Deep Learning based Framework for Real-Time Bird Detection on Jowar Crop in Real Time Environment

Abstract

Bird predation in agricultural farms, particularly in crops such as Jowar (Sorghum) causes high losses in yields, which is a great problem to farmers. Old methods of scaring have become useless, and the surveillance by hand cannot be scaled or real-time. The increasing cases of birds strike and illegal drones in restricted airspaces are threatening the aviation safety and the equilibrium of the ecosystems. To address these difficulties, this study suggests an Internet of Things (IoT) and Deep Learning system of real-time monitoring and detection of birds in changing outdoor conditions. To guarantee the correct identification of birds and other flying objects, this framework also uses lightweight and powerful object detection models YOLOv11n, YOLOv11L,YOLOv8n, and YOLOv8L with a proposed module of Faster R-CNN. The sensors provide the system with field data which is streamed in real time to a cloud deep learning pipeline. Both versions of YOLO bring something to the model in the size, inference speed, and ability to be deployed to the edge. To obtain enhanced accuracy, specific change added emphasis loss by means of stratified sampling method to long-tailed distribution of class and high-resolution region suggestions in response to herding the key points of focus yaw. CNN based YOLO was purposely created to identify dynamic non-stationary objects in our observed world. It is a dynamic IO sensor networks which is anchored using static sensors and that data is sampled and stored as well as transmitted to smart moving nodes having ultra-light flow sensor system. These models had all been trained and tested with our dataset in varying conditions of the environment such as foggy views and our changing and altering lighting and skies filled with moving and stationary objects. The experimental findings in general showed that the single stage YOLO models had a vastly lower detection accuracy in comparison with the proposed Faster R-CNN model. The suggested model had a total real-time field test accuracy of about 98.01%. Simultaneously, both the inference speeds of YOLOv11L and YOLOv8L are competitive and suitable to edge IoT devices. The model is best when used in real-time such as when tracking airports, wildlife and airspace control due to trade-offs between detection latency and accuracy. The current work introduces an effective and versatile architecture of smart bird detection using IoT and deep learning-based methods making a significant contribution to proactive mitigation of the air threats.