metadata

title: Plants Disease
emoji: 😻
colorFrom: pink
colorTo: blue
sdk: gradio
sdk_version: 3.29.0
app_file: app.py
pinned: false
license: cc-by-nc-nd-4.0

Introduction

These days there are many challenges facing humanity in terms of food supply, especially considering the fact that there are over 8 billion of us already. Human society needs to increase food production by an estimated 70% by 2050[1] to meet growing demand. Putting aside the joke that global hunger and overpopulation have the same solution, we need we need to focus on improving harvesting.
Currently, infectious diseases reduce the potential yield by an average of 40% with many farmers and even professionals in plant biology experiencing yield losses sometimes as high as 100%[1]. Modern smartphones can assist in identification and treatment of many diseases without costly equipment or professional education. The described solution is perfect for developing countries as well as for huge agricultural companies and scientific institutions.

Methods

Dataset

Dataset[2] consists of images comprising healthy and diseased leaves of different plants. There are 38 classes overall and over 80k images in total. Numerous augmentations strategies are applied to data(rotation, shift, contrast etc.). We asked a test group(couple of friends) to identify disease of the plant, but even identifying plant type can be tricky

Training experiments

During this project two models were finetuned: MobileNetV2 and InceptionV3. While the former one is ligher is also showed higher accuracy and was chosen as a winner, according to table 1.

3 different experiments were performed with each model:

Training from scratch(randomly initialized weights)
Retraining the whole pretrained model
Training last layer based on pretrained network

Table 1. Model performance on a test set after 10 epochs.

Model	Size (Mb)	Experiment 1	Experiment 2	Experiment 3
MobileNetV2	14	0.151	0.601	0.955
InceptionV3	92	0.738	0.865	0.951