Decision Support System for On-Farm Crop Water Irrigation Scheduling using Machine Learning approaches

Abstract

In India, irrigation is the largest consumer of fresh water and is drawing about 90% of groundwater. The requirement of irrigation system is much needed for the region such as Central Punjab, which occupies nearly 97.95% of the gross irrigated area for agricultural production. The water consumption is very high in commonly cultivated crops in Punjab that are Wheat, Maize, and Rice. This requires modern technologies in water management to meet the agricultural challenges. Hence, this system referred as Irrigation Water Management (IWM), is poised to be a key driver of smart farming to meet crop water requirement with a sufficient economic return without any damage to land and soil. The major challenge in agriculture sustainability and dawdling is to utilize every drop of fresh water effectively and efficiently. The studies on water shortage suggest the development of innovative irrigation methods such as controlled deficit irrigation, partial root drying, and continuous deficit irrigation. In this context, the controlled irrigation, climate, soil fertility, crop quality, and time management are essential to the Decision Support System (DSS) to maximize the crop yield with minimum consumption of water. Advanced Analytics and DSS can help farm managers in taking decision to solve complex irrigation problem. The Reference Evapotranspiration (ETo) is one of the most valuable parameters for hydrological, climatologist investigation, and water resources management. An exact estimate of ETo is necessary to analyze the water demand of irrigated agriculture, crop-water balance and improve the water quality. However, ETo estimation is very difficult to achieve due to its dependency on many input parameters. Therefore, the primary objective of the research is to establish regression models for the estimation of ETo with limited climate parameters. The estimation of daily ETo can help in real-time prediction of crop evapotranspiration and crop irrigation demand. The framework of ensemble-based modeling has been designed in this work using Machine Learning (ML) and Deep Learning (DL) models for estimating the Reference Evapotranspiration ETo, Crop Evapotranspiration ETc, and thus meet crop water requirement in Irrigation Scheduling (IS). The work is carried out with the following objectives: . Estimation of Reference Evapotranspiration (ETo) using H2O framework based on Deep Learning-Multilayer Perceptron’s (DL), Generalized Linear Model (GLM), Random Forest(RF), and Gradient-Boosting Machine (GBM) for classification and regression purposes. . Estimation of single Crop Coefficient (Kc) and Crop Evapotranspiration (ETc) using a novel multilevel ensemble model based on Fuzzy-Genetic (FG) and Regularization Random Forest (RRF) models for Wheat and Maize crops. . Investigations on DSS based crop water requirement, Net Irrigation, Gross Irrigation, and Pumping Time in Irrigation Scheduling using Particle-Swarm optimization with Deep Neural Network (PSO-DNN), and Deep Learning models. The H2O model framework is investigated to determine the daily Reference Evapotranspiration (ETo) for the Hoshiarpur and Patiala districts of Punjab. Daily Meteorological dataset is collected from Indian Meteorological Department (IMD), Pune having combination of six input variables i.e. Tmin, Tmax, RH, u2, Is and Rs. The appropriate missing values have been filled using MissForest algorithms. Four supervised learning algorithms have been investigated to forecast ETo. These are Deep Learning-Multi-Layer Perceptron (DL), Generalized Linear Model (GLM), Random Forest (RF), and Gradient-Boosting Machine (GBM). The FAO-56 Penman- Monteith method is used to calculate the Rs and ETo. A three-layer Deep Learning model with Rectified Linear Unit (ReLU) function and Stochastic Gradient Descent (SGD) via Backpropagation have been developed to obtain the minimum prediction error. The multinomial classification as cross-entropy function has been applied for calculating the loss function. The effectiveness of the developed model is tested using performance metrics to obtain the accurate estimation of ETo using classification and regression purposes. A three-layer multi- model ensemble machine learning approach has been investigated to predict ETo for regression. The three supervised smachine learning models: Extreme Machine Learning (ELM1, ELM2, ELM3, ELM4), Multi-Layer Perceptron’s Neural Network (MLP1, MLP2, MLP3, MLP4), Support Vector Machine (SVM1, SVM2, SVM3, and, SVM4) models have been applied to investigate the abilities and applicability of the ensemble-based model. The investigations on estimation of Crop Coefficient (Kc), and Crop Evapotranspiration (ETc) are carried out using ensemble-based modeling. Two algorithms namely Fuzzy-Genetic (FG), and Regularized Random Forest (RRF) are applied to develop the FG-RRF based (ETc) framework for three crops, namely (Maize, Wheat1, and Wheat2). Fuzzy-Genetic model is applied to simulate the Kc and ETc values using a training dataset. The Kc and ETc prediction probabilities are combined in dataset. Then, the ensembling dataset is used to train the RRF model for predicting the ETc values of each sample of the crop. After getting the best accuracy from a training model, the testing dataset is applied to validate the accuracy of the model. The proposed model is evaluated based on performance metrics to check the accuracy of results. The effectiveness of the developed model (FG-RRF based ETc) is tested and compared with the SVM model including (Sigma and C) parameters. The ensemble-based model is proposed to estimate the DSS-IS using Particle Swarm Optimization with Deep Neural Neural (PSO-DNN), and Deep Learning (DL). The fuzzy forest algorithm is applied for the selection of the best ten input features. The algorithm is based on random forest and designed to reduce and rank the important number of features in regression. These features are chosen based on a feature recursive exclusion function. Using these models, a DSS is developed with improved accuracy by reducing the number of features. The effectiveness of the developed model DSS-IS using PSO-DNN and DL is tested and compared with three samples of irrigation parameters. A decision support system can help the user not only in estimating the ETo but also in selecting the best ( ETo) and ( ETc) (and intermediate parameters). The enhancement of water efficiency requires controlling the high demand for irrigated agriculture, which improves the capabilities to simulate the water cycle and its components accurately. The ensemble learning can enhance the effectiveness of classifiers by blending their decisions individually. A combination of Daily average temperature and solar radiation is the optimal combination for the ETo and ETc estimation. Ensemble models showed great applicability in modeling ETo, and can be highly recommended for estimating (ETc) in Punjab as well as other stations.