Seed localization system suite with CNNs for seed spacing estimation, population estimation and doubles identification

Abstract

Seed placement information for evaluating planters is obtained by tedious manual methods like using a pogostick or a ruler. To obtain more information and reduce the time required. There is a need for an automated system. The goal of this study is to design such a system with the use of GPS and seed detection.A 12-row planter was instrumented with cameras and a GPS. The planter was operated at planting speeds of 9.66 kmph and 12.87 kmph with a seed population of 74,131 and 86,4868 seeds per hectare respectively and seed spacing was measured manually using a ruler. YOLOR-P6, YOLOR-CSPX, YOLOX-S, YOLOX-M, YOLOX-L, YOLOX-TINY, and YOLOV4 were trained to detect seeds on a different dataset. An algorithm was designed to estimate the distance between two consecutive seeds by filtering out old detections from the detector with the use of the IOU metric and GPS stream.The seed spacing estimator was evaluated on Jensen–Shannon Divergence (JSD), mean, standard deviation, ΔCount, and RMS metrics. Results indicate that the developed system along with the algorithm was able to perform better at 9.66 kmph with JSD distances of 0.223 and 0.244 (74,131 and 86,4868 seeds per hectare respectively) compared to 0.298 and 0.258 (74,131 and 86,4868 seeds per hectare respectively) at 12.87 kmph. Moreover, this algorithm was able to reduce the total time taken to detect seed and estimate seed spacing information from 2hrs by manual method to 1 min 14 seconds using YOLOR-CSPX.

Keywords

• Planting systems
• Seed population estimation
• Seed localization
• Deep learning
• Convolutional neural networks
• Artificial intelligence