GIS-Integrated Convolutional Neural Network Framework for Real-Time Border Threat Detection

¹ Assistant Professor & CSE(DS), Vignan’s Institute of Management and Technology for Women, Hyderabad, India, Kumarsir1016@gmail.com
² Assistant Professor & CSE(DS), Vignan’s Institute of Management and Technology for Women, Hyderabad, India, chamundeswari@vmtw.in
³ Assistant Professor & CSE (DS), Vignan’s Institute of Management and Technology for Women, Hyderabad, India, vishnusangeeta9@gmail.com
⁴ Associate Professor & CSE (DS), Vignan’s Institute of Management and Technology for Women, Hyderabad, India, nallasreejavmtw@gmail.com
⁵ Associate Professor & CSE (AI&ML), Vignan’s Institute of Management and Technology for Women, Hyderabad, India, ramyasrivmtw@gmail.com
⁶ Assistant Professor & CSE (AI&ML), Vignan’s Institute of Management and Technology for Women, Hyderabad, India, mvvrao.mca31@gmail.com
⁷ B. Tech Graduate, Vignan’s Institute of Management and Technology for Women, Hyderabad, India, srividyakondoju1812@gmail.com

Abstract

The rapid proliferation of inexpensive unmanned aerial vehicles, cross-border smuggling routes, and illicit incursions demands intelligent surveillance that fuses spatial context with automated visual analytics. We present a GIS-integrated Convolutional Neural Network (CNN) framework that ingests live video streams from fixed or mobile cameras, classifies each frame for “threat” versus “no-threat,” and geo-locates positive detections on an interactive map to support command-and-control decisions. Implemented and evaluated entirely in Google Colab, the prototype achieves 93.8 % accuracy, 0.92 precision, and 0.91 recall on a 4,487-frame benchmark compiled from open-source border videos. The system overlays detections on ESRI- compliant shape-files in real time, demonstrating sub-250 ms end-to-end latency per frame. Extensive experiments confirm that spatial filtering (masking detections outside polygonal border zones) reduces false alarms by 27 % without harming recall. The proposed architecture is low- cost, cloud-agnostic, and readily extensible to additional object classes or sensor modalities.