Do not trust blindly an Artificial Intelligence-based system for identification of insect species

 

Do not trust blindly an Artificial Intelligence-based system for identification of insect species

Govind Gujar

AGRIBIOSYS, New Delhi

agribiosys@gmail.com

The artificial intelligence (AI)-based identification is being used for insect pest identity under field condition so that the data are used for forecasting and decision making in respect of crop protection tactics. This provides for promptness and ease of operation unlike conventional method. However, reliability of the AI will depend upon the database of insect images. Further, most databases are based upon adult stages and hence, this system heavily rely more often on a stage which causes less or little damage to the crop.

Our objective was to test one such AI wherein insect images are uploaded and analysed by the algorithm software that identify these images in respect of names-common as well as scientific name on the basis of similarity with the image(s) of the insect specimen in the database. I used http://pathangasuchaka.in (developed by School of Ecology and Conservation, UAS, GKVK, Bangluru under the leadership of K.N. Ganeshaiah, Project Coordinator and sponsored by Department of Biotechnology, GoI) for this test.

Methodology involved insect images as known to me in my photographic collection. These insect images were identified with peer’s help too and are based upon years of experience. These images were loaded as per instructions and results are tabulated.

Table 1 shows that many aspects like intensity of light for photography, background of insect images, angle of photography-side or top view, resolution intensity as well as cropping of insect images as provided in the software makes lot of difference in identity of images. Correct identity was found for as low as 7.53% similarity for the fall armyworm, while incorrect identity was found for as high as 98.1% similarity for the pink bollworm. The reason for this inaccuracy is that pink bollworm is not listed in the database. Moreover, Brown house moth (Hofmannophila pseudospretella) certainly looks different in terms of spots over its wings, with more wings overlap and lack of border brush hairs.

Conclusion: The AI-based identification as it is, may compliment, but cannot replace conventional system of identification.

PS: Another problem that may come up is variation in the use of common name. For e.g. the software calls ‘cotton leafworm or tobacco cutworm’ as ‘Oriental leafworm moth’. Unfortunately, most common names are not standardized or validated with consensus as found in the developed countries where the concerned apex society approves the name.

Table 1. Identification of moths and one larva with pathangasuchaka.in

 

Insect-as known to author	Orientation	Crop background	AI-based Identity as	% similarity with database	Result
Different light intensity, background, pose, morphology
Fall armyworm (FAW), Spodoptera frugiperda	Top in situ	Non-crop	Pelpidas mathias	15.89%	Incorrect
			Spodoptera frugiperda	7.53%	2nd option, correct
	Top in situ	Maize	S. exigua	5.08%	incorrect
	Top	Maize	S. frugiperda	92.64%	correct
	Mounted in natural pose	Non-crop	S pectin	41.76%	incorrect
FAW female	Cropped	Non-crop	S. frugiperda	60.22%	correct
FAW male	Cropped	Non-crop	S. mauritia	27.22%	incorrect
			S. frugiperda	27.15%	correct
Different light intensity, background and pose
Cotton bollworm, Helicoverpa armigera	Top-cropped	Maize	Helicoverpa armigera	71.59%	correct
	Top-cropped	Maize	H. armigera	52.54%	correct
	Top mounted	Non-crop	H. armigera	95.56%	correct
	Non-cropped-new specimen	crop	Achyra rantatis	77.05%	incorrect
	Top	Non-crop	H. armigera	92.19%	correct
	Left halfwing open-top	Non-crop	H. armigera	78.28%	correct
Different light intensity and pose
Pink bollworm, Pectinophora gossypiella	Top –non-cropped	Non-crop	Hofmannophila pseudospretella	68.19%	incorrect
	Side-noncropped	Cotton	H.pseudopretella	98.1%	incorrect
	Side	Cotton	H. pseudopretella	95.82%	incorrect
	Top	Cotton	H. pseudopretella	85.36%	incorrect
Different light intensity  and pose
Egyptian bollworm	Side	Cotton	Earias insulana	99.71%	correct
	Side	Cotton	E. insulana	73.34%	correct
Effect of cropping the insect image
W&R masked noctuid	Top-cropped	Leaf	Aegocera venulia	99.99%	Not known
	Top-non-cropped 1/5th size	Leaf	A.      Venulia	34.67%	Not known
Larva identified
Cotton leafworm or Tobacco cutworm, Spodoptera litura Larva	Top	Cotton flower	Neogurelca hyas	18.63%	incorrect
Adult moth	Top	Non-crop	S. litura (Oriential leafworm moth)	40.38%	correct

 

Cropped-refers to limiting to image of moth with minimum background as provided in the software, non-cropped-as photographed with its background; non-crop background is plain. All images were in natural pose of adults, and not mounted as required for taxonomy. Not known-to author.

Acknowledgements: Thanks are due to Pathangasuchaka.in for use of this system after logging on it.