ANALYSIS OF COTTON PLANT DISEASE DETECTION USING DEEP LEARNING METHODS

Abstract

Disease detection in agriculture has become very crucial in today’s deteriorating climatic conditions. Cotton production plays a vital role in our country’s economy for being a major cotton producer in the world. Monitoring the large cotton fields manually becomes a tiresome activity for farmers. To leverage advancements in technology, deep learning and image processing techniques are being used for identification of diseases in crops at an early stage and thus preventing them from further harm. In this work, we have done analysis of transfer learning techniques for disease detection and proposed a deep convolutional neural network for accurate classification of fresh and diseased plants. The proposed model gives better accuracy and performs faster than pre-trained models like VGG16, ResNet50, ResNet152V2, InceptionV3 and EfficientNet-B0. Various data augmentation methods were used to improve the training process and reduce chances of overfitting. Fine tuning methods were also implemented on InceptionV3 and EfficientNet-B0 along with data augmentation. A detailed analysis of all the techniques used is done for better understanding of transfer learning techniques along with different aspectslike fine tuning and data augmentation. For comparison, the analysis of performance of a model built from scratch by adding convolutional, max pooling and fully connected layers is also done, which gave training accuracy of 94.52%, validation accuracy of 96.88% and test accuracy of 98.11%. Apart from that, it took approximately 50% less time in implementation. So, monitoring of large fields can be done using the proposed model for early diagnosis and treatment for better production.