EDGESOYBEANNET: A FRAMEWORK FOR REAL-TIME, HIGH-ACCURACY FIELD SOYBEAN POD COUNTING

Author

• Johnbosco Chinonso Nnamso, Southern Illinois University CarbondaleFollow

Date of Award

8-1-2026

Degree Name

Master of Science

Department

Electrical and Computer Engineering

First Advisor

Lu, Chao

Abstract

Effective evaluation of soybean pods on the field plays a critical role in precision agriculture by enabling farmers to optimize crop management and resource allocation. It is key to note that conventional methods face significant limitations, which include high field variability, reliance on manual data collection, and limited computational resources in rural environments where farming remains largely labor-intensive. In order to address these challenges, this study presents EdgeSoybeanNet, a high-accuracy and edge-deployable AI framework for near real-time soybean pod counting. This proposed framework addresses key global challenges by enabling automated real-time pod counting at scale, accurate prediction of soybean pod count in the presence of strong background noise (soils, weeds, debris, etc.), and lightweight model deployment that is suitable for edge devices in environments where smart agricultural technology is required. Moreover, it tackles critical issues such as labor-intensive manual yield estimation, limited access to high-performance computing in rural farming areas, and the need for adaptable, cost-effective, and sustainable agricultural monitoring systems to support global food security and precision farming initiatives. The proposed framework integrates a customized UNet-Lite segmentation network with an adaptive thresholding strategy. The computation process begins with region-of-interest (ROI) extraction from UAV imagery, followed by segmentation by the UNet-Lite to precisely segment soybean pods, ensuring efficient separation from strong background noise elements, including soil, weeds, debris, and thereafter pod detection by adaptive thresholding, using ResNet10-Lite to enable accurate pod detection and counting, which forms an efficient pipeline for high accuracy soybean pod counting. The trained AI models are quantized afterward and then exported to ONNX and deployed with ONNX Runtime or TensorFlow Lite (TFLite) conversion on edge devices, hence eliminating the need for cloud connectivity and enabling near real-time inference in the soybean field. Together, this framework forms a real-time, high accuracy field soybean pod counting system. To the best of our knowledge, this is the first study to integrate adaptive threshold learning into a UNet-Lite segmentation for agricultural applications. The experimental results show a counting accuracy of 89.57% with an inference time of 0.66 seconds on a Raspberry Pi 5 at 300 × 300 input UAV images, and up to 90.43% counting accuracy at 560 × 560 input. Overall, these results clearly validate the practicality and effectiveness of this approach for resource-constrained precision farming. Compared with the state-of-the-art SoybeanNet-S model, this study improves counting accuracy by 5.07% and reduces the number of parameters by approximately 14 times, from 49.6 million down to 3.57 million.