Prediction of Rice Yield via Stacked LSTM

Introduction

Rice is one of the most important food sources for more than half of the world's population (Jeon et al., 2011), it is the second most widely grown cereal crop worldwide (Hirooka et al., 2018) and the demand for rice is expected to grow because of the increasing population on earth (Daniela et al., 2018). In order to ensure the rice yield keeping up with the pace of population growth, some measures to guarantee sufficient rice supply, in other words, food security, are needed necessarily. Crop yield prediction is a representative measure which is vital for food security (Hutchinson, 1991). It can obtain the result whether the future crop yield can achieve the demand of population, therefore, it plays a key role in government's policy making and preparing production plan for following year. Proper decisions of government based on crop yield prediction can make more efficient management of crop production processes. On the other hand, crop yield prediction can provide a reference for farmers and enterprise, thus helping them increase outcome (Na-Udom & Rungrattanaubol, 2015), so rice yield prediction is a matter of importance.

Crop simulation models are used extensively, simulation of plant-growth stages and consequently forecasting the crop yield permits better planning (Inoue, Susan & Horie, 1998). Crop growth simulation model are combined with remote sensing data to estimate rice yield (Abou-Ismail, Huang & Wang, 2004; Inoue, Susan & Horie, 1998). N. Pirmoradian et al. (2006) proposed a very simple model(VSM) to predict rice grain and biomass yields. But the application scope of crop growth models is a little limited, they may only adapt to a few crop species. Meanwhile, the development time and cost of these models are extremely large (Alberto, Juan & Waldo, 2014). Many statistical methods are also used for yield prediction in early stage, such as Multiple Linear regression, logistic regression (Cakir, Kirci & Gunes, 2014), which eliminate the limitation of application range. Liu Qin-pu (2011) proposed grain yield spatio-temporal regression prediction model of Henan Province; Umid Kumar Dey et al. (2017) applied Multiple Linear Regression AdaBoost to the task of predicting rice yield in Bangladesh. R. K. Pal (2012) developed the multiple regression equations for prediction of growth and yield attributes of wheat. The complicated relationships between crop production and interrelated factors cannot be described well by statistical models (Cakir, Kirci, & Gunes, 2014). Due to the aforementioned shortcomings of traditional models, many methods based on machine learning are inspired to solve yield prediction tasks recently. Ratchaphum Jaikla et al. (2008) used the Support Vector Regression method to predict rice yield and acquired comparable performance with Crop Simulation Model. Artificial neural networks (ANN) is a popular method used in crop prediction. For example, Yüksel Çakır et al. (2014) used ANN to predict wheat yield in south-east region of Turkey; Niketa Gandhi et al. (2016) applied ANN with multilayer perceptron to rice yield prediction from year 1998 to 2002 in India.