Hyperparameter-Tuned K-Nearest Neighbor for Herbal Leaf Classification Using Shape, Color, and Texture Feature Fusion in Smart Agriculture

Vol. 8, No. 5, October 2026

6604
Journal of Information Systems and Informatics | ISSN: 2656-5935 | e-ISSN: 2656-4882 | pp. 6604-6629
Published by Asosiasi Doktor Sistem Informasi Indonesia
Hyperparameter-Tuned K-Nearest Neighbor for Herbal Leaf
Classification Using Shape, Color, and Texture Feature Fusion in
Smart Agriculture

Avip Kurniawan1, Agustinus Eko Setiawan2, Vevi Maritha3, Muhammad Abrar Argianata4
1,4Department of Informatics Engineering, Faculty of Technology, Universitas Krisnadwipayana, Jakarta, Indonesia
2Department of Informatics Engineering, Faculty of Technology and Business, Universitas Aisyah Pringsewu,
Lampung, Indonesia
3Pharmacy Study Program, Faculty of Health and Science, Universitas PGRI Madiun, Madiun, Indonesia

Received:
May 20, 2026
Revised:
September 9, 2026
Accepted:
September 27, 2026
Published:
October 24, 2026

Corresponding Author:
Author Name*:
Avip Kurniawan
Email*:
avip.kurniawan@unkris.ac.id

DOI:
10.63158/journalisi.v8i5.1938

© 2026 Journal of
Information Systems and
Informatics. This open
access article is distributed
under a (CC-BY License)

Abstract. This study develops a lightweight K-Nearest Neighbor (K-
NN) classifier to identify five Indonesian herbal leaf species—betel,
bay, guava, soursop, and moringa—for potential smart-agriculture
applications. The initial dataset contained 1,000 leaf images and was
expanded to 6,000 images through offline augmentation. After data
cleaning, 5,884 images were retained for analysis. Each image was
resized to 256 × 256 pixels, converted into grayscale and HSV color
spaces, and segmented using Otsu thresholding. A total of 21
features representing shape, color, and Gray-Level Co-occurrence
Matrix (GLCM)-based texture characteristics were extracted. K-NN
hyperparameters were optimized through Grid Search with
Stratified 5-Fold Cross-Validation using 80:20, 70:30, and 60:40
training–testing splits. Under image-level evaluation, the best
configuration used Manhattan distance, k = 1, and distance
weighting, achieving 92.18% accuracy and a 92.29% macro F1-score
on the 80:20 split. The 70:30 and 60:40 splits achieved accuracies
of 91.79% and 91.21%, respectively. Misclassifications primarily
occurred between soursop and guava leaves. Although the
framework is compact and interpretable, augmentation before
splitting may have produced optimistic results. Future work should
apply source-grouped validation and assess real on -device
performance.

Keywords: Herbal Leaf Classification, K-Nearest Neighbor, Feature
Extraction, Hyperparameter Tuning, Smart Agriculture

Vol. 8, No. 5, October 2026

6605 | Hyperparameter-Tuned K-Nearest Neighbor for Herbal Leaf Classification …..

1. INTRODUCTION

Herbal plants are an integral part of traditional medicine practices in many countries,
including Indonesia. The World Health Organization estimates that approximately 80% of
the global population still relies on plant-based medicine as part of primary healthcare
services. Indonesia alone is home to more than 30,000 medicinal plant species,
approximately 9,600 of which have been shown to possess pharmacological properties
[1]. This potential highlights the need for an accurate, rapid, and widely accessible system
for identifying medicinal plant species to the wider community [2], [3].

In practice, herbal leaf identification is still largely performed by visual inspection, making
the result subjective when species share similar shape, color, or surface texture [4].
Misidentification is not merely taxonomic because an incorrectly selected medicinal leaf
can lead to inappropriate herbal preparation and potential safety risks. An objective
image-based identification aid is therefore relevant for users who do not have ready
access to botanical expertise.

Digital image processing and machine learning provide an alternative for automated
identification. Leaf appearance can be represented by complementary shape, color, and
texture descriptors [5], [6]. HSV separates chromatic information from brightness, while
GLCM characterizes local texture patterns [7], [8]. K-Nearest Neighbor (K-NN) is attractive
for field applications because it is non-parametric, computationally lightweight, and
traceable to stored training samples [9].

Previous studies have explored different feature-model combinations. Meiriyama et al.
[5] obtained 92.67% accuracy using HOG-LBP with K-NN, while Naeem et al. [10] reported
99.01% using multispectral and texture features under laboratory acquisition. Other
studies used GLCM with K-NN [7], [8], weighted K-NN [11], or color-GLCM-shape features
with conventional classifiers [4]. CNN-based approaches [12], [13] generally achieved
higher accuracy but with greater computational requirements, while HSV -K-NN
performance can remain sensitive to illumination [1].

A few researchers have combined all three visual feature groups, and fewer still have
systematically tuned the K-NN configuration. Two gaps therefore remain: incomplete
visual representation when only one or two feature groups are used, and uncertain K-

Vol. 8, No. 5, October 2026

Avip Kurniawan, Agustinus Eko Setiawan, et al | 6606
NN performance when k, weighting, and distance metric are left at default or ad-hoc
values. Because distance choice can materially change K-NN performance [14], both
feature fusion and hyperparameter tuning need to be evaluated together.

Therefore, this research intends to develop a lightweight K-NN model for five Indonesian
herbal leaf species using fused shape, HSV color, and GLCM texture features. The
objectives are to (1) construct a 21-dimensional feature vector, (2) identify the optimal K-
NN configuration through Grid Search with stratified cross-validation, and (3) evaluate
accuracy, precision, recall, F1-score, execution time, and stability across 80:20, 70:30, and
60:40 splits. The study is oriented toward smart -agriculture settings where local
processing and limited resources favor efficient models [15], [16]. The selected species—
betel, bay, guava, soursop, and moringa—are commonly used medicinal plants in Indonesia
[2], [3], [17]. The contributions are the joint use of 21 handcrafted shape-color-texture
features, systematic K-NN hyperparameter tuning across three split scenarios, and class-
level error analysis to identify the most difficult species pair and guide future feature
improvement.

2. METHODS

This study uses an experimental quantitative approach based on the Cross Industry
Standard Process for Data Mining (CRISP-DM). The analytical stages cover business and
data understanding, data preparation, modeling, and evaluation; direct deployment is
outside the present scope [18]. Implementation used Python, OpenCV, scikit-image, and
scikit-learn in Google Colaboratory [19].

As summarized in Figure 1, the reported image-level experimental pipeline consists of
image acquisition, resizing to 256 × 256 pixels, grayscale/HSV conversion, threshold
segmentation, augmentation and cleaning, extraction and concatenation of 2 shape, 3
HSV, and 16 GLCM features, StandardScaler normalization, Grid Search with Stratified 5-
Fold Cross-Validation, and held-out evaluation. This figure describes the protocol that
produced the numerical results reported in this manuscript; Accordingly, the reported
results reflect the original image-level protocol and do not constitute source-level
validation.

Vol. 8, No. 5, October 2026

6607 | Hyperparameter-Tuned K-Nearest Neighbor for Herbal Leaf Classification …..

Figure 1. CRISP-DM-based workflow of the reported image-level experiment from image
acquisition and preprocessing to 21-feature fusion, K-NN tuning, and evaluation

2.1. Data Collection
The data used in this study consist of digital leaf images from five classes. The primary
data were obtained through direct acquisition using a high-resolution digital camera
under natural lighting conditions, while the secondary data were obtained from the public
Kaggle repository. Each class was represented by 100 primary images and 100 secondary
images, resulting in an initial total of 1,000 images with a balanced distribution across
classes. The combination of these two sources was used to enrich variati ons in
background, capture angle, and lighting conditions, so that the model would not learn
only the characteristics of a single acquisition condition. Sample images from both
sources are presented in Figure 2, and the dataset composition is detailed in Table 1.

Vol. 8, No. 5, October 2026

Avip Kurniawan, Agustinus Eko Setiawan, et al | 6608

Figure 2. Sample images of the five herbal leaf species from the primary and secondary
sources

Figure 2 illustrates the visual challenges addressed in this study. Betel leaves can be
clearly distinguished by their heart-shaped blades, while moringa leaves can be identified
by their small, rounded leaflets. In contrast, guava and soursop leaves both have
elongated elliptical shapes and dark green surfaces, making their differences primarily
dependent on subtle details of leaf venation and surface gloss. The figure also shows
differences in background and lighting conditions between the two data source s,
representing variations that the model must be able to tolerate.

Table 1. Composition of the herbal leaf image dataset
No
Leaf Species
(Scientific Name)
Primary
Data
Secondary
Data
Initial
Total
After
Augmentation
1 Betel (Piper betle) 100 100 200 1,200
2
Bay (Syzygium
polyanthum)
100 100 200 1,200
3
Guava (Psidium
guajava)
100 100 200 1,200
4
Soursop (Annona
muricata)
100 100 200 1,200
5
Moringa (Moringa
oleifera)
100 100 200 1,200
Total 500 500 1,000 6,000

Vol. 8, No. 5, October 2026

6609 | Hyperparameter-Tuned K-Nearest Neighbor for Herbal Leaf Classification …..
2.2. Image Preprocessing and Augmentation
The preprocessing stage aims to standardize the quality and format of the images before
feature extraction. All images were resized to 256 × 256 pixels to standardize the input
dimensions while reducing the computational load. The RGB images were then converted
to grayscale for shape and texture feature extraction and to the HSV color space for
color feature extraction, as shown in Figure 3. The leaf objects were separated from the
background using thresholding-based segmentation, which produced binary images as
shown in Figure 4, because an unremoved background could potentially introduce noise
into the calculation of shape features. The threshold was determined automatically using
Otsu's method, which selects the value that maximizes the between-class variance of the
grayscale histogram, so that no manually tuned threshold was required. Because the
polarity of the result depends on w hether the leaf is darker or lighter than its
background, the binary image was inverted whenever more than half of its pixels were
labelled as foreground. Residual speckle and small gaps were then removed by
morphological opening followed by morphological closing, both using an elliptical
structuring element of 5 × 5 pixels.

Figure 3. Color space conversion of a leaf image into grayscale and HSV
representations

Figure 4. Object segmentation result using thresholding on a grayscale leaf image

Vol. 8, No. 5, October 2026

Avip Kurniawan, Agustinus Eko Setiawan, et al | 6610
The number of original images was considered insufficient to train a model that is robust
to variations in acquisition conditions, so data augmentation was performed. Five
augmentation techniques were applied: 90 -degree rotation, horizontal flipping,
brightness adjustment, contrast adjustment, and noise addition, with the results shown
in Figure 5. Rotation and flipping improve the model’s robustness to variations in
orientation, brightness and contrast adjustments simulate differences in lighting
conditions, while noise addition mimics camera sensor interference and image
compression. The exact parameters were as follows. Rotation applied a 90 -degree
clockwise turn, and flipping was performed about the vertical axis. Brightness adjustment
multiplied the Value channel of the HSV representation by a factor of 1.3 and clipped the
result at 255. Contrast adjustment applied the linear transformation g(x) = alpha x + beta,
with alpha = 1.5 and beta = 0. Noise addition used additive Gaussian noise with a mean of
0 and a standard deviation of 10, with the result clipped to the range [0, 255]. Each original
image produced exactly one variant per technique, so that together with the original
image every source photograph contributed six images to the dataset: 1,000 originals
therefore yielded 5,000 augmented images and 6,000 images in total. All images, original
and augmented alike, were resized to 256 × 256 pixels after the transformation was
applied. Details of the augmentation results are presented in Table 2.

Figure 5. Examples of augmentation results applied to a betel leaf image

Table 2. Number of images produced by each augmentation technique, one variant per
technique per source image
No Treatment
Primary
Images
Secondary
Images
Total Images
1 Original images 500 500 1,000
2 Rotation of 90 degrees 500 500 1,000

Vol. 8, No. 5, October 2026

6611 | Hyperparameter-Tuned K-Nearest Neighbor for Herbal Leaf Classification …..
No Treatment
Primary
Images
Secondary
Images
Total Images
3 Horizontal flip 500 500 1,000
4 Brightness adjustment 500 500 1,000
5 Contrast adjustment 500 500 1,000
6 Noise addition 500 500 1,000
Total 3,000 3,000 6,000

The augmentation process produced 6,000 images ready for further processing. The
subsequent quality inspection identified several images that could not be properly
segmented, particularly those with very low contrast between the object and the
background, causing the leaf contour to be poorly formed and the eccentricity value to
be undefined. Such images were removed during the data cleaning stage, resulting in a
final dataset of 5,884 images used for model development.

2.3. Feature Extraction
Feature extraction transforms the visual information in an image into a numerical
representation that can be processed by a classification algorithm. Three groups of
features were extracted from each image, namely shape, color, and texture, with the
details of the measurements calculated for each group shown in Figure 6.

Figure 6. Flowchart of the shape, color, and texture feature extraction stages

Vol. 8, No. 5, October 2026

Avip Kurniawan, Agustinus Eko Setiawan, et al | 6612
Shape features were extracted from the segmented binary images using two measures,
namely metric (circularity) and eccentricity. Metric measures the circularity of an object
based on its area and contour perimeter, as expressed in Equation (1), where a value
approaching 1 indicates a shape that is close to circular. Eccentricity measures the degree
of elongation based on the relationship between the major and minor axes of an ellipse
fitted to the object contour, as expressed in Equation (2), with values ranging from 0 for
a rounded shape to values approaching 1 for a highly elliptical shape.

𝑀𝑒𝑡𝑟𝑖𝑐=4𝜋𝐴/𝑃! (1)
𝐸𝑐𝑐𝑒𝑛𝑡𝑟𝑖𝑐𝑖𝑡𝑦=0(1−𝑏!/𝑎!) (2)

where A denotes the area of the object, P denotes the contour circumference, a denotes
the semi-major axis length, and b denotes the semi-minor axis length of the ellipse.

Color features were extracted in the HSV color space because this color space separates
the pure color component (Hue), color intensity (Saturation), and brightness (Value),
making it more closely aligned with human color perception and more stable to changes
in lighting conditions than RGB. The mean value of each channel was calculated as shown
in Equation (3), resulting in three color features: mean_H, mean_S, and mean_V.

𝑚𝑒𝑎𝑛"=(1/𝑁)𝛴(𝑖=1𝑡𝑜𝑁)𝐶# (3)

where C denotes the color channel under consideration, C(i) denotes the intensity value
of the i-th pixel in that channel, and N denotes the total number of pixels in the image.
Texture features were extracted using GLCM, which analyzes the frequency of
occurrence of pixel pairs with specific gray levels at a given distance and direction. This
study used a distance of d = 1 with four orientation angles, namely 0°, 45°, 90°, and 135°,
at 256 gray levels using a symmetric and normalized matrix. Four statistical properties
were calculated from the matrix: contrast, which measures intensity differences between
neighboring pixels (Equation 4); energy, which measures distribution uniformity (Equation
5); homogeneity, which measures the proximity of the element distribution to the main

Vol. 8, No. 5, October 2026

6613 | Hyperparameter-Tuned K-Nearest Neighbor for Herbal Leaf Classification …..
diagonal (Equation 6); and correlation, which measures the linear relationship between
neighboring pixels (Equation 7).
𝐶𝑜𝑛𝑡𝑟𝑎𝑠𝑡=𝛴#𝛴$(𝑖−𝑗)!𝑃(𝑖,𝑗) (4)
𝐸𝑛𝑒𝑟𝑔𝑦=𝛴#𝛴$𝑃(𝑖,𝑗)! (5)
𝐻𝑜𝑚𝑜𝑔𝑒𝑛𝑒𝑖𝑡𝑦=𝛴#𝛴$𝑃(𝑖,𝑗)/(1+|𝑖−𝑗|) (6)
𝐶𝑜𝑟𝑟𝑒𝑙𝑎𝑡𝑖𝑜𝑛=𝛴#𝛴$D(𝑖−𝜇#)F𝑗−𝜇$G𝑃(𝑖,𝑗)H/F𝜎#𝜎$G (7)

where 𝑃(𝑖,𝑗) denotes the normalized GLCM element at row i and column j, while μ and σ
denote the marginal mean and standard deviation of the GLCM, respectively.

Since each property was calculated at four angles, a total of 16 texture features were
obtained. All features from the three groups were then combined based on matching file
names and class labels, resulting in a single 21-dimensional feature vector for each image,
as detailed in Table 3.
Table 3. Composition of the combined feature vector
Feature
Group
Method Feature Names Count
Shape Binary image contour metric, eccentricity 2
Color HSV color space mean_H, mean_S, mean_V 3
Texture
GLCM (d = 1; 0, 45, 90,
135 degrees)
contrast, energy, homogeneity, and
correlation at each angle
16
Total dimension of the feature vector 21

2.4. Normalization and Data Splitting
The extracted features have substantially different value ranges; GLCM contrast values
can range from tens to hundreds, whereas metric and eccentricity range from 0 to 1. This
difference in scale has a substantial effect on K-NN because the algorithm makes
decisions based on distances between data points, causing features with larger scales to
dominate the calculation while features with smaller scales have almost no influence.
Therefore, all features were normalized using standardization (StandardScaler), which
transforms each feature to have a mean of 0 and a standard deviation of 1. The

Vol. 8, No. 5, October 2026

Avip Kurniawan, Agustinus Eko Setiawan, et al | 6614
transformation parameters were calculated using only the training data and then applied
to the test data to prevent information leakage.

Data splitting was performed using the train_test_split function with the stratify
parameter to preserve the proportion of each class in both the training and test subsets,
and random_state = 42 to ensure reproducibility. Three data-splitting scenarios were
evaluated: 80:20, 70:30, and 60:40. In the primary 80:20 scenario, the dataset was divided
into 4,707 training images and 1,177 test images.

Because offline augmentation was performed before train -test splitting,
augmented variants from one source photograph could occur in both subsets.
The reported results therefore represent an image-level evaluation and may
overestimate performance on independent source photographs, particularly for k
= 1 where a very close variant can determine the prediction. The same mechanism
can also reduce apparent variance across folds and split scenarios. A leakage-
controlled source-level grouped evaluation was not conducted in this study;
consequently, the reported results should be interpreted within the original
image-level protocol.

2.5. Modeling and Hyperparameter Tuning
The classification algorithm used in this study was K-Nearest Neighbor (K-NN). K-
NN assigns the class of a test image based on the majority of labels from its k
nearest neighbors in the feature space, without constructing a parametric model
at the training stage. The behavior of this algorithm is determined by three main
parameters: the number of neighbors (n_neighbors), which controls the
smoothness of the decision boundary; the weighting scheme (weights), which
determines whether all neighbors have equal influence or are weighted according
to their proximity; and the distance metric (metric), which determines how
similarity between feature vectors is measured [14]. These three parameters were
optimized through Grid Search, which was configured to search the combinations
of values within the predefined search space and has been shown to significantly
improve model performance compared with the use of default parameters [20].

Vol. 8, No. 5, October 2026

6615 | Hyperparameter-Tuned K-Nearest Neighbor for Herbal Leaf Classification …..
The validation scheme used was Stratified 5-Fold Cross Validation, with accuracy
as the scoring metric. The search space used is presented in Table 4.

Table 4. Hyperparameter search space for the K-NN algorithm
Algorithm Parameter Tested Values
K-NN
n_neighbors 1, 3, 5, 7, 9
weights uniform, distance
metric euclidean, manhattan

2.6. Evaluation Metrics
Model performance was evaluated using a confusion matrix along with four derived
metrics: accuracy, precision, recall, and F1-score, as defined in Equations (8)–(11) [9].
Accuracy measures the proportion of correct predictions among all observations,
precision measures the correctness of positive predictions, recall measures the model’s
ability to identify all positive instances, and F1-score is the harmonic mean of precision
and recall, which is useful when both aspects are equally important. Because this is a
multiclass classification problem with a balanced class distribution, the last three metrics
were reported for each class and summarized using the macro average [21], [22].

𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦=(𝑇𝑃+𝑇𝑁)/(𝑇𝑃+𝑇𝑁+𝐹𝑃+𝐹𝑁) (8)
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛=𝑇𝑃/(𝑇𝑃+𝐹𝑃) (9)
𝑅𝑒𝑐𝑎𝑙𝑙=𝑇𝑃/(𝑇𝑃+𝐹𝑁) (10)
𝐹1−𝑠𝑐𝑜𝑟𝑒=2×𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛×𝑅𝑒𝑐𝑎𝑙𝑙/(𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛+𝑅𝑒𝑐𝑎𝑙𝑙) (11)

Training and prediction time were measured as computational-efficiency indicators.
Experiments used Python with OpenCV, scikit-image, and scikit-learn; timing was obtained
on an AMD Ryzen 5 7535HS CPU with 16 GB RAM and no GPU. Edge feasibility was
assessed only through classifier prediction time and memory footprint because no
smartphone or embedded board was benchmarked directly. The reported 0.19 ms per
image therefore refers to K-NN prediction after the 21-dimensional feature vector has
been extracted, not to the complete image-to-prediction pipeline. The full 5,884-vector

Vol. 8, No. 5, October 2026

Avip Kurniawan, Agustinus Eko Setiawan, et al | 6616
dataset occupies about 0.94 MiB, while the 4,707-vector training database in the 80:20
experiment requires about 0.75 MiB. These values indicate low model-side storage and
prediction cost, but image preprocessing and feature extraction would also contribute
to latency in a real deployment and should be benchmarked on the target hardware.

3. RESULTS AND DISCUSSION

3.1. Characteristics of Shape, Color, and Texture Features
The distribution of shape features across the entire dataset is presented as a scatter
plot of eccentricity and metric in Figure 7. Soursop and guava leaves occupy the high-
eccentricity region, reflecting their elongated elliptical shapes, whereas betel and guava
leaves account for most of the points with metric values above 0.6 as a consequence of
their more rounded leaf blades. Most points from the five classes remain concentrated
within the metric band below 0.25, clearly indicating that the two shape features can
separate only some of the classes and are not sufficient when used independently.

Figure 7. Scatter plot of the eccentricity and metric shape features for the five leaf
classes

Vol. 8, No. 5, October 2026

6617 | Hyperparameter-Tuned K-Nearest Neighbor for Herbal Leaf Classification …..
The distribution of color features in the three-dimensional space of mean Hue, mean
Saturation, and mean Value is presented in Figure 8. The five classes form an overlapping
data cloud, with the highest density in the Hue range of 40 to 80 and the Value range of
100 to 200, which is reasonable considering that all objects are green leaves. Bay leaves
appear to exhibit the widest spread toward higher Value levels, whereas moringa leaves
dominate the low-Saturation region. The mean value of the Hue channel re mains
relatively stable under geometric augmentation, whereas the Value channel shifts in
response to brightness and contrast adjustments, consistent with the nature of the HSV
color space in separating pure color information from light intensity.

Figure 8. Three-dimensional scatter plot of the HSV color features for the five leaf
classes

The relationships among the GLCM texture features are presented as a pairplot in Figure
9. The contrast feature has a right-skewed distribution, with most values below 300,
whereas correlation is narrowly concentrated around values close to 1. The energy and
homogeneity features exhibit an almost linear positive relationship, while contrast is
negatively correlated with correlation, energy, and homogeneity, consistent with the

Vol. 8, No. 5, October 2026

Avip Kurniawan, Agustinus Eko Setiawan, et al | 6618
interpretation that high contrast indicates sharp intensity differences, whereas energy
and homogeneity indicate uniformity. Moringa leaves are most clearly separated in the
region characterized by low contrast and high energy and homogeneity, which explains
why this class is the easiest for the K-NN model to recognize during classification.

Figure 9. Pairplot of the GLCM texture features contrast, correlation, energy, and
homogeneity

In terms of robustness to augmentation, the correlation values of the original images as
well as the rotated and flipped images ranged from 0.97 to 0.98 across all orientations,
indicating that the texture structure was preserved under geometric transformations.
These values decreased slightly in images subjected to contrast adjustment, brightness
adjustment, and noise addition, indicating that the GLCM features are relatively robust
to mild disturbances but remain sensitive to changes in pixel patterns cau sed by
photometric augmentation.

Vol. 8, No. 5, October 2026

6619 | Hyperparameter-Tuned K-Nearest Neighbor for Herbal Leaf Classification …..
3.2. Hyperparameter Tuning Results
The Grid Search results in Table 5 consistently selected Manhattan distance and distance
weighting, with k = 1 for the 80:20 split and k = 3 for the 70:30 and 60:40 splits. Manhattan
(L1) sums absolute deviations, whereas Euclidean (L2) gives a large deviation on one
feature greater influence through squaring. This distinction is important in the
heterogeneous 21-dimensional feature space: metric and eccentricity are bounded shape
measures, HSV means are intensity statistics, and GLCM contrast is right -skewed.
Standardization equalizes scale but does not make these distributions identical.
Consequently, an atypical contrast or illumination value can dominate an L2 distance
more strongly than an L1 distance, while Manhattan preserves the contribution of the
remaining dimensions. The result is especially relevant for leaf images because local
glare, shadow, segmentation residue, or mild noise can produce a large standardized
deviation in only one or two texture/color coordinates without changing the overall class
identity. Under L2, those coordinates receive disproportionate influence; under L1, their
contribution remains linear, so agreement across the other descriptors continues to
matter.

This provides a plausible technical explanation for Manhattan being selected in all three
splits rather than only one favorable partition, and it agrees with evidence that absolute-
difference measures can be more robust to noise and outliers [14]. For k = 1, uniform and
distance weighting yield the same class decision because only one neighbor votes;
weighting becomes meaningful in the k = 3 scenarios, where closer samples receive
greater influence [11]. The selection of k = 1 must nevertheless be regarded as provisional.
Under the present protocol the training set contains augmented siblings of the test
images, and such a sibling is by construction an extremely close neighbor, which
systematically favours the smallest possible k. Once all variants of one source
photograph are confined to the same partition, those siblings are no longer available, the
distance to the first neighbor increases, and a larger k may well be selected. A source-
level grouped evaluation would therefore require re -optimizing the complete
hyperparameter grid rather than transferring the current configuration to a new split.

Vol. 8, No. 5, October 2026

Avip Kurniawan, Agustinus Eko Setiawan, et al | 6620
Table 5. Hyperparameter tuning results and K-NN performance under the original
image-level splitting protocol
Algorithm Split Best Parameters
CV
Accuracy
(Training)
Accuracy
(Testing)
Training
Time (s)
Prediction
Time (s)
K-NN 80:20
metric =
manhattan;
n_neighbors = 1;
weights = distance
0.9191 0.9218 6.18 0.2198
K-NN 70:30
metric =
manhattan;
n_neighbors = 3;
weights = distance
0.9094 0.9179 3.12 0.1753
K-NN 60:40
metric =
manhattan;
n_neighbors = 3;
weights = distance
0.9051 0.9121 3.82 0.2257

Table 5 shows that test accuracy changed by 0.97 percentage points, from 92.18% (80:20)
to 91.21% (60:40), and declined as the training set became smaller. This difference is
deliberately not interpreted here as evidence of strong stability. Under image -level
splitting, augmented variants of the same source photograph are distributed across the
training and test subsets in every scenario, which compresses the variance between
scenarios by construction; a small spread is therefore an expected property of the
protocol rather than an independent finding about the feature representation. Whether
the representation is genuinely stable across splitting proportions can only be
established under source-level grouped validation. Test accuracy was slightly above
cross-validation accuracy in all scenarios, showing a small difference between cross-
validation and test performance within this protocol but not proving generalization to
independent source images.

K-NN also showed low computational cost. In the 80:20 scenario, parameter search and
fitting required 6.18 s and prediction of 1,177 images required 0.2198 s, or about 0.19 ms

Vol. 8, No. 5, October 2026

6621 | Hyperparameter-Tuned K-Nearest Neighbor for Herbal Leaf Classification …..
per image. As a lazy learner, K-NN shifts cost to inference because each test vector is
compared with stored training vectors; consequently, memory and prediction cost scale
with the size of the feature database.

3.3. K-NN Model Evaluation
The confusion matrix of the K-NN model with the best configuration under the 80:20
scenario is presented in Figure 10. The majority of predictions lie along the main diagonal,
with 1,085 correct predictions out of 1,177 test samples, resulting in an accuracy of 92.18%.

Figure 10. Confusion matrix of the K-NN model under the original image-level 80:20
splitting protocol

Details of the metrics for each class are presented in Table 6. Moringa leaves achieved
the best performance with an F1-score of 98.67%, followed by betel leaves at 97.47% and
Bay leaves at 95.74%. Conversely, guava and soursop leaves achieved the lowest F 1-
scores, at 85.06% and 84.52%, respectively.

Vol. 8, No. 5, October 2026

Avip Kurniawan, Agustinus Eko Setiawan, et al | 6622
Table 6. Per-class performance of the K-NN model under the original image-level 80:20
protocol
Leaf Class TP FP FN
Precision
(%)
Recall (%)
F1-score
(%)
Guava 205 42 30 83.00 87.23 85.06
Moringa 222 4 2 98.23 99.11 98.67
Bay 225 7 13 96.98 94.54 95.74
Betel 231 3 9 98.72 96.25 97.47
Soursop 202 36 38 84.87 84.17 84.52
Macro average – – – 92.36 92.26 92.29

3.4. Discussion
The optimal k = 1 configuration in the 80:20 split is consistent with dense local
neighborhoods in the standardized feature space, where the nearest sample often
already has the correct label. The augmented dataset also creates compact clusters of
visually similar variants, so enlarging k can recruit samples from adjacent class
boundaries rather than add useful information. This is particularly relevant for the guava-
soursop boundary, where the two classes overlap strongly. At k = 1, uniform and distance
weighting are functionally equivalent; the result therefore reflects locality rather than
an independent advantage of weighting. When the training share is reduced to 70% and
60%, the feature space becomes more sparsely sampled and k = 3 is selected, suggesting
that modest smoothing becomes beneficial when the nearest sample is less reliably
representative. The trade-off is that k = 1 remains more sensitive to anomalous vectors
and to the image-level augmentation protocol discussed in Section 2.4.

The confusion matrix analysis reveals that the largest source of classification errors is
concentrated in a single class pair, namely soursop and guava leaves. A total of 33
soursop images were predicted as guava, while 27 guava images were predicted as
soursop, resulting in 60 errors, or 65.2% of the total 92 errors, originating from these
two classes alone. This highly uneven distribution of errors is important to note because
it indicates that the overall accuracy of the model is primarily limited by one specific
discrimination problem rather than by weaknesses distributed across all classes.

Vol. 8, No. 5, October 2026

6623 | Hyperparameter-Tuned K-Nearest Neighbor for Herbal Leaf Classification …..
This dominant error pattern is also visually interpretable. The two species share
elongated-elliptical contours, dark-green coloration, and glossy surfaces, which causes
their metric, eccentricity, mean HSV, and global GLCM descriptors to overlap. In the
sample images, guava also shows more pronounced secondary venation and surface
relief, whereas soursop appears smoother. The current representation is not designed to
exploit that distinction because the HSV descriptors are global channel means and the
GLCM statistics are aggregated over the entire blade. Higher-order HSV statistics such
as standard deviation, skewness, and normalized histograms could preserve within-leaf
color variation that is lost in the mean, while venation descriptors based on ridge or edge
filtering could encode vein density, dominant orientation, and branching structure. These
additions directly target the most frequent confusion while retaining the computational
advantages of handcrafted features.

For deployment-oriented smart agriculture, K-NN offers a compact stored-feature
representation, no conventional iterative retraining, and traceable predictions linked to
nearest training samples. New labeled examples can be incorporated by updating the
stored feature database rather than retraining a deep network, which is useful when
field collections grow incrementally. The measured classifier inference time and sub-1-
MiB feature storage indicate potential for offline edge implementation. Traceability is
another practical advantage: an operator can inspect the nearest stored examples
supporting a prediction and identify problematic reference samples when an error
occurs. Nevertheless, these benefits remain deployment -oriented rather than
deployment-proven because the reported timing excludes image preprocessing and
feature extraction and no smartphone or embedded board was tested directly [15], [16].

The findings of this study in relation to several previous studies in the same domain are
summarized in Table 7. It should be emphasized that this comparison is indicative rather
than directly equivalent, because each study used different datasets, numbers of classes,
and evaluation protocols.

Table 7. Position of this study relative to previous research
Researcher Features Used Algorithms Best Accuracy
Meiriyama et al. [5]
HOG and LBP (shape,
texture)
K-NN 92.67%

Vol. 8, No. 5, October 2026

Avip Kurniawan, Agustinus Eko Setiawan, et al | 6624
Researcher Features Used Algorithms Best Accuracy
Noviana and
Nugraha [4]
Color, GLCM, shape
Logistic Regression,
Decision Tree
78.31%
Naeem et al. [10]
Multispectral and
texture (14 selected
features)
Multi-Layer
Perceptron
99.01%
Basri et al. [12]
Fourier Descriptor
(shape)
K-NN, CNN 95.00%
Zahirah et al. [13] GLCM (texture) K-NN, CNN 96.00%
This study
Shape, HSV color, and
GLCM texture (21
features)
K-NN with
hyperparameter
tuning
92.18% (image-
level)

The 92.18% image-level accuracy is comparable to K-NN with handcrafted features
reported by Meiriyama et al. [5] (92.67%) but below multispectral laboratory-based results
[10] and CNN-based studies [12], [13], [23]. These values should be interpreted as indicative
rather than directly comparable because datasets, class counts, image acquisition
conditions, and validation protocols differ. The present study instead emphasizes the
integration of shape, HSV, and GLCM features, systematic optimization of k, weighting,
and distance metric, evaluation under three train-test proportions, and class-level error
analysis [24], [25]. The CNN-based studies achieved roughly 95–96% accuracy, a few
percentage points above the present result, but their training and deployme nt
requirements are generally higher. K-NN, by contrast, stores a compact feature database
and makes predictions that can be traced to specific neighboring training samples, which
may be useful for resource-constrained field applications, although direct edge-device
benchmarks are still required. K-NN does not involve iterative parametric model training
after feature construction, but the pipeline as a whole is not free of computational
preparation: feature extraction, standardization, grid search, cross -validation, and
construction of the stored feature database all consume resources, and feature
extraction in particular dominates the per-image cost at inference time. The advantage
is therefore low model complexity and compact storage rather than zero preparation
cost. The practical contribution is therefore a trade-off between predictive performance,

Vol. 8, No. 5, October 2026

6625 | Hyperparameter-Tuned K-Nearest Neighbor for Herbal Leaf Classification …..
computational simplicity, and interpretability rather than a claim of superiority over deep
learning.

Internal validity is limited by augmentation before image-level splitting, which can place
variants of the same source photograph in training and test subsets and make the 92.18%
estimate potentially optimistic relative to source-level validation. This issue is especially
relevant to k = 1 because a highly similar variant can become the decisive nearest
neighbor, and it may also reduce apparent variance across folds. Therefore, the reported
stability across 80:20, 70:30, and 60:40 scenarios should be interpreted as stability under
the present image-level protocol rather than proof of source -independent
generalization. Construct validity is constrained by illumination sensitivity: HSV Value
changes with brightness, Saturation can vary with lighting conditions, and grayscale
GLCM can respond to shadows and specular highlights. The current color representation
also uses only channel means, so two leaves with similar average color but different
within-leaf distributions can appear artificially close in feature space. Severe backlighting,
hard midday reflections, deep shade, and mixed artificial illumination were not
represented comprehensively and may degrade color and texture descriptors more
strongly than the geometric shape descriptors. External validity is limited to five species
and relatively controlled backgrounds, while conclusion validity is limited by the absence
of same-dataset statistical comparison with competing classifiers. Finally, deployment
validity is limited because inference timing was measured on an AMD Ryzen 5 CPU rather
than on a smartphone or embedded board, and the reported classifier time excludes
image preprocessing and feature extraction. Future work should therefore use grouped
source-image splitting, stronger illumination tests, higher-order color moments or
histograms, explicit venation descriptors, statistical comparisons, and direct mobile or
embedded-device evaluation.

4. CONCLUSION

This study developed a lightweight K-NN classifier for five herbal leaf species using a 21-
dimensional fusion of shape, HSV color, and GLCM texture features. Grid Search with
Stratified 5-Fold Cross-Validation selected Manhattan distance, k = 1, and distanc e
weighting for the 80:20 scenario, yielding 92.18% accuracy and 92.29% macro F1 as an
image-level estimate, while the 70:30 and 60:40 splits achieved 91.79% and 91.21%

Vol. 8, No. 5, October 2026

Avip Kurniawan, Agustinus Eko Setiawan, et al | 6626
accuracy under the same protocol. These figures are reported as image-level results
throughout and are not claimed as source-independent performance. Prediction required
about 0.19 ms per image on the tested CPU. Error analysis showed that 65.2% of
misclassifications occurred between soursop and guava, indicating that richer color-
distribution and venation descriptors are the most relevant next improvements. Because
augmentation preceded image-level splitting, the reported accuracy may be optimistic
relative to source-level grouped validation and should not be interpreted as
independent-source generalization; the selection of k = 1 is the component most exposed
to this effect, and the proportion of soursop-guava errors may likewise change once
augmented siblings no longer cross the partition boundary. Source-level grouped
validation was not conducted in this study; therefore, independent-source performance
remains unverified.

ACKNOWLEDGMENT

The authors would like to express their gratitude to the Informatics Engineering Study
Program, Faculty of Technology, Universitas Krisnadwipayana, and the Informatics
Engineering Study Program, Faculty of Technology and Business, Universitas Aisyah
Pringsewu, and Pharmacy Study Program, Faculty of Health and Science, Universitas PGRI
Madiun, Madiun for the support, facilities, and contributions provided throughout the
implementation and completion of this study.

REFERENCES

[1] A. Riska, Purnawansyah, H. Darwis, and W. Astuti, "Studi perbandingan kombinasi GMI,
HSV, KNN, dan CNN pada klasifikasi daun herbal," The Indonesian Journal of
Computer Science, vol. 12, no. 3, pp. 1201–1215, 2023, doi: 10.33022/ijcs.v12i3.3210.
[2] S. Ramadhani, J. Iskandar, and T. Husodo, "Studi etnobotani pemanfaatan tumbuhan
obat di Desa Cintakarya, Kabupaten Pangandaran, Jawa Barat," in Prosiding Seminar
Nasional Masyarakat Biodiversitas Indonesia, vol. 6, no. 1, pp. 518–524, 2020, doi:
10.13057/psnmbi/m060107.
[3] S. F. Fatonah et al., "Pemanfaatan tanaman pekarangan untuk pengobatan herbal,"
JPPM (Jurnal Pengabdian dan Pemberdayaan Masyarakat), vol. 4, no. 2, pp. 247–256,
2020, doi: 10.30595/jppm.v4i2.6030.

Vol. 8, No. 5, October 2026

6627 | Hyperparameter-Tuned K-Nearest Neighbor for Herbal Leaf Classification …..
[4] L. P. R. Noviana and I. N. B. S. Nugraha, "Perbandingan klasifikasi citra daun herbal
menggunakan metode logistic regression dan decision tree classifier berdasarkan
fitur (warna, GLCM, bentuk)," JITU: Journal Informatic Technology and
Communication, vol. 7, no. 2, pp. 126–133, 2023, doi: 10.36596/jitu.v7i2.1241.
[5] M. Meiriyama, S. Devella, and S. M. Adelfi, "Klasifikasi daun herbal berdasarkan fitur
bentuk dan tekstur menggunakan KNN," JATISI (Jurnal Teknik Informatika dan
Sistem Informasi), vol. 9, no. 3, pp. 2573–2584, 2022, doi: 10.35957/jatisi.v9i3.2974.
[6] K. Pushpanathan, M. Hanafi, S. Mashohor, and W. F. Fazlil Ilahi, "Machine learning in
medicinal plants recognition: A review," Artificial Intelligence Review, vol. 54, no. 1,
pp. 305–327, 2021, doi: 10.1007/s10462-020-09847-0.
[7] A. Z. Alfarizi and E. I. Sela, "Klasifikasi rimpang menggunakan metode k -nearest
neighbor dan ekstraksi ciri gray level co-occurrence matrix," Jurnal Fasilkom, vol. 14,
no. 1, pp. 88–94, 2024, doi: 10.37859/jf.v14i1.6832.
[8] F. M. Fathoni, C. A. Putra, and A. L. Nurlaili, "Klasifikasi penyakit daun anggur
menggunakan metode k-nearest neighbor berdasarkan gray level co-occurrence
matrix," Biner: Jurnal Ilmiah Informatika dan Komputer, vol. 3, no. 1, pp. 8–15, 2024,
doi: 10.32699/biner.v3i1.6332.
[9] M. Vakili, M. Ghamsari, and M. Rezaei, "Performance analysis and comparison of
machine and deep learning algorithms for IoT data classification," arXiv,
arXiv:2001.09636, 2020, doi: 10.48550/arXiv.2001.09636.
[10] S. Naeem et al., "The classification of medicinal plant leaves based on multispectral
and texture feature using machine learning approach," Agronomy, vol. 11, no. 2, Art.
no. 263, 2021, doi: 10.3390/agronomy11020263.
[11] S. Patil and M. Sasikala, "Segmentation and identification of medicinal plant through
weighted KNN," Multimedia Tools and Applications, vol. 82, no. 2, pp. 2805–2819, 2023,
doi: 10.1007/s11042-022-13201-7.
[12] H. Basri, Purnawansyah, H. Darwis, and F. Umar, "Klasifikasi daun herbal
menggunakan k-nearest neighbor dan convolutional neural network dengan
ekstraksi Fourier descriptor," Jurnal Teknologi dan Manajemen Informatika, vol. 9,
no. 2, pp. 79–90, 2023, doi: 10.26905/jtmi.v9i2.10350.
[13] D. Zahirah, Purnawansyah, N. Kurniati, and H. Darwis, "Digital image classification of
herbal leaves using KNN and CNN with GLCM features," Jurnal Teknik Informatika
(JUTIF), vol. 5, no. 1, pp. 61–67, 2024, doi: 10.52436/1.jutif.2024.5.1.1162.

Vol. 8, No. 5, October 2026

Avip Kurniawan, Agustinus Eko Setiawan, et al | 6628
[14] H. A. Abu Alfeilat et al., "Effects of distance measure choice on k-nearest neighbor
classifier performance: A review," Big Data, vol. 7, no. 4, pp. 221–248, 2019, doi:
10.1089/big.2018.0175.
[15] M. Pintus, F. Colucci, and F. Maggio, "Emerging developments in real-time edge AIoT
for agricultural image classification," IoT, vol. 6, no. 1, Art. no. 13, 2025, doi:
10.3390/iot6010013.
[16] A. T. Khan, S. M. Jensen, A. R. Khan, and S. Li, "Plant disease detection model for edge
computing devices," Frontiers in Plant Science, vol. 14, Art. no. 1308528, 2023, doi:
10.3389/fpls.2023.1308528.
[17] F. Azzahra and A. D. Astuti, "Comparative study of the use of natural ingredient in
ethnomedical study in Indonesia on the herbal formulary of the RI Ministry of
Health," Stannum: Jurnal Sains dan Terapan Kimia, vol. 5, no. 2, pp. 58–74, 2023, doi:
10.33019/jstk.v5i2.3895.
[18] C. Schröer, F. Kruse, and J. M. Gómez, "A systematic literature review on applying
CRISP-DM process model," Procedia Computer Science, vol. 181, pp. 526–534, 2021,
doi: 10.1016/j.procs.2021.01.199.
[19] W. Vallejo, C. Diaz-Uribe, and C. Fajardo, "Google Colab and virtual simulations:
Practical e-learning tools to support the teaching of thermodynamics and to
introduce coding to students," ACS Omega, vol. 7, no. 8, pp. 7421–7429, 2022, doi:
10.1021/acsomega.2c00362.
[20] D. M. Belete and M. D. Huchaiah, "Grid search in hyperparameter optimization of
machine learning models for prediction of HIV/AIDS test results," International
Journal of Computers and Applications, vol. 44, no. 9, pp. 875–886, 2022, doi:
10.1080/1206212X.2021.1974663.
[21] P. R. Sihombing and I. F. Yuliati, "Penerapan metode machine learning dalam
klasifikasi risiko kejadian berat badan lahir rendah di Indonesia," MATRIK: Jurnal
Manajemen, Teknik Informatika dan Rekayasa Komputer, vol. 20, no. 2, pp. 417–426,
2021, doi: 10.30812/matrik.v20i2.1174.
[22] S. Latif, F. F. Dola, M. M. Afsar, I. J. Esha, and D. Nandi, "Investigation of machine
learning algorithms for network intrusion detection," International Journal of
Information Engineering and Electronic Business, vol. 14, no. 2, pp. 1–22, 2022, doi:
10.5815/ijieeb.2022.02.01.

Vol. 8, No. 5, October 2026

6629 | Hyperparameter-Tuned K-Nearest Neighbor for Herbal Leaf Classification …..
[23] M. H. Ahmad, F. M. Hana, T. G. Pratama, and H. Aulida, "Klasifikasi empat jenis daun
herbal menggunakan metode convolutional neural network," Jurnal Ilmu Komputer
dan Matematika, vol. 4, no. 2, pp. 69–76, 2023.
[24] M. F. Kurnia, S. S. Solihat, G. Windarsih, and D. Usmadi, "Identifikasi otomatis lima jenis
resak (Vatica spp.) berdasarkan beberapa karakter morfologi daun dan algoritma
pembelajaran mesin," Buletin Kebun Raya, vol. 26, no. 1, pp. 26–37, 2023, doi:
10.55981/bkr.2023.740.
[25] R. Muliono, J. H. Lubis, and N. Khairina, "Analysis k-nearest neighbor algorithm for
improving prediction student graduation time," SinkrOn, vol. 4, no. 2, pp. 42–46, 2020,
doi: 10.33395/sinkron.v4i2.10480.