Feature Fusion of GLCM, LBP, and Color Histogram for Rice Leaf Disease Classification

The proposed system's stages are shown in Figure 1. In general, the system consists of four main steps: image dimension adjustment, feature extraction, combining features, and classification. The image dimension adjustment is performed during preprocessing to ensure all images have uniform dimensions, enabling each extraction algorithm to operate consistently. In the feature extraction stage, three complementary descriptors, GLCM, LBP, and CH, are applied in parallel. GLCM extracts macrotexture patterns, such as large spots or leaf tissue damage. LBP captures microtextures, such as fine dots and surface irregularities. CH maps color variations arising from physiological changes in diseased leaves. Each descriptor generates a feature vector that is then combined in the feature combination stage to form a more comprehensive image representation. This combined representation strengthens the model's ability to distinguish subtle differences between disease classes. In the final stage, the combined features are fed into an RF classifier to determine the disease category for each image sample. Performance evaluation is conducted using accuracy, precision, recall, and F1-score metrics to assess the system's overall effectiveness.

1.jpg

Figure 1. The stages of the proposed system for rice leaf classification

2.1 Image dimension adjustment

Image dimension adjustment changes an image's size. This study utilizes two image datasets. The image in the first dataset measures 1,440 × 1,920 pixels, while the image in the second dataset measures 1,600 × 1,600 pixels. Both datasets are resized to 128 × 128 pixels. During preprocessing, we use the original images and resize them only to preserve the characteristics of the rice leaves.

2.2 Feature extraction process

The descriptors used in this feature extraction include GLCM, LBP, and CL. GLCM detects coarse patterns, such as spots or leaf tissue damage, effectively highlighting them. LBP analyzes leaf surfaces, such as wrinkles and spots. CL assesses color changes associated with rice leaf diseases.

2.2.1 Gray Level Co-occurrence Matrix

GLCM is a widely recognized technique for extracting texture features, developed by Haralick et al. [19]. GLCM calculates the frequency of pairs of pixels with specific gray levels that occur together at a given direction and distance within an image. This descriptor generates a G × G matrix, where each element represents the frequency with which pairs (i, j) appear together. Suppose P(i,j) is the probability of pixels with intensities i and j appearing together, and G is the number of gray levels. In that case, it is possible to derive statistical features [20]. Contrast measures the difference between pixels and is expressed as in Eq. (1). homogeneity assesses how closely the element distribution aligns with the GLCM diagonal and is expressed using Eq. (2). energy is evaluated based on the compactness or uniformity of the matrix and is calculated as shown in Eq. (3). correlation, on the other hand, measures the similarity of patterns between neighboring pixels, as defined by Eq. (4). In these formulas, $\mu_{\mathrm{i}}$ and $\mu_{\mathrm{j}}$ represent the mean values of row index i and column index j in the GLCM, respectively, and $\sigma_i$  and $\sigma_j$ are the standard deviations of the distributions of values i and j.

$Contrast=\sum_{i=1}^G \sum_{j=1}^G(i-j)^2 \cdot P(i, j)$                 (1)

$Homogeneity=\sum_{i=1}^G \sum_{j=1}^G \frac{P(i, j)}{1+|i-j|}$                (2)

$Energy=\sum_{i=1}^G \sum_{j=1}^G[P(i, j)]^2$                  (3)

$Correlation=-\sum_{i=1}^G \sum_{j=1}^G \frac{\left(i-\mu_i\right)\left(j-\mu_i\right) \cdot P(i, j)}{\sigma_i \cdot \sigma_j}$                (4)

2.2.2 Local Binary Pattern

The LBP technique describes image textures by analyzing the neighborhoods of each pixel. In its simplest form, the LBP operator compares each of the eight surrounding pixels to the center pixel and assigns a binary value (0 or 1) based on whether the intensity of the neighbor is greater than or equal to that of the center pixel. These binary values are combined into a single LBP code that captures local texture information. An illustration of the LBP computation process is shown in Figure 2. The LBP feature extraction was performed using the standard configuration with eight sampling points evenly spaced around a circular neighborhood of radius 1 pixel. The rotation parameter was set to upright, preserving the spatial relationship between pixels without rotation invariance. This configuration was chosen due to its balance between computational efficiency and the ability to capture meaningful texture patterns.

2.jpg

Figure 2. Example of LBP calculation

Note: LBP = Local Binary Pattern.

2.2.3 Color Histogram

CH is a statistical method frequently used to illustrate the distribution of colors in digital images. This method works by calculating the frequency of each color value or combination within the image, then organizing these frequencies into a histogram [21]. The color histogram is considered a global feature because it ignores spatial information about pixels. Instead, it focuses on the relative count of pixels for each color.

Generally, when an image is represented in a specific color space, such as RGB or HSV, a histogram is generated for each color channel. For example, in the RGB color space, a color histogram can be calculated by dividing each channel (Red, Green, and Blue) into several bins (color intervals) and then counting the number of pixels in each bin. Suppose a digital image has color channels C $\in${R, G, B}, and each channel is divided into n bins; then, the histogram value H_c(k) for channel c and the k^th bin can be defined as in Eq. (5). I_c(x,y) is the pixel intensity at coordinate (x,y) for channel c, B_k is the range of color values in the k^th bin, d(a, B_k)=1 if a $\in$ B_k, and zero otherwise, and W×H is the image resolution.

$H_c(k)=\sum_{x=1}^W \sum_{y=1}^H \delta\left(I_c(x, y), B_k\right)$            (5)

2.3 Combine features

The result of GLCM feature extraction (f_glcm) can be expressed as in Eq. (6), where f_gi is the i^th GLCM feature and m is the total number of features. The LDB feature extraction (f_ldb) output is given by Eq. (7), where f_li is the i^th LDB feature and n is the total number of LDB features. Meanwhile, the CH feature (f_ch) is given by Eq. (8), where f_ci is the i^th CH feature and p is the total number of features. The integration of the three features is shown to coincide in Eq. (9).

$f_{g l c m}=\left[f_{g_1} f g_2 \ldots f g_m\right]$                (6)

$f_{l d b}=\left[f l_1 f l_2 \ldots f l_n\right]$                (7)

$f_{c h}=\left[fc{_1} f c_2 \ldots f c_p\right]$               (8)

$f_{c o n}=\left[f_{g l c m} f_{l d b} f_{c h}\right]$                (9)

2.4 Classification

The descriptors used in this feature extraction include GLCM, LBP, and CL. GLCM detects coarse patterns, such as spots or leaf tissue damage, effectively highlighting them. LBP analyzes leaf surfaces, such as wrinkles and spots. CL assesses color changes associated with rice leaf diseases. Random Forest (RF) is a highly effective ensemble learning technique for classification tasks. This algorithm operates by constructing multiple decision trees simultaneously and combining their predictions via majority voting to achieve more accurate, stable outcomes [22]. The primary advantage of RF is its capacity to minimize overfitting via bagging and random feature selection at each split.

The classification process starts with creating several bootstrap samples from the original dataset. Each sample is used to train a decision tree. During tree building, at each node, the algorithm considers only a random subset of features to find the best split. This method ensures that each tree in the forest has enough variation, making the combined prediction more reliable. To evaluate the quality of the split, RF usually uses Gini Impurity or Entropy. Gini Impurity measures how often a randomly chosen element will be misclassified, using Eq. (10), where p_k is the proportion of samples in class k. Meanwhile, Entropy measures the uncertainty in the data split, calculated with Eq. (11). The final prediction for a sample is made through majority voting, calculated using Eq. (12), where T trees predict $\hat{y}_1, \hat{y}_2, \ldots, \hat{y}_T$.

$G=1-\sum_{k=1}^K p_k^2$                (10)

$H=-\sum_{k=1}^K p_k \log \left(p_k\right)$                 (11)

$\hat{y}={mode}\left\{\hat{y}_1, \hat{y}_2, \ldots, \hat{y}_T\right\}$                (12)

2.5 Performance evaluation

The evaluation of the proposed method was conducted using accuracy, precision, recall, and F1-score, as defined in Eqs. (13)-(16), respectively [23]. Since the classification task involves multiple classes, the metrics were computed using the one-vs-all (OvA) strategy, in which each class is compared against all others. The results were then aggregated using macro averaging, assigning equal weight to each class.

Classification evaluation metrics are derived from the confusion matrix. In multiclass classification, TP_i (True Positive) is the number of data points correctly identified as class i. TN_i (True Negative) is the number of data points correctly identified as not belonging to class i. FP_i (False Positive) indicates data points incorrectly labeled as class i, when they belong to another class. Lastly, FN_i (False Negative) represents data points that should be classified as class i but are incorrectly assigned to another class.

$ Accuration =\frac{\sum_{i=1}^K T P_i+T N_i}{\sum_{i=1}^K T P_i+T N_i+F P_i+F N_i}$                (13)

$Precision_i=\frac{T P_i}{T P_i+F P_i}$                (14)

$Recall_i=\frac{T P_i}{T P_i+F N_i}$               (15)

$F 1- score _i=2 \times \frac{ { Precision }_i \times { Recall }_i}{ { Precision }_i+ { Recall }_i}$                    (16)