Automatic measurement of stomatal density from microphotographs

Abstract

Key message

An automated process using a cascade classifier allowed the rapid assessment of the density and distribution of stomata on microphotographs from leaves of two oak species.

Abstract

Stomatal density is the number of stomata per unit area, an intensively studied trait, involved in the control of CO₂ and H₂O exchange between leaf and atmosphere. This trait is usually estimated by counting manually each stoma on a given surface (e.g., a microphotograph), usually repeating the procedure with images from different parts of the leaf. To improve this procedure, we tested the performance of a cascade classifier to automatically detect stomata on microphotographs from two oak species: Quercus afares Pomel and Quercus suber L. The two species are phylogenetically close with similar stomatal morphology, which allowed testing the reuse of the cascade classifier on stomata with similar shape. The results showed that a cascade classifier trained on only 100 sample views of stomata from Q. afares was able to rapidly detect stomata in Q. afares as well as in Q. suber with a very low number of false positives (5 %/1.9 %) and a small number of undetected stomata (14.8 %/0.74 %), when partial stomata near the edge of the microphotographs were ignored. The remaining undetected stomata were due to obstacles such as trichomes. As an example of further applications, we used the positions detected by the cascade classifier to assess the spatial distribution of stomata and group them on the leaf surface. To our knowledge this is the first time that a cascade classifier has been applied to plant microphotographs, and we were able to show that it can dramatically decrease the time needed to estimate stomatal density and spatial distribution.

Appendix

Data preparation

The command line to apply distortions on each microphotograph was:

opencv_createsamples -img./Pos/img.tif -num 100 -bg bg.txt -vec samples.vec -maxxangle 0.5 -maxyangle 0.5 -maxzangle 6.7 -maxidev 10 -bgcolor 120 -bgthresh 0 -w 9 -h 9.

Options	Values	Descriptions
-img	./Pos/img.tif	A positive sample view
-num	100	The number of positive samples to be generated
-bg	bg.txt	Background description file; contains a list of images, which are used as a background for randomly distorted versions of the object
-vec	samples.vec	Name of the output file containing the positive samples for training
-maxxangle	0.5	The maximum rotation angle in x-direction in radians
-maxyangle	0.5	The maximum rotation angle in y-direction in radians
-maxzangle	6.7	The maximum rotation angle in z-direction in radians
-maxidev	10	The desired maximum intensity deviation of foreground samples’ pixels
-bgcolor	120	The background colour considered as transparent
-bgthresh	0	All pixels within bgcolor±bgthresh are interpreted as transparent
-w	9	The resulting sample width
-h	9	The resulting sample height

The command line to merge all binary files into one (merge.vec) used as input for a text file containing all filenames (samples.txt):

mergevec samples.txt merge.vec -w 9 -h 9.

Options	Values	Descriptions
-w	9	The resulting sample width
-h	9	The resulting sample height

Training

The command line for training cascade classifier was:

opencv_traincascade -data./traininghaar -vec merge.vec -bg bg.txt -numPos 10,000 -numNeg 3,000 -numStages 18 -precalcValBufSize 500 -precalcIdxBufSize 500 -featuretype haar -w 9 -h 9 -minHitRate 1 -maxFalseAlarmRate 0.45 -mode BASIC -maxWeakCount 200 -weightTrimRate 0 -maxDepth 2

Options	Values	Descriptions
-data	./TrainingHAAR	The directory for the output files
-vec	merged.vec	Name of the input file containing the positive samples
-bg	bg.txt	Background description file; contains a list of images, which are used as a background for randomly distorted versions of the object
-numPos	10,000	Number of positive samples used in training for every classifier stage
-numNeg	3,000	Number of negative samples used in training for every classifier stage
-numStages	18	Number of cascade stages to be trained
-precalcValBufSize	500	Size of buffer for precalculated feature values (in Mb)
-precalcIdxBufSize	500	Size of buffer for precalculated feature indices (in Mb). The more memory you have, the faster is the training process
-featureType	HAAR	Type of features
-w	9	The resulting sample width
-h	9	The resulting sample height
-minHitRate	1	Minimal desired hit rate for each stage of the classifier. Overall hit rate may be estimated as (min_hit_rate^number_of_stages)
-maxFalseAlarmRate	0.45	Maximal desired false alarm rate for each stage of the classifier. Overall false alarm rate may be estimated as (max_false_alarm_rate^number_of_stages)
-mode	BASIC	Selects the type of haar feature set used in training. BASIC use only upright features
-maxWeakCount	200	Maximal count of weak trees for every cascade stage. The boosted classifier (stage) will have as many weak trees (≤maxWeakCount) as needed to achieve the given-maxFalseAlarmRate
-weightTrimRate	0	Specifies whether trimming should be used and its weight
-maxDepth	2	Maximal depth of a weak tree