Detection of coffee leaf disease and identification using deep learning

Motivation

• (1)

  To provide a new method that can handle a huge amount of datasets, as well as low image dimensions, whereas normally a high image to provide a perfect outcome.

• (2)

  To propose a highly effective method for classifying coffee leaf disease that has the ability to classify with less complexity than the traditional methods.

• (3)

  To support farmers, provide an easy-to-use and low-cost system that can be applied to mobile devices to enhance productivity and reduce losses resulting from diseases through early diagnosis.

• (4)

  The proposed model is based on the urgent need for modern technologies in smart agriculture to improve coffee productivity and address disease challenges in a technologically advanced manner, contributing to improving the global agricultural economy. It is also a comprehensive and integrated solution to detect the most dangerous diseases affecting coffee plants accurately and quickly. Without the need for experts in this field to enhance productivity and reduce losses resulting from diseases through early diagnosis

Data preparation and preprocessing

JMuBEN and JMuBEN2 (Arabica dataset): The dataset was collected by disease detection experts and supervised by the dataset is supervised by three institutions: Chuka University, the University of Embu, and Jomo Kenyatta University of Agriculture and Technology. The first group contains three disease categories: Cerscospora, Leaf Rust, and Phoma. The second group contains two types of diseases: Miner and Healthy. The two volumes contain (58,555) images, with a size was 128 × 128 pixels for five classes (Phoma, Cercospora, Rust, Healthy, Miner). The dataset used for training was obtained from publicly available online sources. The source is open access and widely used in research studies related to the detection of coffee disease (Jepkoech et al., 2021).

Dataset collection is an important part of the building blocks of a system that detects plant diseases, especially coffee. Experts in the field determine the type of disease, collect data, and the challenges associated with collecting data are many, such as weather, accuracy, and other important things (Yebasse et al., 2021). Coffee disease experts classified each image accurately; the analysis focuses on the identification and quantification of various coffee leaf diseases. The dataset comprises images of coffee leaves, categorized into different disease types. The primary objective is to classify and count the number of leaves affected by each disease type. The data underscores the importance of monitoring and addressing coffee leaf diseases to ensure crop health and productivity. The significant number of healthy leaves is promising, yet the presence of diseases like Miner and Leaf Rust highlights areas requiring attention and potential intervention. The test dataset is then used to show how well the recommended architecture can summarize the new information or to show how well the model performs under real conditions. Datasets were obtained from a plantation of Arabica coffee with the help of a camera and a plant pathologist. The images were then edited through cropping focusing on parts of interest, along with samples and numbers for each of them. Image augmentation was done to increase the dataset size and prevent overfitting challenges during the training and validation of the model. Table 1 shows the types of diseases and a description of each disease, along with samples and numbers for each disease.

In addition, we selected another dataset to test the proposed model and generalize it. The dataset is open source data from the Kaggle platform called coffee leaf disease. It containing 1,700 images, divided for training and testing. The dataset was specifically designed for detecting coffee diseases, consisting of three channels with a resolution of 2,048 × 1,024 pixels. It includes four types of diseases (Miner, No Disease, Phoma, and Rust). This dataset focuses on the problem of detecting healthy and diseased coffee leaves (Murhaban & Suhendra, 2023).

Figure 2 shows the image illustrates three basic steps in data processing used for coffee leaf classification. Original images: The image selected from the dataset contains significant variations in lighting, color, and the number of images per disease class, demonstrating an imbalance in the number of samples across classes. After balancing: Data balancing techniques, such as oversampling for underrepresented classes, were applied to achieve an even distribution of images across all classes. This step is important in reducing bias in the proposed model and improving its accuracy. After normalization: The images then undergo a normalization process, where pixel values are converted to a uniform numerical range, such as 0 to 1. This helps the model learn better from the images and reduces the impact of lighting variations or color contrast.

To address the class imbalance, we balanced the dataset, resulting in an equal number of samples for each class. We used algorithm random undersampling balance data undersampling is a technique used to balance an imbalanced dataset by reducing the number of samples in the majority class. This method helps ensure that the model does not become biased toward the majority class during training. Each class now has around 6,572 samples. This balancing ensures that the machine-learning model receives an equal representation of each class, thereby improving its ability to generalize across different disease types. The main steps for undersampling are to: identify the majority and minority classes, randomly remove samples from the majority class, and combine the dataset. These images demonstrate a variety of diseases, highlighting the diversity and challenges in classification balancing the dataset to address the class imbalance each class was limited to a minimum count of 6,572 images. This approach ensured equal representation of each class, improving the model’s fairness and performance. The images were normalized to have pixel values between 0 and 1 by dividing by 255.0. This normalization step is crucial for ensuring consistent input ranges for the machine-learning model. Labels were converted to one-hot encoding using label binaries, transforming categorical labels into a binary matrix format; this is essential for multi-class classification tasks. The preprocessed dataset was split into training and testing sets using an 80–20 split: training data shape: (26,284 samples); testing data shape: (6,571 samples). Figure 3 represents the number of samples in each class before and after balancing. The classes are evenly distributed, which should help improve the performance and accuracy of our mode.

To enhance the model’s generalization ability and increase its robustness, a set of extension techniques was used to improve image diversity and reduce the likelihood of overgeneralization. These techniques included flipping the images horizontally and vertically, rotating them using affine transformations at an angle of 20 degrees, adjusting the brightness by 1.2, and applying a Gaussian blur with a standard deviation (σ) of 1.0, Table 2 shows all the lists the mathematical operations for the augmentation techniques. These operations aim to simulate the variability of real data and improve the model’s generalization ability. These manipulations aim to simulate realistic changes in leaf appearance, improving the model’s ability to detect diseases under different environmental conditions (Maharana, Mondal & Nemade, 2022). Figure 4 shows augmentation for five categories of diseases with the approaches utilized.

Feature extraction

Local binary patterns (LBP): Feature extraction is a crucial step in many machines learning and image processing tasks, including your coffee disease detection project. It involves identifying and extracting important and relevant features from raw data to improve the performance of models from color features, texture features, LBP and others. CNNs automatically learn complex features directly from the image data, often outperforming traditional handcrafted features. In addition to preprocessing, feature extraction was performed to enhance the analysis. Grayscale image and LBP. The grayscale version of the image is created to reduce computational complexity and highlight texture details. LBP applied grayscale images to capture texture features. LBP is a powerful feature descriptor that provides the spatial structure of local image texture. The histogram of LBP values computed to quantify the texture information. This histogram is used as a feature vector for classification. Figure 5 shows an example of the application of the LBP technique.

Extended center-symmetric local binary pattern (XCS-LBP): XCS-LBP applied further to capture more detailed texture information. This technique helps in highlighting finer features of the leaf surface, which are crucial for disease identification LBP used to ex-tract texture features from an image, each pixel’s neighborhood encoded into an LBP histogram, which represents the texture information classification. The XCS algorithm applied to these LBP histograms to classify the images or patterns. The XCS system evolves classifiers based on the features extracted by LBP, improving its classification accuracy over iterations, XCS-LBP leverages LBP for extracting texture features and combines it with the XCS algorithm for effective classification, making it a powerful tool for pattern recognition and image analysis and all these processes are shown in Fig. 6.

CNN models

CNN architectures specifically designed for image analysis and object detection (Khushi et al., 2024). These architectures often consist of multiple convolutional layers followed by pooling layers to extract features from images. The final layers usually consist of fully connected layers for classification (Biswas, Saha & Deb, 2024). Activation functions are a critical component of CNNs and used in the introduction of nonlinearity into each layer is output. Most real-world data is complex and cannot be modeled well with a linear function. Non-linearity is therefore important (Perumal et al., 2024). There are several activation functions used in CNNs, but some of the most popular ones are rectified linear unit (ReLU) (Rastogi, Dua & Dagar, 2024). This function is simple and computationally efficient in setting all values that are negative to zero leaving all values that are positive unchanged.

The CNN1 architecture was built using a sequential stack of layers. The CNN designed in this model consists of a set of concatenated layers that aim to extract features from input images (100 × 100 × 3), and classify them into five classes. We divided the data into 20% test and 80% training. The input layer receives color images of size 100 × 100 × 3. The first convolutional layer (Conv2D) contains 32 filters of size 3 × 3 with a ReLU activation function. The first pooling layer (MaxPooling2D) reduces the dimensionality to 49 × 49 × 32 using 2 × 2 pooling. The second convolutional layer (Conv2D) applies another 32 filters of size 3 × 3 with ReLU. This is followed by the second pooling layer (MaxPooling2D), which reduces the size to 23 × 23 × 32. The dropout layer then drops 50% of the units to reduce the possibility of overfitting. The flatten layer transforms the output from 3D to a vector of size 23 × 23 × 32. The fully connected layer (Dense) contains 256 units with ReLU. The second dropout layer is responsible for aggregating the discovered features, and it drops 70% of the units during training to reduce overfitting. It has no parameters. The output layer (Output Dense) has five neurons (one for each class) and uses SoftMax to give the probability of each class, epochs = 5, with Adam being one of the most widely used optimizers. It combines the advantages of SGD and RMSProp. It automatically adapts the learning rates for each weight.

The second model CNN2 is built on a deep Conv2D neural network architecture, consisting of three consecutive Conv2D layers, each followed by a MaxPooling 2D pooling layer, with two batch normalization layers added to improve training stability. The model starts with an input layer of size 100 × 100 × 3, followed by a first convolutional layer containing 25 5 × 5 filters with “same” padding. This is followed by a second convolutional layer after pooling, containing 50 5 × 5 filters. The first batch normalization layer is applied to 50 channels. The third convolutional layer contains 70 3 × 3 filters. Then, after a series of convolutional layers, a second batch normalization layer is applied. The output is transformed into a vector using a flatten layer. The data is then passed through two fully connected (Dense) layers. A dropout of 0.5 is applied to reduce overlearning. Finally, the output layer contains five classification units with a SoftMax function, a batch size of 32, epoch of 5, and Adam optimizer.

The VGG16 architecture is pre-trained on ImageNet but excludes its top fully connected layers. The VGG16 model is loaded with weights and configured to accept input images of size 100 × 100 with three color channels. After loading the pre-trained VGG16 model, custom layers were added on top. The output from the VGG16 base model is flattened to convert the 2D feature maps into a 1D vector. Following this, three dense layers were added, each with ReLU activation. The first dense layer has 46 neurons, the second has 40 neurons, and the third has 10 neurons. The final layer is a dense layer with a number of neurons equal to the number of classes in the target data, using soft-max activation to produce class probabilities. The complete model is then created by specifying the VGG16 input and the custom output layers. The model was compiled using the Adam optimizer, categorical cross-entropy as the loss function, and accuracy as the performance metric. The total number of layers in the model is 23 layers. The optimizer used in the model is Adam.

Hybrid Learning: In addition to designing three basic models using convolutional neural networks (CNN1, CNN2, and VGG16), a set of hybrid models was developed to combine the power of deep learning in visual feature extraction with the effectiveness of traditional machine learning algorithms in classification. The final classification layer (Softmax) was removed from CNN1, CNN2, and VGG16, and their final hidden layers were used to extract deep feature representations of images. These features were then passed as input to two traditional classifiers: support vector machine (SVM) and Random Forest (RF). This process resulted in six hybrid models: CNN1 + SVM, CNN1 + RF, CNN2 + SVM, CNN2 + RF, VGG16 + SVM, and VGG16 + RF. The main idea behind this approach is to evaluate the extent to which traditional classification algorithms benefit from the deep features extracted by convolutional networks and compare their performance with direct classification results via the Softmax layer.

In addition to the previous models, two transfer learning algorithms were included: ResNet50 and EfficientNetB0, which are deep models pre-trained on the ImageNet dataset. These models were reused without modification to their basic architecture, except for replacing the final layer with a new one corresponding to the number of classes in the coffee leaf data. The goal is to evaluate the actual performance of these two models when used as baseline transfer learning, and to compare their results with those of existing and hybrid models. Neither ResNet50 nor EfficientNetB0 used traditional SVM or RF classification algorithms, relying solely on direct prediction via the final Softmax layer.

In the process of disease detection in plants, the classification step using image processing and computer vision is very important (Upadhyay et al., 2025). The proposed system was designed based on the process of detecting and extracting features in the affected leaf. The model’s parameters are modified so that it correctly identifies and classifies images, and then the model is evaluated. After the training is completed, the model’s performance is evaluated using the test dataset (Ding et al., 2024). The accuracy of the model and its ability to recognize diseases in health are measured. After verifying the model’s performance, it can be used to classify new images of coffee plants and identify various diseases based on the patterns it learns (Shrma & Rahiman, 2024). We implemented the proposed method using Python, where the application was via the cloud using Jupiter Notebook, where libraries such as Pandas, NumPy, Matplotlib, and Seaborn were called to work with data consisting of images. In addition, the data was requested from the Kaggle website.

Hardware specifications used during experiments

The proposed model was implemented on the Kaggle platform, which provides a suitable cloud training environment for efficient model development and testing. The proposed CNN1 model was designed to be lightweight in terms of computational complexity, making it suitable for implementation on low-cost devices. Hardware specifications used during experiments: (1) Platform: Kaggle (Notebook Environment). (2) graphics processor (GPU): NVIDIA Tesla P100 (16 GB VRAM). (3) Memory (RAM): 13 GB). (4) Processor: Intel Xeon (approximately 2.3 GHz). (5) Software Environment: Python 3.10—TensorFlow 2.x.6). In this section, we provide a detailed explanation of the metrics used to evaluate the performance of the models. Moreover, the performance of the proposed models is also given and compared to other models published in the literature. Our experimentations focused on the performance and the detection classification.

To conduct a comprehensive analysis of the performance of the proposed models, confusion matrices were created for all proposed algorithms and presented in a unified table. This representation allows for comparison of the behavior of different models in classifying each category, and for identifying categories that exhibited confusion or misclassification across multiple models. The main diagonal displays the number of correct predictions for each category, while the other values represent the number of errors between categories. This detailed analysis helps explain differences in model performance, especially when overall accuracy values are close. Table S2 shows the confusion matrix for all algorithms.

To analyze the generalization ability of the proposed model, a new dataset consisting of approximately 1,700 images was tested using the highest-accuracy model (CNN1). The confusion matrix in Table 3 shows the classification results on this dataset, with an overall accuracy of 98%. Class balance was maintained by randomly adding a fifth class from the training data, which helped maintain the consistency of the distribution.

This section, we provide detailed experimental results presented for the proposed models, which include convolutional neural networks (CNN1, CNN2, and VGG16), as well as hybrid models (CNNs with SVM and Random Forest), and transfer learning models (ResNet50 and EfficientNetB0). The proposed approach demonstrated that all models performed well, with accuracy ranging between 93% and 99% on the coffee leaf test dataset. CNN1 achieved the highest accuracy among the models, demonstrating its ability to efficiently extract distinctive visual features. Table S3 shows the results of each model for each class, with the models evaluated using four basic metrics: Accuracy, Recall, Precision, F1-score, and the average for each metric.

To demonstrate the comprehensiveness of the proposed approach, the model was tested on a second external dataset of 1,700 images distributed across only four disease classes, to evaluate its ability to generalize in a new data environment. To achieve class balance in the test set, images for the fifth class (Cercospora) were randomly called from the training data, so that the number of samples in each class was approximately equal. The results showed that the model maintained high performance, with an overall accuracy of 98% on this dataset, demonstrating the model’s effectiveness in handling previously unseen data and its ability to reliably distinguish between different types of diseases. Table 4 shows the results we obtained by testing the second data set. The results show very high accuracy, which proes that the proposed approach is reliable and effective.

Performance assessment

Developing an adequate system for the detection of coffee leaf diseases requires the accurate classification of different plant abnormalities. To achieve this, we performed a series of experiments, which were conducted in two stages. In the first stage, we focused on the classification performance of the proposed CNN models. This performance is also observed, where the proposed CNN models perform very well across all standard metrics, the confusion matrix shows a very high accuracy across all classes, with only With a few errors’ misclassifications out of thousands of samples. The models perform exceptionally well in classifying the different categories of coffee leaf diseases, as evidenced by the near-diagonal matrix, indicating accurate predictions aligning with the true labels.

The receiver operating characteristic (ROC) curve was used to evaluate the performance of the trained models in classifying the five categories of coffee leaf diseases. The ROC curve resembles a graph showing how well a classification model performs. The curve is interpreted to understand how the model makes decisions at different levels of certainty. The curve shows the relationship between the true positive rate and the false positive rate for each category. Multiple ROC curves were calculated for each of the proposed models and for the hybrid models, allowing for comparison of their performance in distinguishing different categories. The values and results in the auv indicate that the approach is effective on the test data, demonstrating a high ability to classify healthy and diseased leaves. Figure 7 shows the curves for each algorithm.

Deep learning is a powerful machine learning technique that you can use to train powerful object detectors (Maruthai et al., 2025). There are several deep learning techniques for object detection, including VGG16 to identify the area affected by diseases and determine its type (Karthik, Alfred & Kennedy, 2023). The performance values for the proposed methods to detect coffee leaf diseases are presented in Table 4. In this case, CNN1 has shown the best option to be implemented in a low-cost device for the detection of coffee leaf diseases.

CNN models have proven effective in classifying coffee diseases with high accuracy and speed detection. The performance of the proposed detection systems, built on the CNN1 and VGG16 architectures, exhibited a notable convergence in their effectiveness. Both architectures consistently delivered high accuracy in detection tasks, demonstrating their robustness and reliability. This convergence highlights the complementary strengths of the two algorithms, as they each successfully captured and processed intricate features within the dataset, leading to precise and consistent detection outcomes. Finally, a comparative study with other state-of-the-art methods was made, whose results are As previously mentioned in the introduction, there are many other methods published in the literature, but most of these studies were made using a relatively small size of datasets, and therefore those results are not very convenient for a real-world application. This more exhaustive study, using a large variety of real images and diseases, gives more reliability to the results achieved by the proposed methods for a real application.

Training and testing time models

Figure 8 shows a comparison of training and testing times for different models. The differences between models is shown in terms of time efficiency. This means that the left axis displays the training time, and the right axis displays the testing time, both using different metrics on the same graph. The left blue axis represents the training time in seconds (s). Each blue bar represents the time it took the model to learn from the data. The right green axis represents the testing time in seconds (s). The green barsrepresent the time it took the model to predict outcomes on new data. Figure 8 depicts the training and testing time performance of the proposed models.

The VGG16 model with Random Forest and CNN achieved the shortest training and testing time, reaching 44 and 2.2 s, making it the most computationally efficient model. In contrast, the EfficientNet model required the longest training time, at 199 s, although its testing time remained within the acceptable range (3.5 s). Hybrid models that combine neural networks (such as CNN and CNN2) with traditional machine learning algorithms (such as SVM and RF) demonstrated a good balance, such as the CNN2 + RF model, which delivered fast performance in training (122 s) and testing (2.5 s). These results demonstrate that hybrid models and convolutional neural networks can be an effective choice when real-time processing speed-dependent applications are required. Table 5 illustrates this.

After training, the model uses a dataset containing images of coffee leaves classified into five classes (Intact, Rust, Phoma, Miner, and Cercospora), the trained model was used to predict which class of new images have never been seen belonging to before, as shown in Fig. 9.

A visual comparison of the model’s interpretation decisions for five images of plant leaves is presented, using different interpretation techniques. Each row represents an interpretation of one image: the original image that was classified. The first technique is gradient-weighted class activation mapping (Grad-CAM), which focuses on the regions that the model focuses on in making its decision. The second technique is Grad-CAM++, which is an improved version of Grad-CAM, which focuses on reviewing features in the image for more than one object. Local interpretable model-agnostic explanations (LIME) is used to show the boundaries that most influence the model’s prediction using local segmentation. The fourth technique is the SHapley Additive exPlanations (SHAP) bar plot, which illustrates each feature, such as color, texture, or a specific pixel location, for prediction. Analysis of the results on the selected images represents the first and second rows, representing clearly affected images. Grad-CAM++ and LIME highlight the affected area, while SHAP highlights important features that contribute to the prediction. This indicates that the model focused on diseased areas, not random ones, indicating good interpretation. Similarly, rows three and five show healthy images. Heat maps are low in concentration, almost empty. SHAP bar plots appear absent or have weak effects. This indicates that there are no strong signs of disease, which is positive. However, row four shows a leaf with a mild effect. Grad-CAM++ and LIME identify a specific area, while SHAP shows variations in the importance of features. Figure 10 shows the model’s interability of the model to concentrate on the region of the diseases.

Limitations

• (1)

  Environmental constraints or different environmental conditions such as rain or shadows may affect the quality of the images and make it difficult to distinguish between disease patterns.

• (2)

  The model may not be able to detect diseases that were not included in the training data. Therefore, there is not enough labeled data available for all possible types of coffee diseases, which limits the model’s ability to recognize all cases.