A fundamental change in the application of chemical pesticides must be accompanied by an increase in the level of automation and reliability in pest monitoring. The application of alternatives to synthetic chemical pesticides requires more detailed insights about pathogen induced crop stress in the field along with a spatially precise perimeter of pathogen occurrence. A high-resolution monitoring, both spatially and temporally, can be accomplished utilizing UAVs (unmanned aerial vehicles, often referred to as ‘drones’) in combination with AI-based data analysis.

The core of this research project is the development of an AI-based pathogen detection pipeline based on MLOps principles, especially for the crops considered in NOcsPS (wheat and soybean). We start off with the already available hyperspectral data from CP 10, a database that will be further extended in the course of the project. For the analysis of these data, modern image processing techniques based on AI (computer vision) will be developed and integrated into a data-centric AI pipeline comprising algorithms for data management, data preprocessing and predictive Deep Learning models.

An analysis pipeline for the purpose of automated pathogen detection based on hyperspectral images is developed taking advantage of novel systematic development and deployment principles for data-centric AI applications (MLOps). First, an in-depth manual analysis of the existing database as well as the underlying pathogen monitoring process is conducted (WP 1: Process & Data Understanding). Subsequently, necessary pre-processing steps and data augmentation approaches are assessed with respect to the identified characteristics of the available and further collected hyperspectral data and the technical (i.e. programmatic) implementation of the AI pipeline is initiated (WP 2: Data Engineering). This is followed by the systematic selection and configuration of appropriate Deep Learning models from the state of the art in computer vision, which is performed in an iterative model development and validation process (WP 3: Model Engineering). An encompassing part of the outlined work program constitutes the iterative technical implementation of the AI pipeline, which ensures both ongoing systematic data management and accordingly a continual model adaptation to data becoming available over time (WP 4: MLOps).

Expected Results:

At the end of the project, a robust and automated pathogen detection pipeline should exist that can efficiently adapt to changing operational conditions and reduce future costs incurred for the essential task of pathogen monitoring in integrated pest management through the use of AI.