1 of 1

From Diagnosis to Advice: A Deep Learning (AI) System for Sugarcane Leaf Health in Eswatini – A Case Study of the Eswatini Sugar Association

​

By: Khanyisile Tapiwa Magagula�Department of Computer Science, Faculty of Science and Engineering

University of Eswatini, Kwaluseni, M201, Eswatini

​

Khanyisile Tapiwa Magagula

Email: khanyitapiwamagagula99@gmail.com

References

A QR code can be used for the references or they can be written down like below (please refrain from underlining weblinks)

[1] Sindh Agriculture Department. (n.d.). About sugarcane. Retrieved from https://agri.sindh.gov.pk/about-sugarcane

[2]Iqbal, Muhammad Aamir, and Asif Iqbal. "Sugarcane production, economics and industry in Pakistan." Am. J. Agric.Environ. Sci14.12 (2014): 1470-1477.

[3]Hossain, Md Imam, et al. "Current and Prospective Strategies on Detecting and Managing Colletotrichumfalcatum Causing Red Rot of Sugarcane." Agronomy 10.9 (2020): 1253.

[4]Kumar, Arpan, and Anamika Tiwari. "Detection of Sugarcane Disease and Classification using Image Processing." Int J ResAppl Sci Eng Technol 7.5 (2019): 2023-30

[5]Springer. (n.d.). Chapter: Convolutional Neural Networks. In Lecture Notes in Computer Science (Vol. 8689, pp. 592–601). Springer. Retrieved from https://link.springer.com/chapter/10.1007/978-3-319-07353-8_53AC

​

Introduction

  • Deep learning is a machine learning technique that teaches a computer to filter inputs through layers in order to learn how to predict and classify information. (IBM, 2021).
  • Diseases found in Eswatini’s sugarcane plants especially those that affect the leaves are a major concern and a threat to farmers and that triggers an economic impact, affects the sugarcane yields and production if not detected on time.
  • Detecting and diagnosing sugarcane leaf diseases as soon as possible is a very critical initiative to the farmers.
  • Deep learning-based approaches have shown promising results in image-based sugarcane leaf disease detection and diagnosing, providing high accuracy and efficiency in detecting multiple types of sugarcane leaf diseases.

Problem Statement

  • In the case of Eswatini Sugar Association, sugarcane leaf diseases are detected using traditional techniques(naked-eye observation), which is done by a crop disease and infection expert.
  • It usually requires a continuous observation to confirm and validate the diagnosis.
  • Infected sugarcane leaves are manually recognized based on the leaf spot characteristics on a daily bases.
  • It is a very time-consuming process which is also not always accurate.
  • To overcome the setbacks of the traditional techniques, a need for having an automated detection and diagnosis system of the sugarcane leaf disease is needed.

Objectives

  • To review existing methods used for sugarcane leaf disease detection and diagnosis in Eswatini Sugar Association.
  • To review existing computational models that are used for sugarcane leaf diseases detection and diagnosis.
  • To develop a model that will detect and diagnose sugarcane leaf diseases using deep learning algorithms based on Eswatini Sugarcane leaf disease dataset.
  • To implement and evaluate the performance of the proposed model.

​

Conclusion

In conclusion, the computerized Eswatini e-sugarcane leaf disease detection is an essential tool for detecting and diagnosis of sugarcane leaf diseases. The web app was designed using Flask and includes a user-friendly interface to enhance the user experience. The web app provides information about sugarcane diseases and their control and includes a list of over five detections of sugarcane leaf diseases caused by fungi, bacteria, viruses, and nematodes. The app also includes a sugarcane disease detector that provides recommendations for diagnosis of the detected disease.

Furthermore, by using both black box testing and white box testing, we can ensure that the sugarcane disease detection website is accurate, efficient, and easy to use for end-users. Overall, the sugarcane disease detection and diagnosis system has the potential to improve the yield and productivity of sugarcane plants. It is a valuable resource for sugarcane farmers, researchers, and other stakeholders in the sugarcane industry in Eswatini.

Results

Each model trained was capped at 50 epochs due to the lack of sufficient resources. In this section, the visual analysis of each model during training was captured and the DL algorithms used were compared for purposes of selecting the best amongst them and deploying onto a web based system.

​

Evaluation Metrics

​

​

Materials and Methods

​

​

​

​

​

Architecture of Proposed system

�

(1)

(2)

(3)

The bar chart compares the performance of deep learning algorithms, YOLOv8 CNN, SE-ResNet50 and SA-ResNet50 across four metrics: precision, recall, mAP@50, and mAP@50-95. SA-ResNet50 outperforms in all metrics, with slightly higher values for precision, recall, mAP@0.50 compared to the SE-ResNet50 model, which had a better performance with mAP@50-0.95. These results position SA-ResNet50 as the most optical model for Eswatini’s sugarcane leaf disease detection.

The images show the web application that was built. The best weights retrieved after training SA-ResNet50 were downloaded and used to build the final product.

International Conference on

EVOLVING CITIES

University of Southampton

SEPT. 12-14