On convergence of neural networks for applications in agriculture

Abstract

Neural networks, as universal function approximators, have gained increasing popularity due to the availability of abundant computational resources and data from IoT devices. This thesis investigates the applications of neural networks in agriculture, focusing specifically on seed detection, distance estimation, row detection, and multi-robot pesticide spraying tasks. The first study aims to develop an automated system for obtaining seed placement information and reducing the time required compared to manual methods like using a pogo-stick or a ruler. A 12-row planter was instrumented with cameras and a GPS, and various object detection models (YOLOR-P6, YOLOR-CSPX, YOLOX-S, YOLOX-M, YOLOX-L, YOLOX-TINY, and YOLOV4) were trained on a dedicated dataset. An algorithm was designed to estimate the distance between consecutive seeds by filtering out old detections using the Intersection over Union (IoU) metric and GPS stream. The system’s performance was evaluated using metrics such as Jensen–Shannon Divergenceplanting (JSD), mean, standard deviation, δCount, and RMS. Results indicate that the developed system and algorithm performed better at a planting speed of 9.66 kmph compared to 12.87 kmph, reducing the time taken to detect seeds and estimate seed spacing from 2 hours (manual method) to 1 minute 14 seconds using YOLOR-CSPX. The second study addresses the challenges of multi-robotic pesticide delivery tasks, where traditional task planning and allocation strategies require prior knowledge of crop infection levels, typically obtained through preliminary field scouting. As this assumption becomes obsolete with autonomous robots equipped with integrated sensing and spraying capabilities, the study reformulates the problem within the context of a Partially Observable Markov Decision Process (POMDP). A novel architecture, similar to Cross Transformer, is designed to address these challenges, and comparative evaluations against expert models trained on out-of-distribution datasets are conducted. The study demonstrated that the proposed model performs almost as well as the expert model,with a 0.5% difference in performance, and is more robust to out-of-distribution data. The evalutation metrics are the total distance travelled by the robots as well as the total reward obtained by the robots. The third study focuses on under-canopy robot navigation, which has seen significant demand for crop phenotyping, crop water stress assessment, and liquid application. Traditional GPS fails under canopies due to multi path issues, posing challenges. The study introduces a novel approach using smooth polynomials for row and canopy identification, leveraging image space prediction and a customized loss function to enhance generalization. This method delineates navigation boundaries with polynomial functions in image space and extrapolates seamlessly into real-world navigation due to these polynomials being smooth, down streams task such as control and state estimate become easier. The approach achieves 1ms latency on edge devices and is evaluated using a metric similar Intersection over Union named MPD(Mean Polynomial Distance), ensuring high accuracy for under-canopy navigation tasks.