[Paper Review] Deep Learning-Based Computational Model for Disease Identification in Cocoa Pods (Theobroma cacao L.)

Darlyn Buenaño Vera, Byron Oviedo+2 more|arXiv (Cornell University)|Jan 2, 2024

Cocoa and Sweet Potato AgronomyAgricultural and Biological Sciences3 citations

TL;DR

This paper proposes a lightweight deep learning model, EfficientDet-Lite4, for detecting and classifying 'monilia' and 'black pod' diseases in cocoa pods using image analysis. Trained on a curated dataset of 500 annotated images, the model achieves 35% average precision and is integrated into an Android mobile app for real-time, user-friendly disease diagnosis by farmers.

ABSTRACT

The early identification of diseases in cocoa pods is an important task to guarantee the production of high-quality cocoa. The use of artificial intelligence techniques such as machine learning, computer vision and deep learning are promising solutions to help identify and classify diseases in cocoa pods. In this paper we introduce the development and evaluation of a deep learning computational model applied to the identification of diseases in cocoa pods, focusing on "monilia" and "black pod" diseases. An exhaustive review of state-of-the-art of computational models was carried out, based on scientific articles related to the identification of plant diseases using computer vision and deep learning techniques. As a result of the search, EfficientDet-Lite4, an efficient and lightweight model for object detection, was selected. A dataset, including images of both healthy and diseased cocoa pods, has been utilized to train the model to detect and pinpoint disease manifestations with considerable accuracy. Significant enhancements in the model training and evaluation demonstrate the capability of recognizing and classifying diseases through image analysis. Furthermore, the functionalities of the model were integrated into an Android native mobile with an user-friendly interface, allowing to younger or inexperienced farmers a fast and accuracy identification of health status of cocoa pods

Motivation & Objective

• Address the critical challenge of early detection of 'monilia' and 'black pod' diseases in cocoa pods, which cause up to 80% production losses.
• Develop an accessible, low-cost solution for smallholder farmers with limited technical expertise to identify disease symptoms using mobile technology.
• Design and implement a deep learning-based computational model that accurately localizes and classifies disease-affected cocoa pods from images.
• Integrate the trained model into a user-friendly Android application to support real-time diagnosis and treatment recommendations in the field.

Proposed method

• Selected EfficientDet-Lite4 as the primary object detection architecture due to its efficiency, lightweight design, and strong performance in feature extraction and bounding box prediction.
• Prepared a custom dataset of 500 cocoa pod images, including healthy, 'monilia', and 'black pod' infected pods, through image cropping, normalization, and expert validation by the 'Asociación de Productores Agrícolas Divino Niño'.
• Trained the model using Google Colab’s computational resources, optimizing for both disease classification and object localization with a focus on minimizing inference time.
• Implemented a mobile application using native Android development to deploy the trained model, enabling users to capture images, receive probabilistic diagnoses, and access treatment guidance.
• Designed the app interface with intuitive navigation, including options to view diagnosis results, access detailed information, and provide feedback via a 'Not the result' button for model refinement.
• Evaluated model performance using precision, recall, and mean average precision (mAP) across three classes: healthy, monilia, and black pod (Phytophthora).

Experimental results

Research questions

• RQ1 Can a lightweight deep learning model effectively detect and localize 'monilia' and 'black pod' diseases in cocoa pods from standard smartphone images?
• RQ2 How does the performance of EfficientDet-Lite4 compare to other state-of-the-art object detection models in classifying cocoa pod health status?
• RQ3 To what extent can a mobile application integrating a deep learning model improve disease diagnosis accuracy and accessibility for non-expert farmers?
• RQ4 What is the impact of dataset quality and expert validation on the model’s predictive reliability and real-world usability?

Key findings

• The model achieved an average precision of 35% across all classes, with 42.3% precision for 'monilia', 27.3% for 'black pod', and 34.4% for healthy pods.
• The model demonstrated progressive improvement in training, with loss decreasing over epochs, indicating effective learning on the curated dataset.
• The mobile application successfully delivered real-time diagnosis, showing a 96% probability of a pod being healthy when correctly classified.
• The app interface enabled user interaction with diagnosis results, treatment suggestions, and feedback mechanisms, enhancing usability and model adaptability.
• Expert validation by the 'Asociación de Productores Agrícolas Divino Niño' confirmed the accuracy and reliability of the dataset annotations.
• The model's performance was satisfactory despite the limited dataset size, suggesting potential for improvement with larger, more diverse data.