A plant disease diagnosis method based on processing images that display a sick plant part, is tested here on four pear diseases. In the original classification method employed, an extensible feature ranking method had been adopted. Specifically, a discrete grade was assigned to each invariant feature depending on whether its value was found within predefined strict or loose limits. The potential classes that may correspond to the test sample, were sorted according to the sum of the grades of the features. In the current implementation, a Gaussian behavior is assumed and an analog feature grade is assigned depending on the distance of the extracted feature value from the mean. Modeling feature values with Gaussian distribution would eliminate the need of heuristic selection of the strict/loose feature limits but would make more difficult the extension of the supported set of diseases by the end user. However, the experimental results show that the feature values do not follow a Gaussian distribution since comparable classification precision results are obtained only when a normalization scheme is applied to the input image.
Plant monitoring is important in precision agriculture and similar Internet of Things
(IoT) applications [
In this paper, an alternative implementation of the smart phone application
originally presented in [
The alternative classification method tested here, assumes that the values of each feature found in images displaying the same disease, follow a Gaussian Distribution (GD). During the training phase, the various values of a specific feature are used to estimate its mean and variance that are stored in the signature (recognition rules) of a specific disease. The grade assigned to each feature depends, in this case, on the distance of the current feature value from the mean. Two cases are examined: a) use of the original images without normalization and b) use of Linear Dynamic Range Expansion (LDRE normalization) for each one of the RGB colors to moderate the differences in the light exposure of each photograph. Sensitivity, specificity, precision and accuracy metrics are used to evaluate the LT and GD classification methods. The pear diseases tested with the methods described above include: Fire Blight, Pear Scab, Mycosphaerella and Powdery Mildew. Fifty photographs displaying the upper surface of a pear leaf have been tested from each disease.
The LT, GD classification methods are described in Section 2. The experimental results are presented and discussed in Section 3.
The photographs of plant parts captured by the user must display indicative lesions
(spots) of a specific disease. The plant parts can be leaves or fruits or even roots,
branches, flowers, etc. The user initially selects the user interface language, the image
normalization type and the halo zone around the lesion spots with its thickness
expressed as a number of pixels (see Fig. 1a). Then, the photograph is analyzed by
potentially configuring the thresholds for the separation of the background, the normal
plant part and the lesion. In the present application version, separate thresholds are used
for each color component of the RGB color scale (red, green, blue) as shown in Fig. 1b.
Additional information can be given by the user to assist the disease classification (e.g.,
weather data of the specific rural region where the plant exists, located by Global
Positioning System-GPS). The smart phone application was initially developed for
Windows Phone and then ported to Android platforms. The classification methods are
described in detail and the reader can refer to [
The following invariant features are extracted by the segmented photograph: number of lesion spots, their area, the average gray level of each region (normal plant part, spots, halo). A separate histogram is generated for each color component (red, green, blue) and each ROI (normal, spots, halo). It is assumed that photographs displaying the same disease would have similar features like the ones listed above. Instead of attempting to match full histograms, their similarity is simply checked by comparing where they start, where they end and the where their peak is (the histograms usually consist of a single lobe).
When the input photograph is analyzed two triplets of thresholds are used to
separate the background and the spots. If all the red, green, blue values of a pixel (pr,
pg, pb, respectively) are higher than the corresponding background thresholds (BGr,
BGg, BGb) then, this pixel is mapped to the background since the background in this
implementation is simply assumed to be much brighter than the plant part. The rest of
the pixels are initially mapped as normal plant part (e.g., the leaf in the cases examined
here). The lesion spots can either be darker or brighter (the user indicates this fact to
the application through the Invert checkbox shown in Fig. 1b), than the normal leaf. If
all the pr, pg, pb values of a leaf pixel are lower than the corresponding thresholds THr,
THg, THb and the lesion is darker than the normal leaf color, this pixel is mapped to
the lesion region. If the pr, pg, pb values of a leaf pixel are higher than the corresponding
thresholds THr, THg, THb and the lesion is brighter than the normal leaf, again this
pixel is mapped to the lesion region. Finally, the pixels existing in a zone around the
spots with thickness defined by the corresponding field shown in Fig. 1a, are mapped
to the halo region. The user can also select an image normalization method from the
page shown in Fig. 1a. The supported normalization methods are described in detail in
[
The matrix BGW1 is defined with the same size as the original image and each one of its cells can have four distinct values corresponding to normal leaf, spot, hallo and background. Matrix BGW2 can be derived from BGW1 by marking pixels belonging to the same spot with the same identity. BGW2 can be used to estimate all of the features used, in a simple way. The maximum spot identity is the number of spots while the area covered by the spots is estimated by the number of spot pixels divided by the total number of pixels belonging to the plant part. The average gray level of each region can be estimated by averaging the gray level of the pixels mapped to this region in BGW2. Spots consisting of a very small number of pixels (less than the threshold “Min Area” of Fig. 1a) are considered as noise and are not taken into consideration.
During the training phase of the application for a new disease, a small number of representative photos with similar sick plant parts can be analyzed even by an end user that is not aware of the application architecture and implementation details. The range of each feature fi as determined by the training photographs is used to define the strict limits [fi_s1..fi_s2]. This narrow range can be heuristically extended in order to define the loose limits [fi_l1..fi_l2] of this feature for a specific disease. The application lists all these limits to the end user as shown in Fig. 1d, to assist the definition of new diseases or for the customization of the existing diseases. All the feature strict and loose limits are stored in a disease signature exploited during the test phase. When a new photograph is examined in the test phase, all the features fi are extracted and are compared to the predefined limits stored in each disease signature. A rank Gr is estimated for each potential disease using the following equation: 1, !_#% ≤ ! ≤ !_#&.
0, ℎ !_$ = ( 1, !_$% ≤ ! ≤ !_$&.
0, ℎ
The parameters wi_l and wi_s are the individual grades (weights) assigned to feature fi if it is found within the loose or strict limits, respectively of a specific disease. The diseases are sorted according to the rank Gr that they have received for a specific photograph. The three diseases with the highest rank are listed. The simple classification method described above is the LT.
The new classification method (GD) tested in this paper assumes that the feature values that are extracted from the photographs of a disease training set, follow a Gaussian distribution. The mean feature value fm,i, is estimated by the training samples. The rank Gi given for this feature is inversely proportional to the distance of the value of this feature from fm,i. If σ2 is the variance for this feature (also estimated by the training samples), then the disease rank Gr can be estimated by:
'! = ∑!+-,%% (&)*" ,($!%$&,!)" ") " .
(1) (2) (3) (4)
The LDRE normalization employed in the experiments performed in the context of this paper, stretches the values of each color component in the RGB scale according to the following equation: = − − + (5)
PrevCol is the original color value, MinGray and MaxGray are the minimum and maximum pixel values of the gray version of the image (excluding the background), Range is the desired final range and StartCol is the desired offset where Range starts.
Indicative photographs from the pear diseases that have been tested in the framework of this paper are shown in Fig. 2. The metrics used to compare the LT and GD classification methods are the sensitivity, specificity, precision and accuracy. Let True Positives (TP) be the number of photographs that are correctly recognized as positive to a disease. True Negatives (TN) are the number of photographs that are correctly recognized as negative to a disease. False Positives (FP) are the photographs that are falsely recognized as positives and False Negatives (FN), the ones that are falsely recognized as negatives to a disease. The aforementioned four metrics are defined as: = = + +
+ + (6) (7) (8) (9) =
+ + +
As already mentioned, 50 photographs of each pear disease have been tested. They have been retrieved from pear trees in the Achaia prefecture of Greece. About half of these photographs have been captured under sunlight and the rest, under a canopy in order to test different light exposure. Eight representative (training) photographs have been used to extract the disease recognition signatures in either LT (containing the loose and strict limits of each feature), or GD (containing the mean and variance of each feature). In the first test case, the photographs have not been normalized. The experimental results for this case are shown in Tables 1 and 2.
As can be seen from the results listed in Tables 1 and 2, GD is much worse than LT since 2 of the 4 diseases (fire blight and powdery mildew) cannot be recognized at all: their sensitivity is 0 although the specificity achieved is more acceptable. Using LT as the classification method, only pear scab has a low sensitivity but all the average metrics are better than GD.
If image normalization with LDRE is used, the results obtained are listed in Tables 3 and 4. Although the average values of the 4 metrics obtained for GD are still worse than the ones achieved by LT, they are much closer with LDRE normalization. In any case, these results show that although Gaussian seems a more natural way to describe the distribution of the various feature values, the classification accuracy achieved by GD is worse than the one achieved by LT. The positive aspect of this conclusion is that the rules in the disease signatures used by LT can be defined in a simpler way by an end-user that is not familiar with the implementation details of the developed application: it is easier for an end-user to determine the limits in a list of feature values than estimate the mean values and variances required by the GD. For this reason, the employed LT classification method allows the extension of the supported set of diseases by an end user that is not particularly qualified in computer science or statistics.
Two classification methods are tested in the framework of a mobile phone application capable of recognizing pear diseases. The experimental results showed that the classification method based on the assumption that the invariant image feature values follow a Gaussian distribution in images displaying the same disease, does not achieve a better accuracy than the heuristic classification method that compares if the feature values reside within predefined strict and loose limits. The benefit from the classification method based on Gaussian distribution is that it favors the extensibility of the supported set of diseases by a non-expert. Image normalization significantly improves the accuracy of this classification method.
Based on the experimental results presented in this paper, future work may focus on supporting multiple ranges of feature values for comparison e.g., too strict, strict, loose, too loose, etc, in order to improve further the accuracy.
Acknowledgments. This work is protected by the provisional patents 1009346/13-82018 and 1008484/12-5-2015 (Greek Patent Office). The author wishes to thank the student Olympia Sakorafa for her contribution in the experiments carried out for this paper.