Deep Learning Approach for Water Management in Agriculture

Abstract

Water is an indispensable natural commodity, a fundamental human need, and the most critical source for the existence of the human race and its development. Presently, the world is grappling with water scarcity issues which have risen to an alarming mark. Climate change projection shows that the situation will become grimmer and more prevalent in the near future. The agriculture sector is the primary consumer of water and uses almost 75-80% of the freshwater available worldwide. India, an agricultural country, requires a huge amount of water for irrigation. Still, it has only 4% of the world’s freshwater to serve the food demands of about 18% of the global population. Thus, precise estimation of the irrigation requirements is a significant task to manage the available water resources efficiently and optimize water usage in agriculture. Researchers are trying hard to develop efficient irrigation techniques. The water requirement problem requires more efficient, intelligent, and sustainable techniques

to address the problem. Evapotranspiration (ET) is a significant factor in determin- ing crop water requirements. The precise estimation of reference evapotranspiration

(ET0) and crop evapotranspiration (ETc) is necessary to determine the irrigation

requirements to maintain crop-water balance. Deep learning (DL) models are effi- cient in handling complex non-linear relationships with a massive amount of data.

It shows the ability to handle time-series forecasting problems. Estimating evapo- transpiration shows similar characteristics, such as complex non-linear relationships

among meteorological parameters and the time-dependence nature of these param- eters.

Therefore, the work reported in this thesis is carried out to develop deep learning- based models for the precise estimation of ET0 and ETc values. These proposed DL

models are further investigated to handle the limited availability of meteorological data required for their reliable estimation. The main contributions of this thesis are as follows: Estimation of reference evapotranspiration using a hybrid deep neural network approach for limited meteorological data.

Estimation of reference evapotranspiration using deep reinforcement learning- based ensemble approach to consider the effect of time-varying characteristics

of the ET0 process.

Estimation of daily crop evapotranspiration using limited climate data and to forecast future changes in wheat and rice ETc. Two-hybrid deep neural network models, i.e., Convolution-Long Short Term Memory (Conv-LSTM) and Convolution Neural Network-LSTM (CNN-LSTM), are

investigated for the estimation of reference evapotranspiration. Conv-LSTM per- forms the convolution operation in LSTM cells and CNN-LSTM uses the convolu- tion layer for feature extraction of input data, and then the extracted features are

fed to LSTM layers. The climate dataset of two stations in India: Ludhiana and Amritsar, is adopted to develop the proposed models. It includes daily maximum temperature (Tmax), minimum temperature (Tmin), wind speed (WSP) measured at the height of 2m, solar radiation (Rs), relative humidity (RH), vapor pressure (VP), and sunshine hours (SSH) data from the period 2003 to 2015 of Ludhiana station and 2000 to 2016 of Amritsar station. The study also focuses on climate data scarcity conditions, and thus, different input combinations of climate parameters have been used to investigate the minimum required parameters to model the ET0 process. Several performance measures are used to assess the precision of the model and to perform sensitivity analysis. Temperature and radiation are observed as the prime data inputs required to estimate ET0 values. The proposed hybrid models are then compared with existing temperature and radiation-based empirical models such as Hargreaves, Makkink, and Ritchie. The comparison reveals that CNN-LSTM and Conv-LSTM outperform these existing models. Moreover, Conv-LSTM performs better than CNN-LSTM.

To investigate the time-varying characteristics of the ET0 process, a deep rein- forcement learning (DRL) based ensemble approach, DeepEvap, is proposed for the

estimation of ET0 values by using three climate parameters as input variables i.e.,

Tmin, Tmax, and Rs. The modeling procedure of the proposed ensemble model con- sists of three phases. In phase I, the data preprocessing technique is performed on

the meteorological data to clean the existing impurities(e.g., outliers, missing data) as it affects the performance of any machine learning (ML) based approach. In phase II, four different deep neural network-based models are used to build the estimation model of ET0 and calculate the prediction results. In phase III, the DRL approach is used to ensemble the prediction results of these four models. The meteorological datasets of two stations of India: Ludhiana and Patiala, are selected to validate the proposed approach. Simulation results show that the proposed DeepEvap approach is competitive for ET0 prediction by achieving a coefficient of determination (R2 )

= 0.96, mean sqaure error (MSE) = 0.0018 mmd−1

, and explained variance score

(EVS) = 0.96. It significantly outperforms four baseline models. Moreover, DeepE- vap also integrates four deep neural network models and works better than existing

stack based and weighted ensemble approaches. vi

Two-hybrid DL models, i.e., Convolution Neural Network-eXtreme Gradient Boosting (CNN-XGB) and Convolution Neural Network-Support Vector Regression (CNN-SVR) are proposed to estimate daily ETc values of wheat and rice crops. Further, limited climate data (Tmin, Tmax, mean temperature (Tmean), and (Rs)) is used for the prediction of ETc values to handle data-scarce situations. Also, the

future climate data obtained using two emission scenarios: Representative Concen- tration Pathways (RCP) 4.5 and RCP 8.5 for the time period 2023-2033, is used to

project changes in ETc. The results demonstrate that the proposed hybrid models provide satisfactory performance with the Nash-Sutcliffe Efficiency (NSE) = 0.95 and 0.976 for rice and wheat ETc values, respectively. The simulation of future data (IPSLCM5A-LR (Institute Pierre-Simon Laplace Circulation Model 5A - Low Resolution) and HadGEM2 (Hadley Centre Global Environmental Model version 2) obtained from MarkSim GCM) reveals the increase in Tmin by 7.03%, 7.33%, and Tmax by 10.5%, 11.5% for RCP 4.5 and RCP 8.5 respectively. Also, an increase in ETc of rice crop has been reported by 20-22% while increment of wheat ETc has been noticed by 3-4%. Thus, using limited climate data, the proposed approach efficiently estimates ETc of wheat and rice crops and could assist water resource managers in achieving agricultural water sustainability.