Contribution

DOPS: Drone Optimized Performance Score for Evaluating Real-Time Tomato Ripeness Detection

---

Abstract

In recent years, deep learning (DL) has emerged as a promising tool to detect ripeness or diseases in different types of plants, which helps farmers monitor crop health and determine the optimal harvest times. However, a significant challenge is the integration of these DL models into drones (UAVs) due to low onboard computing capacity, forcing the images captured by UAV cameras to be transmitted to groundbased processors, introducing delays relying on wireless data transmission that compromise real-time identification and affect the accuracy and efficiency of real-life classification. In this study, we present a new metric called Drone Optimized Performance Score (DOPS) to optimize the performance of real-time Tomato Ripeness Detection, taking into consideration accuracy, frames per second (FPS), and latency. We use a systematic methodology where our research includes an approach in the model training phases and also in the deployment phase of two CNN models, MobileNetV2 and ResNet50, with a main focus on evaluating key performance metrics for classification from drones and integrated cameras. Initially, the lighter model MobileNetV2 proves to be more effective for real-time applications based on DOPS evaluation, but after applying a series of optimizations to ResNet50, which is a resourceintensive model, we can maintain its superior accuracy of 98%, but also outperform MobileNetV2 in DOPS evaluation with higher FPS and lower latency, proving that resource-intensive models can also be optimized for real-world deployment.

How to Cite:

Rexhaj, Y., Kasemi, R. & Lammer, L., (2025) “DOPS: Drone Optimized Performance Score for Evaluating Real-Time Tomato Ripeness Detection”, ARW Proceedings 25(1), 43-48. doi: https://doi.org/10.34749/3061-0710.2025.7