, is gaining attention in electronic warfare, marine environment monitoring, and agriculture intelligence for land cover recognition. In contrast to prevailing studies involving two-dimensional images such as synthetic aperture radar (SAR) or inverse SAR (ISAR) images, one-dimensional High-Resolution Range Profile (HRRP) data offers advantages of easy access and simple processing. Nonetheless, its potential application has yet to be explored, as current research on it remains insufficient. Meanwhile, deep learning methods that specialize in classification tasks but encounter challenges in modern electronic warfare, given their heavy reliance on the number of labeled samples. To tackle these issues, a feature fusion-based land cover recognition approach is proposed, which introduces Convolutional Embedding Sequence Encoder (CE-SE) to capture the complex clutter characteristics of HRRP, achieving land cover recognition with a small number of labeled HRRP samples and eliminating the reliance on two-dimensional image data. Experimental results validated on Mini-SAR data from unmanned aerial vehicles demonstrate effective land cover recognition using HRRP. This method significantly improves the recognition accuracy for five land cover types, exceeding 92%, even with only half the sample size compared to traditional deep learning methods. Additionally, the model's inference time for a single HRRP sample is just 1.3 milliseconds, demonstrating its capability for real-time land cover recognition.