Android Based-App Papaya Leaf Disease Identification Using Convolution Neural Network

Abstract

Papaya plant disease can lead to substantial harvest and financial losses for crop owners, potentially affecting the overall income of Malaysia’s agriculture sector. Integrating Artificial Intelligence into the agriculture sector can be a significant leap in attracting young agropreneurs and assisting both existing and young farmers in identifying papaya plant diseases. In line with these challenges, this project proposes papaya plant disease identification using a Convolutional Neural Network (CNN). A total of 225 images were collected from Google sources and real-life captured images, consisting of 3 different classes: Healthy, Ring Spot, and Black Spot. After the preprocessing and augmentation process, a total of 600 images were obtained. Out of 25 Keras API pre-trained CNN models, InceptionV3 was selected as the best base model, as it achieved the highest validation accuracy during a 10-epoch run through Google Colab. Hyperparameters were tuned to obtain the best results by inputting the training images from the top of the base model and extracting 2048 output features from the last layer of the Inception V3 model. The extracted features were saved in the form of NumPy arrays to be employed in the pipeline for hyperparameter tuning, thereby improving the tuning efficiency. The results show the tuned training data achieved a validation accuracy of 1.0 using a batch size of 4, a learning rate of 0.01, and 50 epochs with the SGD optimizer. With this set of hyperparameters, a full model training was conducted with training images as input, resulting in a training accuracy of 1.0 and validation accuracy of 0.96. The trained model was exported in the form of tflite file format and used in the app development through Android Studio. Testing the app’s accuracy involved importing selected 15 images from each class into the designed app, resulting in a precision of 0.8889, recall of 0.8889, f1-Score of 0.8889, and accuracy of 0.8889. These results demonstrate that the accuracy of identifying papaya plant disease through papaya leaves using the designed Android-based app was relatively high.

Keywords

• agriculture
• agropreneur
• android studio
• convolution neural network
• papaya