Introduction

Tea is one of the most widely consumed non-alcoholic beverages in the world. And in the context of Sri Lanka, it plays a pivotal role in the country’s economy. Tea is one of the main foreign exchange earners of the country. However, tea production is often affected by various pests and diseases, which can cause significant losses in production and quality .

From them, tea leaf diseases are more likely to cause significant damage as it is the harvesting part of the tea plant. Among the leaf diseases of tea Algal leaf spot, Black blight, Blister blight, Gray blight, and Spider mite attack are the most common types of diseases in Sri Lanka . The traditional methods of tea pest and disease detection rely on manual observation and identification by experts or farmers, based on the symptoms and signs on the tea leaves. However, these methods are time-consuming, labor-intensive, and not reliable. Moreover, they require a lot of domain knowledge and experience in order to early detection and prevention of diseases.

Therefore, there is a need to develop a smart technique for tea diseases detection. Using computer vision and ML based system can be more beneficial in this scenario as the end user has not a lot of technicality to hassle with. Computer vision is the field of study that enables computers to process and understand visual information, such as images and videos. ML is the field of study that enables computers to learn from data and perform tasks without explicit programming. By combining computer vision and machine learning, it is possible to create a system that can automatically and accurately identify and classify various diseases on tea leaves from images.

Background

As of now for plant disease detection usage ML and computer vision have become a major trend in agriculture. This is called Precision agriculture also known as smart farming which has emerged as an innovative tool to address current challenges in agricultural sustainability. This is due to the robust and effective workflow that it provides to the user compared with other disease identification and early detection methods. In developed countries, these methods are used at the commercial level.

ML techniques are used in five main categories in agriculture. They are Disease detection, Weed detection, Crop recognition, Crop quality and Yield prediction. Other than these cases there are many usages of ML in livestock agriculture also. From these categories, Disease detection is crucial as they are most unpredictable and not easy to mitigate. In tomato farming , Rice farming , Potato farming , and apple farming these techniques are heavily used. With usage and the importance of the crop that we are maintaining the importance of building new technologies becomes higher.

When it comes to Sri Lanka the most common crops are tea, rubber, rice, and coconut . These crops have the most economic value compared to other crops as they play a crucial role in the export market. From them, tea is a plant that is used to make tea drinks which is a world-popular non-alcoholic drink and it provides a large sum of export income to Sri Lanka. Hence, the importance of protecting and maintaining the tea crops is crucial.

Currently, there is no working end-to-end system developed in Sri Lanka for identifying the diseases of Tea crops. The detection of diseases is crucial as currently, diseases like Blister blight, Algal leaf spot, Black blight, and etc do huge damage to the tea crops in the country. If there is a system that can detect and analyze diseases it is going to be hugely beneficial for the tea industry in Sri Lanka.

Problem Statement

In the Sri Lankan context, the current tea leaf disease identification relies on subjective human observation which is not unreliable. Even the experts, may not identify the diseases accurately through a manual intervention. In a large-scale context, some diseases need to be identified through laboratory tests by taking samples from the plantation. It is not an efficient way since it consumes a lot of time and also a considerable cost. Even after identifying a plant as a diseased plant, pesticide application may not be efficient in cases where the disease has been developed to an extent such that it can not be recoverable through pesticides. Therefore, early detection of diseases is a crucial aspect of improving crop health and economic benefits. And also there is no efficient way for the tea cultivators to get an idea about to which area the disease has been spread.

This approach is to create a novel strategy that makes use of computer vision and ML to identify the tea leaf diseases accurately through a user-friendly and easily approachable information system for the end users and take prompt actions for the pesticide strategies etc. The integration of disease distribution profiles via GIS locations will be a comprehensive solution to obtain insightful information about disease distribution and the necessary steps that need to be taken to mitigate the issues.

Objectives and Scope

Development of an accurate ML model for tea leaf diseases detection and isolation

This project aims to develop a robust ML model capable of accurately classifying tea leaves as diseased or healthy. Furthermore, the system will predict specific diseases in cases of diseased leaves, leveraging datasets containing prevalent tea leaf diseases in the Sri Lankan context.The most suitable machine learning algorithm will be chosen by comparing the accuracy of a set of algorithms and the model will be trained on both existing and novel dataset. It should be noted that the system will focus on predicting the most common leaf diseases exclusively.

Explore the possibility of design the system to detect tea leaf diseases at their early stages of disease development.

This project aims to detect diseases at the early stages of the disease development cycle to enhance the disease detection capabilities beyond what is achievable through manual intervention alone, thereby facilitating proactive measures to mitigate disease spread and minimize crop damage.

Development of a real-time user-friendly information system (mobile device-centered) for potential stakeholders.

This study aims to design and implement a user-friendly mobile application accessible on both iOS and Android platforms, utilizing cross-platform frameworks. The application will incorporate features for image input and real-time disease detection, leveraging cloud infrastructure for seamless processing. The primary goal is to provide a technically simple yet effective tool for end users, such as tea cultivators, and industries, to easily access and utilize for tea leaf disease detection and management.

Explore the possibility of generating the disease distribution profile

This project will explore the possibility of generating a disease distribution profile of the plantation, providing insightful data about diseased plants along with their Geographic Information System (GIS) locations and processed timestamps to make useful decisions for end users.

Methodology

Data Acquisition

In the first phase of the project, a diverse dataset of tea plant leaf images will be acquired that are affected by various diseases. This data acquisition should be done comprehensively to ensure the suitability for the training and testing of the ML models. This novel data set is collected from the tea plantations in Sri Lanka in different regions through field surveys. When collecting this dataset it is necessary to ensure the variance in the images of healthy leaves and also the leaves affected by various diseases in the Sri Lankan context. Environmental conditions were also taken into consideration when acquiring the novel dataset. Based on the resources a high-resolution RGB camera or a NIR camera will be used to gather the data about the extreme ends of the diseases in order to explore the possibility of early detection of diseases. With exception to this novel dataset which is collected through the field surveys, existing datasets will be utilized to train the model from reliable data repositories.

Image Processing

Before the model is trained, the images will undergo several preprocessing methods to ensure the consistency of the dataset. It includes the standardization of image size, resolution, and the quality to reduce the effects of varying scales and distribution of pixel values in images. This will make it easier for machine learning models to learn patterns effectively. Different filtering techniques may have to be utilized to reduce unwanted noise and preserve important features of the images. In addition to that different color correction techniques will be used to reduce the effects of different lighting conditions and camera settings when acquiring the images. After these preprocessing steps, the image dataset will be split into training, validation, and testing datasets in order to facilitate model training and evaluation.

Feature Extraction

Feature extraction involves the process of extracting the relevant information from the raw leaf images. Based on the literature an educated guess is made that different diseases have different leaf characteristics that can be analyzed using image processing for disease detection. In feature extraction important information about color characteristics, shape descriptors, leaf patterns, and other indicators of diseases will be extracted.Since different diseases have unique surface characteristics that can be identified by texture-based features, it is an important factor that can be utilized to classify images with their disease types.

Machine Learning model selection

Several ML algorithms will be compared based on validation criteria like precision, accuracy, recall, and F1 score to determine which algorithm is best suited for tea plant leaf disease detection. The mobile application will incorporate the algorithm that exhibits the best performance in terms of these criteria. In this project, many ML-based algorithms, such as SVM, Random Forest, ResNet, LeNet and YOLOv8, will be assessed in this research. Since the acquired dataset is limited transfer learning methods will be applied to fine-tune the model. Best suited model is selected to integrate with the information system.

Development of the Mobile Application

The project entails the design and development of a scalable and user-friendly mobile application compatible with both iOS and Android platforms. Leveraging the Flutter framework, the application will offer seamless cross-platform functionality, ensuring widespread accessibility. The selected machine learning algorithm for real-time disease detection will be seamlessly integrated into the mobile application. The mobile application’s user interface (UI) is designed to be intuitive and user-friendly, enabling end-users to efficiently manage their tasks. Through this interface, users can swiftly determine the health status of suspected plant leaves and identify any associated diseases within a short period. To expedite processing time, the application leverages necessary cloud computing concepts. These techniques optimize processing time, delivering reliable and precise results to users in real time. For large-scale plantations where capturing images of each plant individually is impractical, users are provided with the capability to capture multiple images of the plantation at various locations. These images are geotagged to record their precise location within the plantation. Leveraging this geotagged data, the application generates insightful data visualizations that depict the spread of diseases across the plantation. By analyzing the geotagged images, the application constructs a distribution profile that highlights areas affected by diseases, enabling users to identify disease hotspots and prioritize intervention efforts accordingly.