FASTCON: Food, Ag & Supply Chain in Tech Conference

As a leading supplier in the institutional food service industry, Land O’Lakes provides various processed cheese products via annual competitive contracts. Each bid requires a custom price, set individually per contract, trading off win probabilities (low prices) and profit margins (high prices).

To provide accurate estimates of win probabilities, the Land O’Lakes Data Science team set out to build a classification model, predicting likelihood of winning a bid at a given price, but training such a model is challenging due to limited data and overfitting. In this presentation we will share two innovative techniques in data augmentation and de-overfitting, which enable a classification model reliable enough to optimize prices and maximize expected profit.

Seagate Technology manufactures and excess of 2 million HDD heads per day. Robust manufacturing of this many units provides unique challenges that are well aligned with machine learning analytics strengths. In this session, we will describe a successfully demonstrated defect inspection application and discuss challenges in scaling the application to a high volume production environment. The discussion will highlight organizational, hardware, and software learning that is valuable for scaling machine learning applications.

IRI had developed and operates an ETL Framework that currently leverages internal infrastructure running Hadoop ecosystem services such as HDFS and Hive. There is continued interest in leveraging cloud-based resources providing utilization-based pricing and scalability to meet business needs. A prototype was performed to assess the compatibility and performance of migrating the existing IRI framework to Azure Databricks.

Developing Deep Learning solutions to improve quality and yield of wheat harvester using computer vision technology.

The first part the presentation describes the recent development of machine learning algorithms for rapid determination of deoxynivalenol (DON) levels in wheat and barley seeds from different genetic lines using hyperspectral imaging. As a widely cultivated crop and staple food in the world, wheat (Triticum aestivum L.) is crucial for ensuring food security. Unfortunately, its production is limited by many diseases. Fusarium head blight is one most devastating fungal disease for wheat. Fusarium is a plant pathogen and DON-producer. Many environmental factors affect fungal growth and therefore cereal damage and toxin accumulation, such as temperature, water activity, pH and nutrient composition. It is very important to breed the disease-resistant varieties. To accurately identify such variants, it is particularly important to expose germplasm lines to field pathogens and to assess disease responses for breeders. However, it is impractical to use conventional methods to rapidly and non-destructively assess DON contents in large-scale wheat cultivation. In this study, the hyperspectral image data of over 1000 wheat and barley samples (each sample about 20 g) from over 500 genetic lines with different DON levels were collected, then the DON contents were measured with gas chromatograph-mass spectrometry (GC–MS). The spectral data in reflectance obtained from the images were used to develop multivariate models including partial least-squares regression (PLSR), support vector machine regression (SVMR), locally weighted partial least square regression (LWPLSR), and Artificial neural network (ANN) models. The prediction accuracy of the full wavelength models was tested. Feature wavelengths were chosen based on successive projection algorithm (SPA), and competitive adaptive reweighted sampling (CARS). The selected wavelengths were examined on evaluation of the prediction accuracy of models. The results demonstrate that the hyperspectral imaging and machine learning can be conveniently used to determine DON in wheat and barley seeds.

The second part of the presentation describes the recent development of a novel, systems approach to automated weed recognition in vegetable crops grown under high weed density conditions, typical of organic production systems. It describes the development of a novel seed treatment developed to create a machine-readable vegetable crop plant and matching robotic computer vision system that together can address current challenges that traditional computer vision and machine learning approaches face when attempting to detect and identify vegetable crop plants growing in fields with high weed densities and significant levels of foliage occlusion. Vegetable crops are very susceptible to damage from weed competition, with early season weeds emerging within 3 weeks after crop emergence a control priority to prevent significant yield loss. The objective was to establish an effective computer vision method to rapidly differentiate weeds using Crop Signaling, by identification of Rhodamine B uptake in vegetable crop seedlings as a machine-readable identification trait. This presentation describes a novel computer vision system that was designed with a custom illumination system designed specifically to excite the fluorescence properties of Rhodamine B and image them. Rhodamine B was selected for study because it can be used as a fluorescent tracer, has good systemic properties in plants, and is included on the USA EPA List 4B of inert pesticide ingredients “for which EPA has sufficient information to reasonably conclude that the current use pattern in pesticide products will not adversely affect public health or the environment.” and has been used as a tracer in drinking water and for many biological applications. Study results show that the system can detect and allow visualization of the Rhodamine dye internal to the crop system. The research demonstrates that a Crop Signaling approach, using Rhodamine B can be used by a computer vision system to automatically discriminate weeds from crops.