Deep Learning Solutions for Agricultural and Farming Activities

Introduction

With the ever-increasing global population, smart agriculture emphasizes (i) enhancing agricultural productivity and quality of food and (ii) safeguarding the natural ecosystem. It is comprised of automation of the identification of plant diseases, yield prediction, weed detection, insect detection, crop identification, and livestock management using modern techniques, such as cloud-based services, machine learning (ML), Internet of things (IoT), image processing, big data analytics, and many more. Smart agriculture promotes automated farming and the collection of field data using various means, including cameras, micro-controllers, actuators, and others. The data is analysed by IoT or machine learning to deliver useful information for decision making. Traditional machine learning needs to extract the domain features of input image data followed by classification. Feature extraction expects domain expertise as a prerequisite. Furthermore, traditional machine learning methods are not robust enough to handle high volumes of high-dimensional data. Both of these issues are handled by Deep Learning (DL). DL is widely used for automating various aspects of smart agriculture for two major reasons: it can handle huge amounts of data and does feature engineering on input images on its own (Zhu, 2018; Tseng, 2018). The popularity of this technique and its applicability over the years can be seen in Figure 1, which provides information on papers published in 1985 and between 2010 and 2019 on the application of Deep Learning to different techniques.

Figure 1.

Papers published on different applications of deep learning, by year

The next section covers Deep Learning and its advantages and disadvantages, followed by a brief explanation of commonly used DL algorithms: CNN, RNN, LSTM, and GRU. Finally, the authors elaborate on Deep Learning-based agriculture applications and present the conclusions of the study.

Background

Deep Learning is a collection of machine learning algorithms that models high-level abstractions in data by means of architectures comprising multiple nonlinear transformations. DL circumvents the prerequisite of feature extraction needed for classification tasks. DL is a deep multi-layered Neural Network (NN), with large numbers of neurons and the objective of capturing complex, nonlinear relationships in input image data (Kamilaris, 2018; Lecun, 1995; Tseng, 2018).

The major advantages of DL are:

• •

  It has the capacity to carry out feature engineering on its own.

• •

  It can handle huge amounts of unstructured data and can be applied for different domains. including computer vision, time series, language and speech processing, games, and many more.

• •

  It provides precise results compared to traditional machine learning.

• •

  The pre-processing required is much lower as compared to traditional classification algorithms.

The major disadvantages of DL are:

• •

  It performs well only with a very large amount of data for training.

• •

  It takes more time for training (because it does the task of feature engineering on its own).

• •

  There is a lack of transparency in interpreting outcomes (i.e., the system is a black box, and researchers do not know how it drives solutions.)

• •

  It requires high-end machines equipped with expensive high-performance GPUs and a large amount of storage to train the models.