Weather-Driven Agricultural Decision-Making Using Digital Twins Under Imperfect Conditions

TL;DR

This research develops a modular framework named Cerealia that allows end-users to check for data inconsistencies when perfect weather feeds are unavailable and aids end-users in informed decision-making, and develops a prototype of Cerealia using the NVIDIA Jetson Orin platform.

Abstract

By offering a dynamic, real-time virtual representation of physical systems, digital twin technology can enhance data-driven decision-making in digital agriculture. Our research shows how digital twins are useful for detecting inconsistencies in agricultural weather data measurements, which are key attributes for various agricultural decision-making and automation tasks. We develop a modular framework named Cerealia that allows end-users to check for data inconsistencies when perfect weather feeds are unavailable. Cerealia uses neural network models to check anomalies and aids end-users in informed decision-making. We develop a prototype of Cerealia using the NVIDIA Jetson Orin platform and test it with an operational weather network established in a commercial orchard as well as publicly available weather datasets.

Figures (7)

• Figure 1: High-level schematic of this work. Cerealia stores historical weather traces and simulated anomalous data that is used to detect runtime inconsistent data generated by the remote weather stations deployed in the field. The end users can use Cerealia to analyze how inconsistent data impacts the target agriculture decision-making process.
• Figure 2: Workflow of Cerealia. We make Cerealia modular that allows designers to integrate various noisy data that can be used with historical weather traces to train machine learning model(s). A runtime consistency checker module uses a trained model to check for imperfect measurements and their impact on targeted decision-making applications (for instance, fruit heat/frost prediction).
• Figure 3: Visualization of the data inconsistencies incorporated with Cerealia. The plot shows temperature readings over time in the presence of the following inconsistencies: (a) random (top-left), (b) malfunction (top-right), (c) drift (bottom-left), and (d) bias (bottom-right).
• Figure 4: Average inference time (left) and average memory usage (right) per sample for the Quincy network.
• Figure 5: Performance of Cerealia with multiple instances. The plots show average inference time (left) and average memory usage (right) for the Quincy network with varying container configurations. Even with 30 instances of Cerealia running concurrently, we can get inference results in less than a second and no more than 0.4 MB of memory consumption.