CropDoc: A Deep Ensemble Learning Approach for Plant Pathology Detection

Abstract

"Sri Lanka is a country with a rich background in agriculture that was called the ‘Granary of the east’, which resembles the rich agricultural status of the country. Agriculture plays a vital role in the economic growth and balance of the country however, the agriculture domain faces many challenges like diseases, pest attacks, nutrition deficiencies, climate changes, etc. Due to these challenges, about 30% of crop cultivation gets destroyed annually in Sri Lanka. These challenges can be identified and controlled at an initial stage by correct identification and management, however, due to the lack of knowledge among the agricultural community and difficulty in communication between agricultural officers and farmers, these challenges still pertain in the current society.

A novel deep ensemble neural network-based approach is proposed for plant diseases, pests, and nutrition deficiencies classification in this paper using a customized dataset. As the base models of the ensemble approach, a few Keras models were imported and modified for an optimized performance using transfer learning and a few other deep learning techniques. Few Keras models, ResNet50, VGG16, DenseNet121, EfficientNet, XCeption and InceptionV3 were tested for performance in multi-class classification individually for the dataset and the best-performed models were selected as the base models in the ensemble approach. Stacking, averaging and weighted averaging ensemble techniques were considered in the research and stacking and averaging techniques were utilized for the approach considering its relatively high performance.

All base models and ensemble model were trained and tested using training and testing datasets and the ensemble model was able to outperform all the base models by achieving 99.5%, and 99.1% training and testing accuracies respectively. Furthermore, the proposed approach was benchmarked with a common dataset and the results have shown that the proposed deep ensemble approach transcends the performance of the existing work."