Decision Support Systems An Overview

2.1.2 Machine learning algorithms for wheat yield prediction
Decision support system (DSS): A DSS is a system that provides information and maintains the administrative activities of a business or organization. A DSS addresses organizational administration, operations, and development levels to help in making the decisions about quickly changing issues that may not be easily identified in advance. A DSS can be fully automatic, human-powered, or a combination of each. A DSS has inputs, such as factors, numbers, and characteristics for analyzing user information and proficiency. Users need to analyze the inputs and outputs manually. Transformed data produces a DSS “decision.” Outcomes produced by DSS are built on specific criteria. DSSs that bring out certain cognitive decision-making functions are assembled with artificial intelligence (AI) or intelligent agent technology and known as intelligent DSS. The DSS for harvesting high superiority wheat holds the idea of the united method in making multidisciplinary DSSs. This DSS is taken into account and provides selection provision for all key elements of the production chain, from strategic selection to tactical operations (Rossi et al., 2010).

Regression method: Regression evaluation is a hard and fast statistical technique for estimating the relationships among a based variable and one or more independent variables. If there is no reasonable dependency among variables, one can attempt mathematical equations in order to find a link between them (Niedbała, 2018). The maximum common regression evaluation is linear regression in which a researcher reveals the line that most intently fits the statistics in step with a precise mathematical criterion. Regression evaluation is normally used for two conceptual functions. Initially, regression evaluation is broadly used for predicting and forecasting, where its use has considerable overlap with the field of ML. Then, in some conditions, regression analysis can be used to conclude a fundamental relationship between the independent and dependent variables.

Random forest (RF): Random decision forest is a collaborative learning method that works by constructing a huge number of decision trees during training and outputting classes that are the modes of sorting and grouping or mean prediction (regression), for classification, regression, and other tasks. RF corrects the tendency of overfitting a training set of decision trees. The RF classifier is a collective method that trains numerous decision trees similar with bootstrap, which is collectively called bagging, and subsequent aggregation. RF is an effective and flexible ML technique for crop yield predictions on both a regional and a global scale for its excessive accuracy and exactness, accessibility, and effectiveness in facts analysis. For classification and regression purpose, RF can be used, and when needed, it can be used as regression model also (Jeong et al., 2016).

Support vector machine (SVM): SVM is a set of ML rules developed by way of Vapnik and is primarily based at the principles of statistical learning theory. SVM uses an introduced feature of structural and experimental threat reduction. It has the capability to do the mapping of the functions in high dimensional space though translating the difficult problems to a linearly separable event (Kumar et al., 2019). The determination of the SVM algorithm is to find that the RF classifier (i.e., ensemble method) that trains numerous decision trees in parallel with bootstrap. If you have a set of training samples, each one is marked as association to one of two groupings and the SVM algorithm generates a model that allocates the new example to one of the groups and a nonstochastic binary linear classification. The SVM model represents the examples as points in space, that is, it maps the individual categories into the widest possible gaps. The new examples are then mapped to that identical area and expected to fit to a group primarily built on the aspect of the distance on which they fall.

Neural network: A neural network includes neurons, organized in layers, which translate an input vector into an output vector. Input is taken at each unit after applying numerous features on it, and at that moment passes output to the next layer. Artificial neural networks (ANNs) are extraordinarily crude electronic networks of neurons created on the basis of neural structure of the brain. They learn by analyzing the records one by one and comparing the record’s classification with the identified actual classification of the document. The faults from the preliminary class of the first record is fed lower back into the network and conditioned to adjust to the network’s set of rules. A group of input values (xi) and related weights (wi) and a characteristic function (g) does the summation of weights and draws the result to the output (y). A neural community includes neurons organized in layers and transforms an enter vector into several outputs. Every unit proceeds with an input, puts a function to it, and passes its output to the succeeding layer (Training an Artificial Neural Network, 2020).

Multilayer perceptron neural network (MLP): MLP is a category of feed-forward ANN, or networks consist of several layers of perceptron. MLP are as often colloquially known as “vanilla” neural networks, specifically when they are having a hidden layer. An MLP contains an input layer, a hidden layer, and an output layer. Through the exclusion of input nodes, every node is a neuron that uses a nonlinear activation function. The input layer gets the parameters to control the neurons of the hidden layer(s) and the output layer method and the weighted indicators from the neurons of its preceding layer and calculate an output cost, making use of an activation feature (Kross et al., 2018). MLP makes use of a supervised learning method referred to as backpropagation (BP) for training. It is a combination of layers and nonlinear activation that differentiate MLP from a linear perceptron and can make a distinction of data that is not linearly separable.

Adaptive network-based fuzzy inference system (ANFIS): ANFIS refers to synthetic neural community based on the Takagi-Sugeno fuzzy inference device. By integrating neural networks in addition with fuzzy logic principles, it is possible to combine the advantages of both into a particular framework. The inference system supports a series of fuzzy IF-THEN rules with a learning function that approximates nonlinear functions. One can identify two parts of the network structure: the premise and consequence. In detail, the architecture consists of five layers. The first layer contains input fuzzy rules, the second layer contains input membership functions, the third layer contains fuzzy neurons, the fourth layer contains output membership functions, and the fifth layer contains a summation of all operations (Rusgiyono, 2019).

Self-organizing map (SOM): A SOM is a form of ANN that uses unsupervised learning to provide a two-dimensional discretized view of the input space of a training sample called a map, and it is a way to reduce dimensions. Unlike other synthetic neural networks, SOMs follow competitive learning rather than error-correcting learning, so we have experience using neighborhood features to hold topological assets in the input space. It is common to think of this type of network structure associated to a feed-forward network in which the nodes are imagined as connected, but this kind of structure varies in arrangement and motivation. The SOM models involve input nodes demonstrating the principle features in wheat crop manufacturing, including biomass signs, organic carbon (OC), pH, Mg, Total N, Ca, cation exchange capacity, moisture content (MC), and the output weights characterized the class labels similar to the anticipated wheat yield (Pantazi et al., 2014).

Supervised Kohonen networks (SKNs): SKN models are supervised neural networks, rising from SOMs used for sorting and grouping. In the case of SKNs, the SOM and output layers are amassed collectively to provide a joint layer trained in keeping within the regime of SOMs. In the SKN network, the input map Xmap and the output map Ymap are “combined” to form the joint input/output map (XYmap) as a result of the unsupervised Kohonen network training scheme. (Melssen et al., 2006) (Table 3).

Table 3. Wheat yield prediction summary by machine learning algorithms.

Technique usedMeritsDemeritsConclusionFuture scope•DSS

(Timsina et al., 2008)Estimated yield forecast using:
Climatically driven potential yield•From water balance component

•DSSAT software showed how the yield prediction can be enhanced with increases in CWP and IWP

•Has certain assumptions

•Input parameters have some uncertainty

•Dynamic model simulates crop growth and yield prediction

•Throughout the planting period sowing can be done

•On the basis of atmospheric demand stimulus can be applied

•Effect of weed and pests were not included in the input parameters

•Can add them as they also effect the yield prediction

•Regression method

•RF

•SVM

•Neural network

(Dadhwal et al., 2003)•Predict wheat yield before two months of their maturity

•Assumptions in the results presented

•Uncertainty in the input parameters

•Used static wheat

•growing areas that leads to error

•Input EVI perform better yield prediction that SIF

•Mix of satellite data and climate provides high-performance yield forecast

•Soil information can be taken in order to increase yield production

•MLP

(Bhojani & Bhatt, 2020)•Newly generated algorithms of MLP proves improved output using low RMSE and RAPE

•Recommendation of activation functions for small network structure only

•Activation function provides the results with more accuracy

•DharaSig, DharaSigm, and SHBSig, activation functions was created

•Increase the performance of neural network

•Can also add weather dataset and soil dataset for crop yield prediction

•Regression model

•SVM

•RF

•Neural network

(Cai et al., 2019)•EVI

•achieved better performance than SIF

•Combining climate data with satellite data provides high-performance yield forecasts

•Climate data

•cannot be captured by satellite data, so unable to get 100% accurate result

•Improvement of vegetation index from MODIS and solar-induced chlorophyll fluorescence

•Can add soil factor with temperature, water, and satellite data

•ANN

•Multilayer perceptron

(Kadir et al., 2014)•MLP

•Networks have been proven to be effective with linear and nonlinear data

•Pesticides effect

•Soil conditions

•Diseases

Affect wheat yield but these parameters were not considered•Seven input parameters were there

•MLP could predict wheat yield with 98% accuracy

•Can reduce the input parameter to increase efficiency

•ANFIS

•Used to simulate the hydration properties of wheat grain

(Shafaei et al., 2016)•The simple structure ANN simulation framework was easier to use as compared with the three different structures of ANFIS

•To attain higher moisture content

•Usage of longer hydration times instead of higher hydration temperatures

•Water absorption rate is drastically increased with increasing hydration time with temperature

•Higher hydration temperatures can be used with this procedure to measure water content in a short time.

•Supervised SOM and crop sensors

•Were used to predict soil properties for yield prediction

(Pantazi et al., 2016)•Reduced labor

•Time costs required for soil sampling and analysis

•The output shows that cross-validation-based yield predictions for the low-yield class SKN model surpassed 91%

•Being unable to model continuous output relationship

•The resulting nodes consisted of predicted yield equal frequency classes from three trained networks such as CP-ANN, XY-F, and SKN

•Proposed architecture can be enhanced with smooth interpolating kernels to deals with inability of continuous output kernel