Sensor agnostic AI models for digital phenotyping

OVERVIEW

Digital sensors have emerged as the predominant technology for collecting large-scale data in many industries.  In agriculture, quantification of phenotypes through deep learning models that utilize digital sensor data underlies the development of higher-yielding, more resilient crops as well as novel solutions to control weeds, pests, and disease. However, the rapid evolution of digital sensor technologies has led to the replacement of older sensor types with newer ones.  This transition often requires additional data capture, reannotation, and retraining of preexisting machine learning models to maintain their efficiency and accuracy. 

Corteva Agriscience, the Seeker for this Challenge, is seeking novel methods or models to universalize data collected by different sensors of the same or similar type, without the increased costs associated with re-collecting or reannotating data and retraining existing models.

This is a Prize Challenge which requires a written proposal to be submitted and there will be a guaranteed award for at least one submitted solution. By submitting a proposal, the Solver grants Corteva a right to use any information included in their proposal.

Submissions to this Challenge must be received by 11:59 PM (US Eastern Time) on October 02nd, 2023.

- Login or register your interest to start solving.

ABOUT THE SEEKER & ELIGIBILITY

Corteva, Inc. (NYSE: CTVA) is a publicly traded, global pure-play agriculture company that combines industry-leading innovation, high-touch customer engagement and operational execution to profitably deliver solutions for the world’s most pressing agriculture challenges. Corteva generates advantaged market preference through its unique distribution strategy, together with its balanced and globally diverse mix of seed, crop protection, and digital products and services. With some of the most recognized brands in agriculture and a technology pipeline well positioned to drive growth, the company is committed to maximizing productivity for farmers, while working with stakeholders throughout the food system as it fulfills its promise to enrich the lives of those who produce and those who consume, ensuring progress for generations to come. More information can be found at www.corteva.com. 

Find out more about participation in Wazoku Crowd Challenges.

THE CHALLENGE

Background

Over the past 10 years, digital sensors (e.g., RGB, multispectral, and hyperspectral cameras; LiDAR; etc.) have emerged as the predominant technology for collecting large-scale data in many industries. When placed on remote sensing vehicles like drones, satellites or terrestrial vehicles, digital sensors can collect large amounts of data in a short period of time with very little labor, greatly increasing efficiency. In agriculture, quantification of phenotypes through the analysis of digital sensor data underlies the development of higher yielding, more resilient crops and novel solutions to control weeds, pests, and disease.

Early on, many phenotypes derived from sensors were generated using simple color segmentation (e.g., differentiating crop canopy from soil).  However, recent advances in artificial intelligence (specifically deep learning) have enabled data scientists to generate models that extract more subtle features on an individual level, akin to how a brain might learn to recognize a specific object by learning its shape, proportions, color, etc. leading to a vast expansion of phenotyping solutions.  This opens the door to the analysis of far more features, which can result in superior products.  

The rapid evolution of digital sensor technologies has led to the replacement of older sensor types with newer ones that deliver advanced features at lower costs.   When sensors are replaced, deep learning models developed from previous sensor types often do not perform as efficiently or accurately due to sensor-to-sensor variability (sensor characteristics are different), feature extraction discrepancies (distinct patterns that the model recognizes change from sensor to sensor), and limitations of generalization (the training set and test sets do not have sufficient overlap).  Techniques like domain adaptation, transfer learning, and model fine-tuning do exist to overcome these challenges, but are often expensive requiring additional data capture, annotation, and model retraining. 

Corteva Agriscience is seeking novel methods or models to universalize data collected by different sensors of the same or similar type without the increased costs associated with re-collecting or reannotating data and retraining existing models.

SOLUTION REQUIREMENTS & ACCEPTANCE CRITERIA

Any proposed solution should address the following Solution Requirements: 

Must have: 

• The proposed model or method can universalize RGB sensor data between sensors of the same type. For example, models or methods that can be applied (segmentation, classification, or detection) to images acquired by two RGB sensors in all three of the following scenarios:
  • of the same type and same model but physically different cameras from same manufacturer (e.g. two Sony RGB cameras of same model)
  • of the same type but different models from same manufacturer (e.g., two models of a Cannon RGB camera) 
    • An example data set reflecting a scenario where a model trained on Sensor A data will be used on data collected from Sensor B is accessible after agreeing to the CSA. The data set consists of paired images from each sensor taken at the same location. Images with the same file name found in different “sensor folders” are paired. 
  • of the same type but different manufacturer (e.g., RGB camera from Cannon vs DJI) 
    • An example data set reflecting a scenario where a model trained on Sensor A data will be used on data collected from Sensor C is accessible after agreeing to the CSA. The data set consists of paired images from each sensor taken at the same location. Images with the same file name found in different “sensor folders” are paired. 

Nice to Haves: 

• The proposed model or method can universalize data for the same phenotype collected by sensors of related types (e.g., data from an RGB camera and a multispectral or hyperspectral sensor)
• Methods that can also be used to transfer models developed in one environment type to another, across different sensor resolutions, or images acquired from one camera angle to another
• The proposed model or method is computationally efficient 

Approaches not of interest: 

• Models or methods that require re-collecting a large dataset with the new sensor, reannotating the data, or retraining existing models  

Solutions with Technology Readiness Levels (TRLs) 1-6 are invited; from ideas about WHAT to do, to concepts/solutions with HOW to achieve this.

This is a Prize Challenge with a guaranteed award and has the following features:
 

1. There will be a guaranteed award for at least one submitted solution. The best solution(s) has the opportunity to win all or part of the guaranteed award of $15,000 USD, as solely determined by Corteva.
2. The total payout will be $15,000, with at least one award being no smaller than $5,000 and no award being smaller than $2,500.
3. By submitting a proposal, the Solver grants Corteva a royalty-free, perpetual and non-exclusive right to use any information included in their proposal.
4. Corteva Agriscience may also issue “Honorable Mention” recognitions for notable submissions that are not selected for monetary awards.