Development of a Real-Time Non-Contact Yield Mapping System Using Cascading Neural Networks for Wild Blueberry Harvesters

• A novel non-contact yield monitoring system was developed for wild blueberry harvesters using 3D Time-of-Flight imaging. • Cascading neural networks were used for point cloud segmentation, debris quantification, and mass estimation. • The proposed system achieved high mass prediction accuracy (R² = 0.952; MAE = 4.01 kg). • Real-time field trials demonstrated significant improvement in yield map resolution and estimation reliability. Precision agriculture has introduced technological advancements designed to enhance the monitoring and management of crop yields. An important component of these innovations is the development of accurate yield mapping systems. In wild blueberry ( Vaccinium angustifolium Ait.) cultivation, unique challenges arise due to the crop's growth pattern and the specific characteristics of the terrain. In addressing the limitations of traditional sensors used in the wild blueberry harvesting process, integrating a 3D Time-of-Flight (ToF) camera, along with image and point cloud processing algorithms, establishes a foundation for real-time yield estimation. The research introduces a novel cascading neural network (NN) architecture that surpasses traditional single-variable regression models by simultaneously capturing the complex, non-linear physical interactions between crop volume and embedded debris. Specifically, a cascading system of NNs was developed, initiated with a NN framework for point cloud segmentation in parallel with a NN framework for debris identification and quantification, culminating with a 1D NN for mass estimation. The system was field validated across two commercial wild blueberry fields (3.81 ha and 11.57 ha), collecting and analyzing a comprehensive dataset of 37,196 point cloud files from 95 harvested totes. Comparative analyses against traditional methods including load cell, volumetric flow rate and mass flow rate systems, revealed the proposed NN system achieved superior performance in mass estimation, quantified by higher R 2 values and lower mean absolute error (MAE) in mass estimation (0.952 and 4.01 kg, respectively). This was followed by a significantly lower MAE in yield estimation (725 ± 321 kg ha -1 ). This study demonstrated the efficacy of a cascading NN architecture in addressing the challenges of yield mapping in wild blueberry cultivation, opening for broader application across different crops.