Radar-Based Drone Classification Using In-Context Learning in Large Language Models: A Comparative Study of LLMs, CNNs, and SVMs on FMCW Radar-Based Doppler-Range Maps
Translated title
Radar-Based Drone Classification Using In-Context Learning in Large Language Models
Author
Johnsen, Paw Milwertz
Term
4. semester
Education
Publication year
2026
Submitted on
2026-06-03
Pages
86
Abstract
This project examines whether a Large Language Model (LLM) can be used for radar-based drone classification and how it compares to standard machine learning methods. The data consist of Doppler–range maps from frequency-modulated continuous-wave (FMCW) radar, which visualize a target’s speed (Doppler) and distance. The tasks are to detect drone presence and to classify the drone’s model, motion, and cargo amount. The LLM is applied with few-shot in-context learning and prompt engineering, meaning it is guided by a small number of labeled examples provided in the prompt. Its performance is compared to a Convolutional Neural Network (CNN) and a Support Vector Machine (SVM) under three training regimes: full data, limited data, and limited data with augmentation. Results show that the CNN achieved the highest overall accuracy—exceeding 87.6% across parameters—when trained on the full dataset, while the SVM exceeded 75.8%. The LLM was tested with 5-shot and achieved over 90% for motion and presence, but lower accuracies for cargo and model (58.5% and 81%). Analysis of misclassifications suggests most LLM errors stem from ambiguities and overlapping radar signatures in the dataset. Overall, supervised methods (CNN and SVM) delivered the highest accuracies, while the in-context LLM proved able to interpret Doppler–range maps with very little data, making it the most data-efficient approach and a promising option for radar-based drone classification.
Dette projekt undersøger, om en stor sprogmodel (LLM) kan bruges til radarbaseret klassifikation af droner, og hvordan den klarer sig i forhold til mere traditionelle metoder. Datasættet består af Doppler–afstandskort fra frekvensmoduleret kontinuerlig bølge (FMCW) radar, som visualiserer målers hastighed (Doppler) og afstand. Opgaven er at registrere, om en drone er til stede, og at klassificere dens model, bevægelse og mængden af last. LLM’en anvendes med few-shot in-context læring og prompt-design, hvor modellen får vist få mærkede eksempler i selve prompten. Dens ydeevne sammenlignes med et konvolutionelt neuralt netværk (CNN) og en supportvektormaskine (SVM) under tre datascenarier: fulde data, begrænsede data og begrænsede data med augmentering. Resultaterne viser, at CNN’en opnår den højeste samlede nøjagtighed, over 87,6% på tværs af parametrene, når den trænes på hele datasættet, mens SVM’en når over 75,8%. LLM’en blev testet med 5-shot og opnåede over 90% for bevægelse og tilstedeværelse, men lavere for last og model (58,5% og 81%). Analyse af fejlklassifikationer peger på, at de fleste LLM-fejl skyldes tvetydige eller overlappende radarsignaturer i datasættet. Samlet set gav de superviserede metoder (CNN og SVM) den højeste nøjagtighed, mens LLLM’en med in-context læring viste evne til at tolke Doppler–afstandskort med få eksempler, hvilket gjorde den mest dataeffektive og en lovende tilgang til radarbaseret droneklassifikation.
[This apstract has been rewritten with the help of AI based on the project's original abstract]
Keywords