Prediction of maize growth stages based on deep learning

Highlights

• ConvLSTM encoder-decoder model can forecast daily weather factors correctly.

• Hybrid model and data-driven model can predict maize growth stages.

• Data-driven model outperforms hybrid model in forecasting performance.

• The proposed techniques can be used for the arrangement of agricultural activities.

Abstract

An accurate forecast of daily meteorological factors throughout the year is not only of great significance for the study of climate trends in a certain area but also enables the prediction of crop growth stages. Moreover, the prediction of crop growth stages is related to the scheduling of planting and tillage, the determination of machine harvest time, and the prediction of crop yield. However, highly complex dynamics cause large volatility in meteorological factors, so it is very challenging to predict the crop growth stage accurately, based on weather data. To solve this problem, we propose a data-driven encoder-decoder model, using long short-term memory (LSTM) and convolutional LSTM (ConvLSTM), which can be applied to forecast daily sunshine duration, cumulative precipitation, and average temperature for the coming year. To further test the performance of the ConvLSTM-based model, it is compared with the conventional LSTM encoder-decoder model and the convolutional neural network (CNN)-LSTM encoder-decoder model. The results demonstrate that, the ConvLSTM-based model is more accurate than the others for forecasting temperature (MAE = 2.602 °C, RMSE = 3.456 °C), precipitation (MAE = 3.878 mm, RMSE = 10.503 mm), and sunshine hours (MAE = 3.445 h, RMSE = 4.172 h) in 2014–2016. Furthermore, precise forecasting of meteorological factors allows us to develop a hybrid model and a data-driven model for the prediction of each growth stage separately. The hybrid model combines the ConvLSTM encoder-decoder model with empirical models, whereas the data-driven model comprises the ConvLSTM encoder-decoder model and traditional neural network structures. Finally, we compared the two types of models on a real-world dataset from Dandong, and concluded that the data-driven model is more accurate than the hybrid model for prediction of maize growth stages, with ${\mathrm{R}}^2$ in the range of 0.755–0.883, MAE 0.588–2.205 days, and RMSE 0.978–2.729 days. In the future, these models can also be used to predict the growth stages of other crops.

Introduction

A growth stage model is one of the most important parts of a crop growth model (Ceglar et al., 2011). The accurate prediction of crop growth stages can help agricultural workers predict crop yield effectively, arrange farming activities efficiently, and determine an appropriate harvesting time (Van Oort et al., 2011). In the past few decades, many traditional methods have been used to predict the growth stage of crops, most of which function by simulating the processes of actual crop development. In addition, weather conditions play an indispensable role in crop development. Therefore, many researchers have predicted the growth stage of crops from meteorological factors, such as temperature, precipitation, and solar radiation. Of these factors, temperature is one of the main driving forces affecting crop development (Wang et al., 2017). As a consequence, many studies on temperature-based growth prediction have been reported (Kumudini et al., 2014, Soltani et al., 2006, Yun et al., 2017, Zhang and Tao, 2013, Jones et al., 1986). In addition, research has been conducted on the prediction of crop growth stage based on multiple meteorological factors (Jones, 1986, Yang et al., 2004, Zhao et al., 2018). Among these studies, the climate suitability model (Xu, 2014) is notable: this is based on multiple meteorological factors and has been used to analyze the relationship between climate and the crop growth process; it could also be used to predict the crop growth stage and yield. Beyond meteorological data, satellite remote sensing data are the most important for monitoring large-scale crop conditions and obtaining crop growth information (Meng et al., 2008). Hence, studies on the prediction of the growth stage of crops and vegetation using remote sensing data have also increased in recent years (Hu et al., 2009, Liu et al., 2017, Sakamoto et al., 2005, Yu et al., 2012). Most of the above methods depend on existing domain knowledge, and have drawbacks, such as a fixed framework and a lack of flexibility. In contrast with traditional methods, a data-driven model can be used for modeling directly from the data when there is of insufficient domain knowledge. Models created in this way even have a better non-linear fitting ability than models based on domain knowledge (Alhnaity et al., 2019, Haider et al., 2019, Nevavuori et al., 2019, Reddy and Prasad, 2018, Yalcin, 2017). We can conclude all the methods mentioned above can calculate the growth stage of crops or vegetation under certain conditions. However, most of them have the shortcoming that they cannot provide long-term predictions, and need the support of meteorological factors of the forecast year when calculating growth stages. That is, most of the above models need the weather data at the appropriate timescale to function normally, which also limits the model flexibility to a certain extent.

Therefore, these problems have hindered the longer-term prediction of crop growth stage. The key to solving the problems lies in the long-term forecasting of meteorological factors and establishing the climate model in a generalizable manner. The weather data needed for the growth stage model are mostly subsets of the daily meteorological factors throughout the year. Predicting the growth stage of crops requires the forecasting of several weather factors that play an important role in crop growth in the future year. Moreover, the daily weather of a whole year is a time series with large random variation, whose internal characteristics are difficult to learn. Therefore, we need to find a method of providing stable weather data for the growth stage model, based on annual day-to-day meteorological factors. Fortunately, the implicit relationships within the time series can be extracted well by deep learning methods, and they can establish more complex models for data prediction. Consequently, we hope to predict the crop growth stage far in advance by means of deep learning and meteorological data. For the forecast of short-term meteorological factors, a recurrent neural network (RNN) (Mikolov et al., 2010) is a suitable model. This model and extended models have proven to perform well in time series forecasting problems, such as meteorological data prediction (Kumar et al., 2019, Poornima and Pushpalatha, 2019, Qing and Niu, 2018, Wang et al., 2018). Convolutional long short-term memory (ConvLSTM) is a special-purpose variant of RNN that was established by Shi et al. for precipitation nowcasting (Shi et al., 2015). ConvLSTM can be used to construct a good temporal sequence relationship and can perform local feature extraction, like a CNN. Experimental results showed excellent performance in the forecasting of approaching precipitation (Souto et al., 2018). However, most of the above neural network structures are suitable only for short-term weather forecasting. It is difficult to forecast the daily meteorological factors for a whole year independently by using these structures. The research described in this paper addresses this challenge.

In view of the aforementioned characteristics of daily meteorological factors, an encoder-decoder model, combined with both ConvLSTM and long short-term memory (LSTM), was developed in this study. The model can be used to forecast the daily mean temperature, daily sunshine duration, and daily accumulated precipitation, which are suitable for growth stage models, throughout the whole year. In addition, the model is combined with a climate suitability model to form a hybrid model, and combined with back-propagation (BP) neural networks to form a data-driven model. These two types of models are then applied to predict maize growth stages, and their forecasting performance is compared.