Litcius/Paper detail

Interpreting atomization of agricultural spray image patterns using latent Dirichlet allocation techniques

Hongfei Li, Steven A. Cryer, John W. Raymond, Lipi Acharya

2020Artificial Intelligence in Agriculture12 citationsDOIOpen Access PDF

Abstract

Breakup patterns of agricultural formulations are explored using unsupervised learning techniques to elucidate the mechanics of atomization for oil-in-water formulations. Previous researchers have shown these formulations succumb to a different breakup mechanism than conventional formulations, beginning with inhomogeneities within the liquid sheet that nucleate holes within the material being sprayed, beginning the mechanism responsible for breaking the sheet into droplets. The Latent Dirichlet Allocation (LDA), a Bayesian hierarchical model, is used to explore unsupervised learning relationships between image analysis metrics on spray video data and the resulting atomization droplet size. Latent factors discovered by LDA were used for classification of video segments and achieved 99.9% accuracy (3-fold cross validation). Seventy-five videos were used for regression where each video had a unique measured droplet size distribution (D10, D50, and D90 values) for atomization. Experiments using the features learnt by Latent Dirichlet Allocation used with regression have extremely good results (R2 ~ 0.995 in 3-fold cross validation, R2 ~ 0.963 on never-seen videos), which serves as evidence for the potential use of this model in image analysis of agricultural spray patterns. LDA has huge potential in both learning and predicting atomization patterns [e.g., driftable fines (drops <150 μm)] when used with images based on the breakup phenomena in agricultural spray. These small drop sizes that occur during atomization have the greatest propensity for off-target movement through wind induced drift. LDA proved useful in characterizing current and future formulation designs using only images as witnessed by observations and excellent predictions summarized in this paper. In fact, these methods offer potential for use under field conditions to address spray performance based upon images of spray patterns at the nozzle without the need for expensive light scattering equipment often used to measure this phenomenon.

Topics & Concepts

Latent Dirichlet allocationBreakupBayesian probabilityComputer scienceArtificial intelligenceTopic modelMathematicsMechanicsPhysicsPlant Surface Properties and TreatmentsFluid Dynamics and Heat TransferAerosol Filtration and Electrostatic Precipitation
Interpreting atomization of agricultural spray image patterns using latent Dirichlet allocation techniques | Litcius