Digital imaging technique to detect and count aphids in soybeans
Principle Investigator: Sreekala Bajwa, Professor & Chair, Agricultural and Biosystems Engineering. Co-investigator: Jason P. Harmon, Professor, Entomology; John Nowatzki, Extension Specialist, S. Sivarajan, Research Associate, M.M. Maharlooei, Visiting Scholar, Ag & Biosystems Engineering, NDSU.
Aphid is a serious pest in soybean, with a threshold for economic damage assessed as 250 aphids on a single soybean plant. Currently, aphids on soybean plants are counted manually to make crop management decisions such as insecticide application. Manual counting is a time consuming and laborious task. An automated counting method would make the task faster and easier. The objective of this project was to investigate the use of digital images of soybean leaves captured with regular consumer cameras to accurately detect and count aphids of on soybean leaves.
Data were collected from soybean plants grown on test fields near Fargo under the supervision of NDSU entomologist Dr. Jason Harmon. The soybean varieties grown on the test fields were considered as susceptible to soybean aphids. Images of soybean leaves with aphids on them were acquired with three different consumer cameras, a Sony W80 (7.2 MP) digital camera, AppleTM iPhone 6 (8 MP) and NokiaTM Lumia 1020 (41 MP) with and without a macro lens. The purpose of using macro lens was to have a better focus on images taken at close proximity since there were issues with camera focusing in 2014 green house study when the camera was too close to target. Images were acquired from soybean fields on July 23, August 4 and August 12 of 2015. The protocols followed (camera type, data collection method, etc.) were similar for all three sets of data collection. The leaves were laid straight on a still board that had a ruler taped on it, and images of the leaf were taken.
The acquired images were processed to remove the still board background from the leaf part using leaf color to distinguish the leaf from the background. Once the leaf on the board was identified, the image was processed again to identify the objects on the leaf that included aphids, leaf spot, exoskeleton (dead skin of aphid), etc using their hue (color tone). From these objects, aphids were identified using their distinct hue, size, roundness and aspect ratio (length to width ratio), and the objects identified as aphids were counted. Randomly selected aphids on the leaf processed with image processing method were manually checked to verify how accurately the method identified aphids. The method was tested on images of the top and bottom sides on soybean leaves.
This research showed that the digital imaging technique has the potential to detect and count aphids on soybean leaves in field conditions with reasonable accuracy. There was strong correlations between aphid count estimated with the digital imaging technique and manual count under field conditions. The Sony camera gave the best correlation of 0.93 on the top side of leaves, and 0.87 on the bottom side of the leaves. The correlations were a little less than what was observed for images collected in greenhouse conditions under bright light. The iPhone camera also performed well. However, the Nokia camera performed poorly for detecting and counting aphids. The misclassifications were caused by exoskeleton that had similar size and shape as a small aphid, aphids that are too small, and some leaf spot that looked similar to an aphid. We continue to refine the algorithm. Overall, the method was successful in detecting and counting aphids.