Image-Based Crop Disease And Severity Classification Using Convolutional Neural Network

Abstract

Crop Protection-Related Problems Are The Main Challenges In The Agriculture Sector Since They Have A Large Impact On Crop Yield Reduction. In Particular, Disease Is One Of The Major Protection Problems That Result In Food Insecurity And Agricultural System Unsustainability. To Prevent Crop Disease From Spreading At The Earliest Stage, Modern Technology Is Preferable To The Traditional Approach. In This Study, A Modified Convolutional Neural Network Model Is Proposed To Identify And Classify Disease Types And Severity Levels On Wheat And Maize Crops By Extracting Leaf Features. Six Types Of Disease Classification And Three Levels Of Severity Are Taken Into Consideration In Addition To Healthy Leaf For Each Crop. The Proposed Model Contains A Modified Version Of The VGG16 Architecture And A Pair Of Channel Shuffle Blocks As An Auxiliary Structure, Which Enables It To Perform Multi-Labeled Classification With A Single Model. It Has Been Trained Through Backpropagation And Optimized With The Adam Optimizer Using Field Crop RGB Image Training Datasets. To Provide A More Practical Approach, 65% Of The Data Was Gathered From Research Centers, And The Remainder Was Gathered From Privatedata Sources In Combination With Greenhouse Images With Manual Farm Field Capturing. Also, To Expand The Collected Data Set And The Variation Of Data, Different Augmentation Techniques Are Applied. A Highly Accurate Result Has Been Achieved After Experimenting With Different Regularizers And Different Hyper Parameters. The Proposed Model Is Composed Of The Combination Of The VGG Structure And Shufflenet Structure And It Achieves 0.98 Accuracy For Both Disease And Severity Classification. The Model Shows Its Robustness By Showing A Performance Of High Accuracy With Low Computational Cost Compared To VGG Model And Better Practicability Since It Considers Real Field Data. The Multi-Labeling Classification Ability Makes The Model Multi-Tasking, With Reduced Computational Time And Data Simultaneously.