Recipient Organization
AGSMART INC
712 BUTLER ST
SAFETY HARBOR,FL 34695
Performing Department
(N/A)
Non Technical Summary
Current cucurbit production practices involve multiple broadcast applications of fungicides from planting to harvest. Our goal is to devise an AI-based system that administers fungicides only where the crop canopy or disease is present. The specific objectives are: (1) Create advanced AI models for detection and identification of cucurbit crop canopies, flowers, fruits, and disease symptoms; (2) Develop a full-scale prototype smart spray system, using the AI models and machine vision for precise pesticide applications on cucurbits. We prioritize cucurbits as our initial focus due to an existing image database, disease susceptibility, and their vining growth pattern, which often results in off-target applications. This offers a substantial chance to minimize pesticide use, especially early in the season. Our approach facilitates rapid training for the detection of various cucurbit types in diverse production environments. We can produce ample synthetic image datasets for model training, even with limited original images, with the use of generative AI. This is crucial for early detection of exotic diseases which lack large image datasets, thus preventing potential industry-wide damage if such diseases are introduced. Ultimately, we will commercialize a tractor-mounted or autonomous system that applies fungicides selectively and monitors disease.
Animal Health Component
100%
Research Effort Categories
Basic
(N/A)
Applied
100%
Developmental
(N/A)
Goals / Objectives
The first goal of the project is to incorporate state-of-the-art (SOTA) deep-learning machine vision models into a high-performance computerized tractor sprayer controller for targeted pesticide applications on cucurbits.To achieve this we will train SOTA object detection and instance segmentation models to; 1) detect, identify and count cucurbit crop canopies, flowers and fruit, 2) detect common cucurbit diseases, and 3) utilize generative SOTA models to create abundant synthetic image data for model improvement. The second goal of the project is to design and build a pre-commercialization but full-scale grade prototype smart spray system. We will integrate sprayer controller with pulse width modulation system for valves,data acquisition, artificial lighting, dual boom system, self priming capabilities, estimation of spray usage, reporting and simulation functions into the spray system.
Project Methods
OBJECTIVE 1: Incorporate state-of-the-art (SOTA) deep-learning machine vision models into a high-performance computerized tractor sprayer controller for targeted pesticide applications on cucurbits.Methods: In our pursuit to optimize targeted spraying, we aim to identify not just the cucurbit canopies through segmentation but also concurrent features such as flowers and fruit. These additional features will provide insights into crop development and performance metrics. At the forefront of current advancements, the YOLOv5 and YOLOv8 models by Ultralytics, Los Angeles, CA, represent the state-of-the-art (SOTA) for both object detection and instance segmentation. These models excel in the simultaneous identification of multiple feature classes they are trained on. Notably, YOLOv8 also incorporates an inherent object-tracking capability, facilitating real-time tracking and counting of individual entities.OBJECTIVE 2: Design and build a pre-commercialization but full-scale grade prototype smart spray system. ?The prototype smart sprayer will be built by retrofitting a small commercial sprayer (Figure 3A) with computerized machine vision and spray nozzle control components. In our quest to select the most efficient computational platform for machine vision to power the smart sprayer, we plan to evaluate the performance of NVIDIA's Jetson AGX Orin series--ARM64-CPU and Ampere GPU-integrated embedded systems--against custom-constructed field computers equipped with x86-64 mini-ITX motherboards in NEMA-protected enclosures, powered by 12VDC sources. To illustrate, our lab has pioneered a prototype field computer with x86 architecture, incorporating a mini-ITX motherboard, an Intel i9 CPU (comprising 10 cores and 20 threads), 32Gb of DDR5 RAM, and a cutting-edge NVIDIA RTX A2000 GPU built on the Ampere architecture. Under peak operation, the power consumption of this x86-64 system stands at approximately 200 watts, contrasted against the 60 watts of the Orin ARM system. The energy efficiency of the Orin ARM system, as evidenced by its reduced power consumption, certainly holds appeal. However, when subjected to intense real-time multi-tasking demands, particularly in multi-camera machine vision applications, its performance tends to lag behind the x86-64 architecture.For a comprehensive performance assessment of these two candidate computer systems to be used in a smart sprayer controller, we'll employ standard image datasets and our proprietary YoloV8 models to compute average inference durations (in milliseconds per image). Subsequent efficiency metrics will be derived by factoring in the energy expenditure--measured in kilowatt-hours (kWh)--for each computational task using precision watt meters (http://www.p3international.com/products/p4400.html). To further examine the multitasking and multithreading capabilities of these platforms, this methodology will be extrapolated to process live feeds from four HD USB cameras, capturing concurrent video streams. For each computer platform, we will also leverage NVIDIA's TensorRT- a framework adept at optimizing and compressing models, delivering up to threefold improvements in inference speeds on NVIDIA GPUs. Furthermore, computational efficiencies between the conventional 32-bit floating point inference modes and the expedited 16-bit modes will be compared, the latter potentially offering a twofold increase in speed without accuracy trade-offs.Our computerized smart sprayer system prototype will integrate several enhancements to optimize performance, reliability, and user experience. To enhance machine vision accuracy, an auxiliary LED light will be introduced to augment the ambient lighting. We will implement a robust relational database within the controller, utilizing SQLite to archive vital spray parameters including GPS timestamp, coordinates, speed, as well as camera-captured JPEG images. This database will also store measurements on spray nozzle/valve activations, pesticide pressure, flow rates, and reservoir levels. Pressure and flow sensors will be positioned between the pressure regulator and spray boom, with the tank level sensor placed in the tank's sump. The comprehensive database will facilitate the generation of detailed spray metrics, including volume utilization, area coverage, and as-applied mapping. Archived images will be instrumental for refining AI models and simulating spray scenarios on the controller, enabling, for instance, comparisons of chemical savings across different spray buffer settings without actually running the sprayer in the field.