Real-Time Broccoli Flower Detection through Edge Computing On Aerial Multispectral Imagery

XXX-X-XXXX-XXXX-X/XX/$XX.00 © 20XX IEEE Real-Time Broccoli Flower Detection through Edge-Computing On Aerial Multispectral Imagery Hsin-Hung Tseng Department of Civil Engineering National Chung Hsing University Taichung City, Taiwan [email protected] Ming-Der Yang Department of Civil Engineering National Chung Hsing University Taichung City, Taiwan [email protected] Abstract—This study aims to develop a method for counting and estimating the optimal harvest time of broccoli flowers based on their maturity level using a CNN-based model on multispectral imagery. Broccoli is a widely consumed vegetable with high nutritional value and easy cultivation, storage, and cooking. However, broccoli plants have different sizes and growth stages at the same time, which makes it difficult for farmers to decide when to harvest them for maximum profit. This study presents a complete procedure for processing raw multispectral imagery, including distortion rectification, band alignment, irradiancereflectance conversion, and data normalization. The study uses the model - EDANet and its modified version that adds two latent connections to improve object edge segmentation. The dataset consists of 955 training, 284 validation, and 128 test samples. The experimental result shows that both models achieve more than 80% f1-score when the imagery has matured broccoli flowers ready for harvest. Moreover, the modified model performs better than the original model by at least 5% on the unmatured dates. Furthermore, this study applies model pruning and model quantization to enhance the computation efficiency and enable real-time detection on the UAV with limited resources. The optimized model with integrate-GPU achieves six frame per second, which is 3.3 times faster than the vanilla model with CPU only. The modified model with integrate-GPU is also 3.3 times faster than the non-optimized version (CPU) and 1.3 times faster than the vanilla model (CPU). The proposed method demonstrates the feasibility of real-time detection and size estimation of broccoli flowers on single-frame aerial multispectral imagery. Keywords—multispectral imagery, CNN, model optimization, UAV, edge computing