Deep learning-based detection and classification of water stress in lettuce through vision-based approach

Date of Publication

10-2023

Document Type

Master's Thesis

Degree Name

Master of Science in Mechanical Engineering

Subject Categories

Engineering | Mechanical Engineering

College

Gokongwei College of Engineering

Department/Unit

Mechanical Engineering

Thesis Advisor

Laurence A. Gan Lim

Defense Panel Chair

Jeremias A. Gonzaga

Defense Panel Member

Conrad Allan Jay R. Pantua
Timothy Scott C. Chu

Abstract/Summary

As the global population grows, the demand for food and water is increasing, necessitating the efficient use of limited agricultural resources. Water scarcity is a major issue that affects crop growth, yield, and quality. Traditional plant water stress detection methods are time-consuming, expensive, and impractical for large-scale applications. To address this, researchers are developing non-invasive vision-based techniques that detect plant water stress accurately using computer vision and machine learning. Deep learning, a subfield of machine learning, is effective in plant phenotyping. With this, the application of a vision system and deep learning in agriculture can help improve traditional agricultural practices and advance technology in the agriculture sector.

This study is focused on developing a vision-based approach to detect and classify water stress levels in green ice lettuce plants to aid in the irrigation management of the plants in a pot experiment setup. This system aims to help in the water management in the lettuce cultivation. No water stress, mild water stress, and severe water stress were simulated in the lettuce in the pot setup to gather the dataset of lettuce with varying water stress levels. The dataset was used to train a hybrid classification model wherein a pre-trained MobileNetV2 model was used to extract features from the images, and a Support Vector Machine algorithm was trained to classify the degree of water stress in lettuce. The vision-based classification model was implemented using a Raspberry Pi 4B.

Based on the results, the MobileNetV2-SVM model achieved an accuracy of 88.33%. The model’s training time was 238 seconds, and its inference time in the Raspberry Pi 4 was 0.5571 seconds. The implementation of the system to detect mild water stress in green ice lettuce resulted in an increase in crop yield and a reduction in water consumption. This implies that implementing a vision-based deep learning model to water stress in green ice lettuce can potentially boost water productivity, achieving more yield with less water.

Abstract Format

html

Language

English

Keywords

Deep learning (Machine learning)

Upload Full Text

wf_no

Embargo Period

11-6-2023

This document is currently not available here.

Share

COinS