Deep Learning Algorithm for Atomization Characterization using Shadowgraph Images

Predicting droplet size distribution as a function of chemical composition, input pressure, nozzle geometry, and atmospheric conditions using algebraic approach is extremely complicated due to the number of variables involved, especially for agricultural spray applications. A low cost shadowgraphy technique was used to capture the spray pattern from six different standard ASABE nozzles (ranging from extremely coarse 6515 type to very fine 11001 nozzle type) at different inlet pressures, along with various combinations of adjuvant mixtures. The resultant images were then fed into convolution neural network (CNN) architecture to classify the different nozzle types, pressure, and the adjuvant mixtures. Preliminary results show that the CNN architecture was able to classify the spray patterns with different nozzles, pressures, and adjuvant mixtures with more than 98% accuracy. The architecture was able to identify the subtle changes in the atomization in different configurations which is shown through class activation mapping.