ABSTRACT
Fall armyworm (Spodoptera frugiperda) is an invasive pest that attacks a wide range of plants (Early et al., 2018) It is especially notorious for attacking one of Africa’s most important foods: maize which is a source of livelihood and a staple food for millions of people across the continent. (Day et al., 2017). Current approaches used in fall armyworm monitoring require physical presence of an agricultural expert (agricultural extension officer or plant entomologist) to guide farmers in the identification of fall armyworm damage on maize leaves. Without expert training, farmers could easily confuse FAW attacks with other common maize pests leading to delayed or incorrect intervention measures and can lead to the loss of an entire crop. Meissle et al. (2010). In the recent past, machine learning techniques have been applied in pest detection. (Ebrahimi et al., 2017; Voulodimos et al., 2018). Despite the potential benefits offered by current machine learning approaches in literature, there lacks a CNN based mobile artifact that offers an easy-to-use alternative to classify and localize fall armyworm damage on maize leaves in the natural farm environment. This research compares the performance of two one stage convolutional neural network meta- architectures to develop a FAW damage detection mobile application. Experimental results show impressive performance, with the best performing efficientdet lite model achieving a mean average precision of 85.85% and the best performing yolov4 tiny model achieving a mean average precision of 82.5%.
Table of Contents
DECLARATION i
ACKNOWLEDGEMENTS iii
LIST OF FIGURES vi
LIST OF TABLES vii
LIST OF EQUATIONS viii
ABBREVIATIONS AND ACRONMYN ix
ABSTRACT x
CHAPTER ONE: INTRODUCTION
1.1 BACKGROUND OF THE STUDY 1
1.2 PROBLEM STATEMENT 2
1.3 THE OBJECTIVES 3
1.4 SIGNIFICANCE OF THE STUDY 3
1.5 RESEARCH CONTRIBUTION 4
1.6 SCOPE, ASSUMPTIONS AND LIMITATIONS OF THE STUDY 4
1.7 ORGANIZATION OF THE RESEARCH THESIS 5
CHAPTER TWO: LITERATURE REVIEW
2.1 MAIZE FARMING IN KENYA 6
2.2 COMMON MAIZE PESTS IN KENYA 7
2.3. OVERVIEW OF FAW IN KENYA 8
2.4. FAW LIFE CYCLE 10
2.5. FALL ARMYWORM MONITORING TECHNIQUES 11
2.6 COMPUTER VISION IN AGRICULTURE 12
2.7 ONE & TWO STAGE CNN META-ARCHITECTURES 13
2.8 RELATED WORK 17
2.9 CONCEPTUAL MODEL 20
CHAPTER THREE: RESEARCH METHODOLOGY
3.1 INTRODUCTION 21
3.2 STUDY SET UP 21
3.3 RESEARCH DESIGN 21
3.4 DATASETS 21
3.4 MODEL TRAINING AND TESTING 23
3.5 SOFTWARE DEVELOPMENT METHODOLOGY 25
3.6 ETHICAL CONSIDERATIONS 26
CHAPTER FOUR: SYSTEMS DESIGN AND ARCHITECTURE
4.1 REQUIREMENTS ANALYSIS 27
4.2 SYSTEM ARCHITECTURE 28
4.3 SYSTEM BEHAVIOR MODELING 29
4.4 PROCESS MODELING 33
4.5 SYSTEM IMPLEMENTATION 34
CHAPTER FIVE: RESULTS, EVALUATION AND DISCUSSION
5.1 INTRODUCTION 36
5.2 CONVOLUTIONAL NEURAL NETWORK TRAINING 36
5.3 QUANTITATIVE RESULTS 37
5.4 SYSTEM PROTOTYPE TESTING 38
5.5 PERFORMANCE COMPARISION WITH PREVIOUS STUDIES 40
CHAPTER SIX: CONCLUSION AND RECOMMENDATION
6.1 CONCLUSION 41
6.2 RECOMMENDATIONS 41
References 42
LIST OF FIGURES
Figure 1: Agroecological Zones of Kenya 6
Figure 2:African Maize Stalk borers & Maize Leafhoppers 7
Figure 3:Maize Aphids & Bollworms 8
Figure 4:Cutworms 8
Figure 5:Fall Armyworm Status in Africa 9
Figure 6:Fall Armyworm Status in Kenya 10
Figure 7:Fall Armyworm Life Cycle 11
Figure 8: Fall Armyworm Pheromone Traps 12
Figure 9: One Stage Meta-architecture 13
Figure 10: Two Stage Meta-architecture 14
Figure 11: EfficientDet Architecture 16
Figure 12: EfficientDet Performance on COCO Dataset 16
Figure 13: YOLOv4 Architecture 17
Figure 14: Yolov4 tiny Performance on COCO Dataset 17
Figure 15: CNN Model Training Conceptual Model 20
Figure 16: CNN System Conceptual Model (with selected algorithm) 20
Figure 17: Late and Early FAW Infestation Annotated Images 22
Figure 18: Rapid Application Development 26
Figure 19: System Architecture 28
Figure 20: Use Case Diagram 29
Figure 21: Sequence Diagram 32
Figure 22: Context Diagram 33
Figure 23: Level 1 Data Flow Diagram 34
Figure 24: Paired Sample Statistics 37
Figure 25: Paired Samples Correlations 38
Figure 26: Paired Sample Differences 38
Figure 27: Sample Mobile Application Screenshot 39
Figure 28: Sample Screenshots of FAW Damage Object Detection 39
LIST OF TABLES
Table 1: Train, Validation, Test Split 22
Table 2: Upload Image use case narration 30
Table 3: View Dashboard use case narration 30
Table 4:Perform Object Detection use case narration 31
Table 5: YOLOv4 TINY mAP 36
Table 6: EfficientDet lite04 mAP 36
Table 7: Test Case Results 38
Table 8: Performance Comparison with other Studies 40
LIST OF EQUATIONS
Equation 1: Precision and Recall 24
Equation 2: Intersection Over Union 24
Equation 3: Average Precision 24
Equation 4: Mean Average Precision 25
ABBREVIATIONS AND ACRONMYN
FAW - Fall Armyworm
IPM - Integrated Pest Management
FAO - Food and Agriculture Organization
KARLO - Kenya Agricultural & Livestock Research Organization
CABI - Centre for Agriculture and Bioscience International
CV - Computer Vision
CNN - Convolutional Neural Network
YOLO - You Only Look Once
mAP - Mean Average Precision
CHAPTER ONE
INTRODUCTION
1.1 BACKGROUND OF THE STUDY
Fall Armyworm (Spodoptera frugiperda) is an invasive pest that attacks over 350 plant species. It is native to the tropical and sub-tropical regions of the Americas. (Early et al., 2018) but has over the last five years spread to the African continent (initial reports in 2016) and Asia (initial reports in 2018). The spread and devastating effects of the fall armyworm attacks have been felt especially in Africa since it attacks maize plants considered a major staple food and source of livelihood for millions of farmers in the continent (Day et al., 2017). The food and agriculture organization notes that farmers loose between 20 - 40 % of their yields to pests and diseases threating the state of food security FAO. (2020).
Currently FAW monitoring and detection is done through manual based monitoring (farm scouting) and trap-based monitoring (pheromone traps). Like many other manual pest monitoring techniques, these approaches are subjective, delayed and hard to implement at scale.(Dai et al., 2016; Selvaraj et al., 2019; Thenmozhi & Reddy, 2019) (He et al., 2019). Lack of prompt action in case of a FAW invasion can lead to loss of an entire crop yield (Kassie et al., 2020). Researchers have studied different ways computer vision techniques can be applied in the agricultural discipline (Ghadge et al., n.d.; Paul et al., 2020; Tian et al., 2020) Classical image processing and deep learning techniques have been proposed to detect anomalies in plants.(Ensari et al., 2020; Jayswal & Chaudhari, 2020; J. Liu & Wang, 2021; Patil et al., 2020; Rustia et al., 2021; Syarief & Setiawan, 2020). The adoption of deep learning approaches in object classification and detection has greatly improved the computer vision tasks in comparison to traditional image processing techniques (Voulodimos et al., 2018). Existing models lack generalizability when tested in natural environments. This is mainly because images used to train the said models are taken on plain backgrounds in a laboratory setting. (Selvaraj et al., 2019).
The researcher seeks to develop a convolutional neural network-based solution that provides timely and accurate FAW damage detection on maize leaves. The proposed solution will leverage the power of deep neural network and deep transfer learning (C. Tan et al., 2018) to train and deploy CNN models that unique distinguish fall armyworms damage on maize leaves from infestation by other pests. This research project will compare two algorithms based on one-stage CNN meta-architectures (Huang et al., 2017). The better performing model will be deployed on a mobile application for use in the farm. The resulting artifact will provide a vital integrated pest management tool for stakeholders in the maize value chain including farmers, governmental and non-organizational organizations taking keen interest in integrated pest management strategies against the fall armyworm.
1.2 PROBLEM STATEMENT
Pest and diseases negatively affect crop growth process and the resulting yield that is harvested (Cerda et al., 2017). Current approaches used in FAW monitoring require physical presence of an agricultural extension officer or plant pathologist (farm scouting and inspecting pheromone traps). In addition to that farmer use some empirical knowledge they have acquired along the way to identify presence of FAW in their farms. Though effective in some cases, more often than not, it is subjective and prone to errors (Thenmozhi & Reddy, 2019). Without expert training, farmers confuse FAW attacks with other common maize pests such as Cotton bollworm (Helicoverpa armigera) and Southern armyworm (Spodoptera eridania). Untimely and Incorrect identification of a pest leads to delayed corrective measures and can lead to loss of an entire crop (Meissle et al., 2010).
Previous literature relating to pest identification in computer vision has concentrated more on pest classification and estimating population on pheromone traps. (Thenmozhi & Reddy, 2019) (Wang et al., 2015) proposed a CNN based system to classify 40 classes of insect species found in the Xie1, Xie2 and National Bureau of Agricultural Insect Resources (NBAIR) datasets. (Chiwamba et al., 2018) proposed the use of CNN system in embedded Raspberry Pi in counting the number of FAW moths on a pheromone trap so as to estimate its population. Other techniques employed in pest identification are based on combining traditional image processing techniques and classical machine learning such as histogram of oriented gradient (HOG) or Scale invariant feature transform and Support Vector Machines. While these methods show impressive results in pest classification(Fuentes et al., 2017) they cannot be generalized in identification of other pests. The researcher seeks to extend the current body of knowledge on the application of CNN based algorithms in pest detection. In this case the application will target FAW damage on maize leaves.
The researcher seeks to investigate the application of one stage convolutional neural networks in monitoring FAW attacks. This by extension will improve the existing FAW monitoring techniques.
1.3 THE OBJECTIVES
1.3.1. GENERAL OBJECTIVE
Develop a mobile based fall armyworm (FAW) damage detection system using one stage convolutional neural network meta-architectures.
1.3.2. SPECIFIC OBJECTIVES
1. Collect field data of images of the damage done by fall armyworm.
2. Compare the performance of YOLO v4 tiny and EfficientDet lite CNN meta - architectures.
3. Develop an android mobile application that detects FAW damage on maize leaves using one stage convolutional neural network meta-architectures.
1.4 RESEARCH QUESTIONS
1. Can YOLOv4 tiny or EfficientDet lite CNN algorithms be used to develop an accurate FAW damage detection model?
2. How does the performance of YOLOv4 tiny and EfficientDet lite CNN meta- architectures compare in detecting FAW damage on maize leaves?
3. Can the aforementioned model be integrated into a mobile application for FAW damage detection?
1.4 SIGNIFICANCE OF THE STUDY
The research proposes an accurate and timely solution for identifying FAW invasion on maize leaves. FAW is a serious threat to an already weak food security situation on Africa. Since it was first reported in the continent in 2016, yearly maize losses due to FAW attacks are estimated at 9.4% - 66% (Baudron et al., 2019; Day et al., 2017; Kumela et al., 2019). By using the proposed system in FAW damage detection, different stakeholders will benefit. First the society at large will enjoy increased food security and improve environmental conservation. Early detection of FAW helps farmers take corrective actions to protects their plants from further destruction (Fuentes et al., 2017). Early detection also ensures that correct amount and type of pesticides are used, saving the environment from hazardous effects due to excessive use of pesticides (He et al., 2019). The research also adds on the efforts of non- governmental organizations such as Food and Agriculture Organization and Centre for Agriculture and Bioscience International researching on integrated pest management solutions for dealing with the fall armyworm and other pests (Khatri et al., 2020). The research also provides a robust and accurate tool for farmers that helps them reduce economic losses by detecting FAW invasions early and taking the necessary steps in safeguarding their crops. Finally, the research benefits computer vision researchers interested in deep neural network methodologies in pest detection.
1.5 RESEARCH CONTRIBUTION
The research contribution is the proposed approach of using one stage CNN meta-architecture mainly focused on mobile devices. The application provides an end to end platform that takes images as an input, performs object detection on the images and output an image with bounding boxes and the confidence scores highlighting areas attacked by FAW. The resulting artifact can be installed on mobile devices and used in the field. This provides a step forward in incorporating vision-based application in integrated pest management. The artifact works on devices with diverse camera resolutions and in complex environments such as as varying backgrounds, natural lighting, camera orientation, illumination etc.
1.6 SCOPE, ASSUMPTIONS AND LIMITATIONS OF THE STUDY
The study’s scope is limited to detecting FAW attack on maize leaves. Other crops attacked by the FAW pest are not considered. The images used to train and evaluate the models were taken in Kirinyaga county, Kenya so the application might not generalize well in areas with different environmental conditions. The study will use one stage convolutional neural network-based approach to detect presence of fall armyworm on maize leaf images taken in the natural environment and output an image with bounding box and confidence score around regions where fall armyworm damage is identified. Future versions and configurations of the meta-architectures might show varying results from those presented in this paper.
The researcher utilizes commercial and open-source software packages and libraries in performing this study and is therefore limited to the capabilities of said tools. The researcher will utilize more than one software package/library when deemed necessary to overcome this limitation.
1.7 ORGANIZATION OF THE RESEARCH THESIS
This thesis includes five chapters. The first chapter gives an introduction of the research, the problem statement, research objectives and the research contributions. The second chapter gives relevant literature review related to the FAW, computer vision techniques and algorithms the researcher seeks to explore. Related works in this research will also be highlighted. The third chapter covers the research methodology used to train the models and the approach followed in developing the mobile application. The fourth chapter presents the results and discussion after training and deploying the CNN models on the mobile application. Chapter five covers the Conclusion and recommendations for future research.
Click “DOWNLOAD NOW” below to get the complete Projects
FOR QUICK HELP CHAT WITH US NOW!
+(234) 0814 780 1594
Login To Comment