EdgeSoybeanNet : A framework for real-time, high-accuracy field soybean pod counting

Accurate estimation of field soybean pods plays a critical role in precision agriculture. However, conventional methods face significant limitations, including high field variability, visually complex backgrounds, and the computational constraints of deploying deep learning models in rural edge environments. To address these challenges, we present EdgeSoybeanNet, a high-accuracy, edge-deployable AI framework for near real-time soybean pod counting. The proposed framework integrates a customized UNet-Lite segmentation network with an adaptive thresholding strategy. The computation process begins with region-of-interest extraction from UAV imagery, followed by segmentation and pod detection using adaptive thresholding. The trained AI models are then quantized and exported to ONNX and deployed with ONNX Runtime, TensorFlow Lite (TFLite), or TensorRT on edge devices, eliminating the need for cloud connectivity and enabling near real-time inference in the soybean field. To the best of our knowledge, this is the first study to incorporate 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. These results demonstrate the feasibility and effectiveness of this approach for resource-constrained precision farming. Compared with the state-of-the-art SoybeanNet-S model, our approach 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.