Visualization and characterization of agricultural sprays using machine learning based digital inline holography

Abstract

Accurate characterization of agricultural sprays is crucial to predict in-field performance of liquid applied crop protection products. In this paper, we introduce a robust and efficient machine learning (ML) based Digital In-line Holography (DIH) algorithm to accurately characterize the droplet field for a wide range of commonly used agricultural spray nozzles. Compared to non-ML based DIH processing, the ML-based algorithm enhances accuracy, generalizability, and processing speed. The ML-based approach employs two neural networks: a modified U-Net to obtain the 3D droplet field from the numerically reconstructed optical field, followed by a VGG16 classifier to reduce false positives from the U-Net prediction. The modified U-Net is trained using holograms generated using a single spray nozzle (XR11003) at three different spray locations; center, half-span, and the spray edge to create training data with different number densities and droplet size ranges. VGG16 is trained using the minimum intensity projection of the droplet 3D point spread function (PSF). Data augmentation is used to increase the efficiency of classification and make the algorithm generalizable for different measurement settings. The model is validated using National Institute of Standards and Technology (NIST) traceable glass beads and six different agricultural spray nozzles representing various spray characteristics. The principal results demonstrate a high accuracy rate, with over 90% droplet extraction and less than 5% false positives, regardless of droplet number density and size. Compared to traditional spray measurement techniques, the new ML-based DIH methodology offers a significant leap forward in spatial resolution and generalizability. We show that the proposed ML-based algorithm can extract the real cumulative volume distribution of the NIST beads, where the LD system measurements are biased towards droplets moving at slower speeds. Additionally, the ML-based DIH approach enables the estimation of mass and momentum flux at different locations and the calculation of relative velocities of droplet pairs, which are difficult to obtain using conventional spray characterization techniques.