SPECIFIC AIMS:

We will develop a food system management framework by combining Performance Based Engineering (PBE) and Artificial Intelligence (AI), to increase the efficiency, safety & resilience of agricultural production by improving food yield in terms of both quantity and quality, controlling crop diseases, decreasing resource consumption & waste, and increasing traceability. Traditionally, agricultural management is based on empirical judgement resulting from experiments & experience, which is no longer adequate because of increased complexity and added uncertainty of food systems due to the increasing demands of the world population. Thus, uncertainty quantification in food systems has been recently adopted. However, a robust framework is still lacking and this project aims at filling this gap.

SIGNIFICANCE AND IMPACT:

We will expand the PBE to Performance Based Food Engineering (PBFE) to address challenges faced by the food industry. PBFE will be integrated with Machine Learning (ML) and digital twins to develop a framework for managing food systems. PBFE will have several applications:

1) Evaluation of the performance of existing food facilities (including greenhouses & open fields) under different environmental hazards.

2) Evaluation of the performance enhancement methods (for increasing crop yield, e.g., use of fertilizers & maintaining proper water drainage; for controlling crop disease, e.g., chemical & biological controls, therapy & genetic engineering) in terms of cost and resilience and accordingly ranking these methods.

3) Quantifying the benefits of real-time monitoring, and automated machinery & sensors on the resilience and performance of food facilities. The expected impact is increased efficiency, safety & resilience of various processes.

The project is relevant to two general objectives of AIFS, which are enabling 1) agricultural producers to maximize food quantity and quality, and 2) food processors and distributors to deliver highly traceable and safe food, while minimizing resource consumption and waste.

BACKGROUND:

Crop yield prediction is of great importance to global food production. Policy makers rely on accurate predictions to make timely import decisions for national food security. During the last decade, scientists and engineers have made significant headway in developing and deploying tools and devices that deliver massive, yet too often raw, data streams to food system stakeholders at unprecedented spatiotemporal resolution. Concurrently, AI algorithms repeatedly break benchmarks in Computer Vision (CV) applications, Natural Language Processing (NLP), and automation. Despite its significant benefits, the complex food systems face several challenges in the application and adoption of AI. Frameworks developed in other domains can be integrated with AI approaches to tackle these challenges. PBE is one of these frameworks with a robust formulation for decision making under inherent uncertainty due to the complex nature of food systems. PBE originated in the natural hazards field, particularly earthquake engineering, and is now adopted by wind and fire engineering. 

APPROACH WITH SUCCESS METRICS:

Success metrics for the above three impactful applications are defined as the number of:

1) Evaluated food facilities

2) Identified effective enhancement methods

3) Determined efficient monitoring techniques and sensors

PBFE consists of four main analyses, Figure 1, namely 1) Environmental hazard, 2) Crop growth, 3) Crop yield, and 4) Loss. The hazard analysis (e.g., climate change expressed as Temperature/Humidity Index (THI), soil degradation, deforestation, diseases & drought) is performed for a food facility under different scenarios, quantified by the probability of occurrence of Intensity Measures (IMs) for various hazard levels. 

Crop growth analysis uses IMs along with crop models for different crops and sources of environmental hazards to determine the crop growth condition of the food facility. For the purpose of illustration, we use Stunting Measures (SMs) to represent the crop growth condition. Stunting is one common symptom of plant disease, which significantly influences food production and is responsible for reducing the physical and economic productivity of crops and can be a major impediment. Deep Neural Network (DNN) object localization & segmentation can be implemented on global & local levels making use of satellite or wide-angle images by Unmanned Aerial Vehicles (UAVs) to locate “damages” of crops or to evaluate the health condition of plants. Accordingly, SMs fragilities are determined as a function of the hazard IM.

Figure 1: Overview of the proposed PBE methodology for food systems.

The crop yield analysis uses physical and data-driven models (digital twins) for simulations to predict the yield quantity and quality under hazard scenarios of the given SMs. The outcome is represented in terms of Yield Parameters (YPs), e.g., Total Factor Productivity (TFP), considering changes in crop yield, and cost of production and commodity mixes. This is tied to a collateral loss analysis for annual rate of crop spoilage or waste.

Finally, given the YPs, a set of Decision Variables (DVs) including operability, mitigation costs/duration, and potential for collateral losses are evaluated to inform stakeholders about the future performance of the food facility and to help them make informed decisions for actions to improve the performance. PBFE, from food facility F to IM to SM to YP to DV, involves uncertainty quantified by the total probability theory:

where p[X] is probability density function of X, λ(IM) is Mean Annual Frequency (MAF) of events with IM, and λ(DV) is the MAF in terms of the DV, which is influenced by the rate of collateral loss λ_coll. Each function represents one element of the analysis methodology: λ(IM) from hazard analysis, pSM|IM from growth analysis, pYP|SM from yield analysis, and pDV|YPfrom loss analysis.

Eq. (1) de-constructs the assessment into four analyses using three “intermediate variables”, namely IM, SM & YP, and re-couples these analyses by integrating over all levels of the variables and making use of relevant conditional probabilities. Appropriate variables of IMs, SMs & YPs should be chosen where the conditioning information need not be carried forward, e.g., SMs should be selected so that YPs (& DVs) do not vary with IMs. Forward and backward Uncertainty Quantification (UQ) methods are considered in PBFE. The former UQ method is employed in various phases of PBFE, e.g., to identify which of the YPs provide the largest contribution to uncertainty in the yield analysis. Starting with a target performance, the latter UQ method is used to determine design parameters of the food facility.

Several phases of the developed PBFE framework will be supported by various AI methodologies. One of these is the use of DNNs to predict the crop yield. Moreover, image segmentation will be used to monitor the evolution of crops and to estimate yield, identify crop disease in the hazard and stunting analyses, and detect any potential problem with the agricultural facility or its components (i.e., irrigation systems, drainage systems, greenhouses, etc.) including third-party intrusion to reduce potential risk and improve productivity. For segmentation, images will be collected by crowd-sourcing from UAVs, farmers, and manufacturers. NLP will be used for hazard analysis to reduce the uncertainty of the defined hazards. Considering that ground-truth datasets are privately held in the food industry, getting direct input from farmers using NLP is important to define/update characteristics, frequency and levels of hazards. NLP will also be used to compare PBFE results with the ground truth, e.g., median downtime or economic losses from the PBFE simulations will be compared with the actual values determined by NLP.

CALL FOR COLLABORATION:

We formulated a dictionary file containing two tables. Table 1 lists the possible factors to study for the four stages of the PBFE Framework. Table 2 collects the agricultural terminology related to the PBFE methodology. 

Table 1: Elements in the PBFE framework

Table 2: Agricultural terminology

Once the data are collected, researchers within the AIFS can search for a particular term and extract the description from the document, as shown in Figure 2. 

Figure 2: Dictionary search example

All AIFS project teams are invited to provide contributions to the elements in the PBFE framework and to the agricultural terminology databases using the following links:

Contribute to Table 1: https://forms.gle/e9s8WnAvvz12VRMR7

Contribute to Table 2: https://forms.gle/9VyfqUfUWTqGNgZz9

---

Project Team

Collaboration

The projects are jointly conducted with partners through the AIFS Network:
https://aifs.ucdavis.edu/

UC Berkeley

UC Davis