Practical Strategies for Automated Phenotyping: From Raw UAV Data to Multispectral Time Series for Machine Learning Applications.

This study addresses the transition from raw unmanned aerial vehicle (UAV) acquisitions to a multispectral image time series for machine learning applications in precision agriculture. Traditionally reliant on manual labour, wheat breeding and phenotypic assessment suffer from subjectivity and inefficiency. To harness efficient machine learning methods, accurate datasets are essential. Our objective is to standardise procedures, enhance accuracy and facilitate scalability when creating those datasets. We realise a study on the experimental field at Obersiebenbrunn, Austria, managed by Saatzucht Edelhof. Using a custom hexacopter equipped with a multispectral camera, we acquire image data which are subsequently processed by using Pix4Dmapper to generate large orthomosaics. To be organised into comprehensive data, the time series of images are aligned with expert-assessed ground truths for machine learning training. We address challenges such as data processing, experimental design and geolocation accuracy, including evaluating resampling algorithms. Our benchmark justifies the use of bicubic resampling for balancing computational efficiency and image quality. This study contributes to advancing machine learning applications in remote phenotyping and precision agriculture, offering insights into overcoming technical challenges and enabling standardised, scalable solutions.