A deep learning approach to Identify crop disease, (an image classification task).
Crop disease recognition poses a significant challenge for agriculture. Yet, with advancements in visual computing and enhanced computational hardware, automated disease identification has become feasible. The effectiveness of Convolutional Neural Network (CNN) architectures has been demonstrated on publicly available datasets. To evaluate the performance of state-of-the-art classification models in real-world, uncontrolled on-site conditions, a curated dataset was assembled, encompassing five sugarcane plant diseases. These images, sourced from fields across diverse regions of Karnataka, India, were captured by camera devices with varying resolutions and lighting conditions.
Trained on this diverse sugarcane dataset, the models achieved an impressive accuracy of 93.20% on the test set and 76.40% on images from different reputable online sources. This highlights the robustness of our approach in recognizing intricate patterns and variations inherent in practical scenarios. In summary, leveraging CNNs on a highly diverse dataset emerges as a promising strategy for the development of automated crop disease recognition systems.
The dataset contains 2940 images of sugarcane leaves belonging to 6 different classes (consisting of 5 diseases and 1 healthy). These include major diseases that affect the crop in India. All the images were taken in a natural environment with numerous variations. The images were taken at various cultivation fields including the University of Agricultural Sciences, Mandya Bangalore and nearby farms belonging to farmers. All the images were taken using phone cameras at various angles, orientations, backgrounds accounting for most of the variations that can appear for images taken in the real world. The dataset was collected with the company of experienced pathologists.
Fig 1. Example of leaf images from our dataset, representing every class. 1) Helminthosporium
Leaf Spot 2) Red Rot 3) Cercospora Leaf Spot 4) Rust 5) Yellow Leaf Disease 6) Healthy.
We trained a Resnet50 model in two ways, by training the entire model in one case, and only the fully connected part in another case. Transfer learning was used in both cases starting from pre-trained weights on the ImageNet dataset. Here also, we note that weights obtained from training the fully connected part only were used as the starting point for the training of the entire network. Optimizer used - Stochastic Gradient Descent with Restarts was used in all the cases
Achieved an acuuracy of 93.0 % on test test. All networks were trained on 80:20 Split. Networks ran for a total of 15 epochs when training only the fully connected layers and 25 epochs when training all the layers
Fig 2. Example image of a leaf from our test dataset suffering from Helminthosporium Leaf Spot
Fig 3.Visualization of activations in the initial layer of Resnet-50 architecture depicting that the model has efficiently learnt to activate against diseased spots on the example leaf.
To test the mode you can run the infer file by loading the saved model and testing on the sample data.
run infer.py