deewan07/phythopathology-using-deep-learning

PHYTHOPATHOLOGY-USING-DEEP-LEARNING

Just like human diseases, plant diseases are also evident since the earliest of times. Fossil evidence proves that plant diseases have existed on earth over 250 million years ago and continues to do so. There have been many famines and other economical disasters because of various plant disease outbreaks. Some of the major plant epidemics were late blight of potatoes in Ireland (1845–60), coffee rust in Sri Lanka (1870s), Panama disease of bananas in Asia (1990s) and so on. These epidemics result in major losses for the economy. In this paper we discuss a model using deep learning which would help in getting to know regarding any harm a crop faces before it is damaged completely. This is done by looking closely at a crop’s leaves and further classifying them into categories for better understating of the health of the crop. Here, the dataset is primarily for apple leaves only. Firstly, image preprocessing is done followed by data augmentation. After training and testing of the model the results are out which are then categorised into four classes mainly, healthy leaves, scab, rust and multiple diseased leaves.

DenseNets is a common CNN-based ImageNet that is used by a range of applications such as classification, segmentation, location, etc. For representational capacity, most models before DenseNet relied solely on network depth. DenseNets maximize the network’s ability by feature reuse instead of extracting representational power from extremely deep or large architectures.

EfficientNet is another (newer) common CNN-based ImageNet model that in 2019 was able to perform SOTA in multiple image tasks. With considerably less criteria, EfficientNet executes revolutionary model scale-up to achieve excellent precision.

Proposed Model To make sure that the plant or crop is healthy, detection of infected leaves at an early stage is crucial. This ensures that the farmer gets the harvest they deserve without any hassle. Hence, in this paper we have used Deep Learning DenseNet model for classification of the leaves. We have divided our model in a few steps. Firstly we captured the image of the leaf and then performed image preprocessing. Next, the preprocessed data is augmented by various techniques namely, flipping, convolution and blurring to obtain an increased dataset. Further the edges of the images were detected using the canny edge detection method. And once the model is trained and tested we classify the leaves into four different categories. These classifications are healthy leaves, scabbed leaves, leaves with rust and finally the most infected leaves which are the unhealthy ones with multiple diseases.