Identify Diseases in Crops leaf (planter)

After extreme weather conditions and natural disasters, plant diseases and pests play a bigger role in the loss of our crops annually at least 20 - 40 % globally. Plant diseases oppose a greater threat to small and medium level farmlands because of a few reasons. One thing is lack of knowledge because farmers unable to identify situations in early-stage diseases made great disasters. Other reasons are linked to the first reason. Which are if they need to identify diseases early stage they need experience as human experts. In both rural and developing countries struggling to find expert knowledge, even though they are willing to find human experts, farmers cannot be able to spend a lot of money on hiring human experts. So in rural areas or in developing countries are lack that expert knowledge and also farmers are unable to spend a lot of money on hiring human experts and costly tests to identify plant diseases. Early detections are effective, but they need lots of resources and a bunch of human experts. So needs automatic, inexpensive and effective methods need to address this problem.

Because of the recent advantages of deep learning methods, it is used almost always everywhere in every field. There is no change in computer vision. Deep learning especially convolutional neural networks are the go-to method for classification, detection and segmentation tasks. There are so many convolutional architectures that achieve state-of-the-art situations in computer vision tasks. The most popular architectures are AlexNet[1], VGGNet[2], GoogleNet[3], ResNet[4] and they dominated classification and detection tasks.

There are already many studies done on disease identification and detection. But most of them are done on medium and high-quality images and also millions of images covering more situations. It is best to model but they need high computation power which poor farmers or researchers in poor countries can’t afford because of its cost. Because of that, we need a model that can train on low-quality images and various situations training for a more robust model. Network architecture after VGGNet[2] comes with deeper networks because deep networks are more accurate and increase the performance of the model. But it comes with a few drawbacks. The deeper the networks more the parameters to train and more parameters need long training and more computer power to train. And deeper networks are more likely to override on small datasets and lead to gradient vanishing resulting in stuck training. So more layers do not always improve model performance.

One solution for this problem is to use skip connections. ResNet used this solution to effectively go deeper without dramatically increasing parameter count and training time. Also, they used heavy batch normalization. With all of these discoveries in our hands, we create a hybrid solution to train with low-quality low-resolution images that can work on a mobile platform. It is 34 weight layer deep and used concepts' coming from the residual net, google inception and also VGGNet to a more robust lightweight model. It trains with a dataset combined few well known and widely used leaf disease identification studies. Which are PlantVillage[5], Tea leaf disease dataset[6], Rice leaf disease dataset[7] and Banana leaf dataset[8].

Methodology

Data sources

Data acquisitions are done in two ways. The main way of collecting labelled training samples is using already published leaf diseases data sets. For this work we used four different data sets which are the well-known plant village dataset[], rice leaf disease dataset[], banana leaf disease dataset[], and tea leaf dataset[]. There are 54000 images, 14 plants and 42 subclasses in total in the plant village dataset. But we use a subset of it. Which include 11 plant types and 35 subclasses. In the rice leaf disease, dataset 3355 images belong to four classes. In the banana dataset, there are 1289 images in three subclasses and tea-leaf disease is quite small which only has 60 images belonging to three subclasses. All images are in RGB colour schema but have different dimensions, in the plant village dataset all images are in 256x256 dimensions also the rice leaf dataset has the same dimensions. In tea, leaf images are in 133x200 and banana leaf images are in 150x113 dimension which is a much small dimension than other data we had.

Data pre-processing

Images need to have identical dimensions before being fed into the CNN so we resize those images to a single scale which is 224x224. Before that, most of the images are centered properly and had imbalance classes. We use methods to increase low sample size classes to some balance. One is using crop by using anchor points. Here we use five anchor points to create crops in initial sample points of more than 1000 and less than 3000 samples. Which are the top right corner, top left corner, bottom right corner, bottom left corner and center of the frame. To do this cropping we resize the image into a 256x256 shape. Our target is to increase the sample size to 3000 to 4000 in those classes. For classes that have fewer samples than 1000, we used a different approach which is to use random crops alongside those bounding box cropping. Here we create a 2000 - 3000 sample size for those classes. After cropping then we resize all images to 224x224. Then we apply random rotation between -200 to 200. Also, color space changes such as brightness, hue, and saturation are changed randomly.

Experimental architecture for disease identification

several state-of-art CNN architectures were tried with augmented data such as VGG16, Inception and ResNet. and few hybrid architecture also tried which are created using above mention architectures. finally, hybrid model constructed by adopting inception and residual networks with heavy batch normalization. for preventing from over-fitting as above-mentioned data augmentation were used. other than that dropout layers were used in after each fully connected layer in classifier in each and every architecture. rectified linear unit activation function was used in hidden layer and 42 way softmax activation was used as final layer. because of limited computing power 224x224 size images was used to training.

The Planter android application

Is entirely programmed with the native kotlin language and models created in the early stages to identify plants from other objects and identify diseases that are inherited into this application. Both models are unable to run on mobile devices because of their higher need for computing power so TensorFlow lite API is used to convert those models into mobile platform friendly models. After the plant, the identification model identifies a plant captured by the user that image pass into the disease identification model to identify is it healthy or not and what will be disease or diseases found in that input image. They show the results to users with more details about the findings.

Reference

• Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556. 2014.

• Krizhevsky A, Sutskever I, Hinton GE. Imagenet classification with deep convolutional neural networks. In: Advances in neural information processing systems. 2012. p. 1097–105.

• Szegedy C, Liu W, Jia Y, Sermanet P, Reed S, Anguelov D, Erhan D, Vanhoucke V, Rabinovich A. Going deeper with convolutions. In: Proceedings of the IEEE conference on computer vision and pattern recognition. 2015. p. 1–9.

• He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. 2016. p. 770–8.

• Mohanty SP, Hughes DP, Salathé M. Using deep learning for image-based plant disease detection. Front Plant Sci. 2016;7:1419.

• hu, gensheng; wu, haoyu; zhang, yan; wan, mingzhu (2019), “Data for A Low Shot Learning Method for Tea Leaf’s Disease Identification”, Mendeley Data, V1, DOI: 10.17632/dbjyfkn6jr.1

• rice_leaf_images (1.0). (2021, November 16). [The code in this project is part of A Project which is aimed at detecting diseases and pests in rice using rice leaf images. The model build of this project was embedded in hardware - Arduino(upgraded to a raspberry pi 4).]. https://www.kaggle.com/nizorogbezuode/rice-leaf-images

• hailu, yordanos (2021), “Banana Leaf Disease Images”, Mendeley Data, V1, DOI: 10.17632/rjykr62kdh.1

• GmbH, P. (2015). Plantix | Best Agriculture App. Plantix. https://plantix.net/en/

• Belasque J Jr, Gasparoto M, Marcassa L. Detection of mechanical and disease stresses in citrus plants by fluorescence spectroscopy. Appl Optics. 2008;47(11):1922–6.

• Feng J, Liang My, Zhao B, et al. Multispectral imaging system for the plant diseases and insect pests diagnosis. Spectrosc Spect Anal. 2009;29(4):1008–12.

• Meroni M, Rossini M, Picchi V, Panigada C, Cogliati S, Nali C, Colombo R. Assessing steady-state fluorescence and PRI from hyperspectral proximal sensing as early indicators of plant stress: The case of ozone exposure. Sensors. 2008;8(3):1740–54.

• Zhang M, Qin Z, Liu X, Ustin SL. Detection of stress in tomatoes induced by late blight disease in california, USA, using hyperspectral remote sensing. Int J Appl Earth Observ Geoinf. 2003;4(4):295–310.

• Liew O, Chong P, Li B, Asundi A. Signature optical cues: emerging technologies for monitoring plant health. Sensors. 2008;8(5):3205–39.

• Chaerle L, Van Der Straeten D. Imaging techniques and the early detection of plant stress. Trends Plant Sci. 2000;5(11):495–501.

• MacKenzie DJ, McLean MA, Mukerji S, Green M. Improved RNA extraction from woody plants for the detection of viral pathogens by reverse transcription-polymerase chain reaction. Plant Dis. 1997;81(2):222–6.

• Martinelli F, Scalenghe R, Davino S, Panno S, Scuderi G, Ruisi P, Villa P, Stroppiana D, Boschetti M, Goulart LR, et al. Advanced methods of plant disease detection. A review. Agron Sustain Dev. 2015;35(1):1–25.