Smart Agriculture : Use of Computer Vision, ML and IoT to Improve Crop Productivity

Team

Supervisors

Table of content

  1. Abstract
  2. Related works
  3. Methodology
  4. Experiment Setup and Implementation
  5. Results and Analysis
  6. Conclusion
  7. Publications
  8. Links
  9. Summary

Abstract

Agriculture is the main source of the world’s food supply. It has been the backbone of civilizations for thousands of years. During recent decades, the focus of agriculture has shifted to producing crops in more controlled and optimal environments in order to increase the quality and quantity of the yield at lower costs to fulfill the massive food requirements of a continuously growing population. Modern smart greenhouses are buildings that are equipped with the latest IoT devices, environment control and monitoring systems and automated irrigation and fertilizer systems to provide optimal growth conditions for crops. The latest trend in smart agriculture is to incorporate computer vision and machine learning (ML) capabilities to predict and optimize existing cropping systems to maximize productivity. The research focus on reduce the harm cased by different diseased by introducing disease/anomaly detection and identification so that the earlies and the necessary actions can be taken.

Methodology

By using computer vision techniques, solutions to address the following scenarios are to be done

Apart form these main objectives, it is planned to observe some of the exsisting models performance on the system and to develop a local dataset.

Experiment Setup and Implementation

Data collection

The aerial data was collected using both the cameras, and the amount is 500 and 300 each. Plant leaves data were gathered with smartphone which included 400 images and those images were categorized after running sample tests in Horticultural Crop Reseasech and Development institute in Peradeniya. Then in order to develop more accurate solutions in the preprocessing phase, color calibration with backgroud removal was done alog with the annotation and categorisation of the data.

Results and Analysis

Disease Identification

Disease Detection

Conclusion

In conclusion, the custom CNN model which was developed from scratch gave the best results that can be applied to the real world application that is being developed, considering the disease identification phase.
Considering the disease detection phase, it leave the question of the robustuness due to the only anased feature was color variation and upto the depth the solution is useful in a real wolrd system. As a begining of this kind of an apprach, it was possible to achieve more that 50% of accuracy, and this can be further improved in future research.

Publications

Summary

The research was conducted aiming to introdude computer vision, machine learning techniques to develop a smart agricultural system, and addressing this concept, two approaches were taken to adrress the issue of avoiding/reducing the threat of diseases to improve the quality and the quantity of the yield. the main objective was to develop a disease detection system with the use of aerial images analysing the phenotying characteristics of plants. This was done with a help of a mask RCNN model which we could obtain a result of accuracy 52-28% based on the dataset used. Other approach was to develop a disease identification system which has the ability to identify 3 disease classes in this research. For this apprach the existing inception v3 model was observed with its performance and a custom CNN was developed from scratch which resulted the best performance in comparison which is 87.5% if accuracy. The models gave considerably satisfying results and they are further to be improved ot be used in real world applications.