2.2.1 Prediction of Future Paddy Harvest and Rice Consumption Demand

To maintain a sustainable balance between rice production and demand of a country, it is necessary to implement proper forecasting models to support decision-making processes. Currently, the Government of Sri Lanka is conducting the “the crop cutting survey” to estimate the average yield of paddy in Sri Lanka at the district level, by calculating the means of a sample survey [8]. A sample of 3,000 villagers for the main season “Maha” and 2,000 villagers for the second season “Yala” are considered for this study. This process requires a considerable amount of manpower, resources, and it is time-consuming [8]. Therefore, it is essential to introduce a systematic approach to gather required data and forecast the paddy harvest and rice consumption, so that the supply and demand is well-known in advance.

Several studies have used a machine learning approach to build models to identify these trends on paddy harvest as well as rice demand. The review work of [9] uses machine learning approaches to predict rice crop of India. Their approach used sequential minimal optimization (SMO) as the classifier, adopting the “WEKA” framework for a data set that covers 27 districts in Maharashtra state (India). They have explored a set of parameters: precipitation, min/max/average temperature, crop evaporation, cultivation area, final production, and yield (Kharif season: mid-June to mid-November) from year 1998 to 2002. The authors of [10] have analyzed the trends of paddy harvest of Sri Lanka for past, present, and future and developed a time series model to predict long-term trends (for 3 years). Following their results, the ARIMA (2, 1, 0), was observed as the as the best fitted model to the used data set, with minimum Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC). The authors of [11] have identified one of the main reasons for paddy industry problems; that is, dissatisfaction of farmers, even with the self-sufficiency because of the increase in average yield and extent of cultivation. They have explored several variables which are important for paddy production: extent sown with distinct water suppliers, percentage of crop failure, crop production, min/max/average harvest, etc., to investigate the causes for major problems in paddy industry and used statistical techniques as well as time series analysis.

As observed over past few years, paddy price fluctuation is a huge problem for socioeconomic development, and therefore, it is very crucial to have an accurate approach for rice demand forecast. This will create a sustainable balance in the rice industry and avoid unexpected price raises, especially during rice scarcities and manage the surpluses effectively. The authors of [12] present a platform for information gathering/management and a decision-making process for paddy production targeting decision-making authorities as well as farmers. They considered parameters that control the rice demand and predicted the price that is to be expected for rice in well advance using a systematic approach: a decision tree following classification model. The authors of [13] have studied rice production and consumption from the viewpoint of supply and demand in each region and used population of each region to approximate the rice consumption using simple and multiple regression curves. A study on [14] analyzed of food consumption patterns in Sri Lanka. They have examined the variation of food consumption patterns and economic parameters determining the demand for the period. The analysis has done using ADIS model. They have relieved that food consumption patterns are moving from starch-foods to non-starched foods and that Sri Lanka consumers are more responsive for price changes in rice.

2.2.2 Rice Distribution

The Genetic Algorithm (GA) approach is a well-established algorithmic approach for planning and scheduling [15–23]. However, there are only a limited set of research works related to harvest distribution that have used GAs. The study [24] uses a GA-based approach to identify cost-effective cropping pattern that maximizes profits for an Indian irrigation development project. They have considered a set of constraints: continuity, land/water requirements, crop diversification, and storage restrictions. Their results show that GA approach provides reasonable solutions fast, compared to traditional linear programming approach. The authors of paper [25] discuss a novel approach for the logistics facilities for rice distribution, which is a problem that can be categorized as a Vehicles Routing Problem. They have implemented a Taboo Search–based algorithm and optimize the process of distribution centers based on GA. They have mainly focused on sugar production industry and milled rice distribution centers of Thailand, specially targeting export. They propose a framework to minimize the cost of sugar and milled rice transports from the mills to a seaport. The research work on [26] has used Linear Programming method for the optimization of rice allocation for the poor in Bandung. Their work shows that the proposed model plans the rice distribution with minimum transportation and warehouse cost. The work on [27] explores the design of logistics network, including rice supplier, onsite warehouse, and customer distribution centers in Indonesia: the province of Sulawesi. A mixed ILP model has been used to adopt the problem, to minimize the procurement, inventory, and transportation costs.

2.3.1 Requirements of the Proposed Platform

As the initial stage of developing “isRice” platform, we decided to study the existing paddy harvest as well as rice related processes/methodologies in the world. We initiated the process for the Sri Lankan context and carried out interviews (physical and over the phone) and questionnaires to gather information. We have used a data set that includes 80 farmers and 30 rice mill owners covering three districts of Sri Lanka (Anuradhapura, Kurunegala, and Matara).

During our study, we focused on identifying the feasibility and requirements of a management platform for paddy harvest and rice related processes. We expected three responses: (1) a system is highly required (required), (2) a system is required but tends to rely on traditional methods for predictions and planning (not critical), and (3) a system is not required (not required). The results are summarized at the Table 2.1.

Summarized statistics shows that 56.7% of farmers agreed that a proper system is required; 32.5% of farmers agreed for the system but did not see it as a crucial requirement; 17.5% farmers disagreed with the system; nearly, 50.0% of the rice millers’ owners agreed with the requirement of a system; and 38.9% did not see it as a crucial requirement and only 11.1% has disagreed with the system.

The statistics proved that the management platform for rice related processes is a timely requirement for both farmers as well as for rice mill owners to keep the sustainability between supply and demand and ensure food security and enhance farmers as well as consumers standards of living.

2.3.2 Data to Evaluate the ‘isRice” Platform

For the training as well as evaluation of the platform, we required historical data about paddy and rice related processes, specifically data on paddy harvest and demands. Since it is a challenging task to gather data from entire country, we decided to select few districts in Sri Lanka that represents the paddy harvesting communities of Sri Lanka.

As observed through our survey, “rainfall” was identified as the major parameter that affects the rice production in Sri Lanka. Sri Lanka can be divided into three climate zones: DZ-Dry Zone, IZ-Intermediate Zone, and WZ-Wet Zone, based on the average annual rainfall. Therefore, we conducted the study for three districts: Kurunegala (covering IZ), Anuradhapura (covering DZ), and Matara (covering WZ).

We gathered paddy harvest and demands statistics for the Anuradhapura, Kurunegala, and Matara districts for year 1990 to 2017, from different sources as follows:

• Paddy cultivation statistics (Yala and Maha seasons): Agriculture Department
• Paddy harvest purchasing statistics: Paddy Marketing Board (PMB) and Department of Census and Statistics.
• Producer’s prices for paddy: Sri Lanka Central Bank.
• Factors affecting annual paddy harvest: Rice Research and Development Institute.
• Factors actors affecting paddy harvesting practices and purchasing prices in different regions: Regional farming centers.
• Consumer behavior factors affecting distribution and transportation: Rice mills owners.

2.3.3 Implementation of Prediction Modules

The two main prediction modules of “isRice” management platform were developed using RNN and LSTM models.

2.3.3.1 Recurrent Neural Network

The RNN is a widely used field of artificial neural networks. In a RNN, connections between each network unit are established sequentially following directed graphs [4]. RNNs have hidden units (state vector) to keep the history of all past elements sequentially and process input sequence accordingly, one element at a time. Therefore, RNNs have the capability to identify the dynamic temporal behaviors of time sequences. With RNN, each element of the sequence will be allocated to the same task, and each node’s output is directly affected by the previous nodes computational value. As RNN utilizes the memory (internal state) to handle the input sequences that are connected, during the training itself, RNN keeps records on relationship of these sequences. Therefore, with the help of these relationship records of all previous inputs, RNN predicts an accurate output. Handwriting recognition, speech recognition, etc., are among the applications that use RNNs to achieve better results.

The structure of RNN is shown in Figure 2.2.

S (a neural network block) takes I_t (the input) and outputs O_t (the output). A loop is used to pass the information passed from step x to step (x + 1). A RNN can be thought as a collection of n copies of the network a, and each network forwards an output to network a + 1 which is the successor (as output feedback).

An unfolded RNN loop is shown in Figure 2.3, where the chain structure detones that RNNs are generally applied as sequences (and lists).

2.3.3.2 Long Short-Term Memory

The LSTM concept was introduced in 1997 by research Hochreiter and Schidhuber, and they are a special class of RNNs which have the capacity to learn long-term dependencies [5]. Generally, RNNs have a chain of neural network modules, which are repeated, and each repeated module has one tanh layer (as shown by Figure 2.4).

Following the same, LSTMs follow the structure of a chain, with a different structure for repeating modules. As for the LSTM network, a module generally comprises of four neural network layers and three gates (as shown in Figure 2.5).

The first gate (forget gate) decides on information that should be removed from the cell state. Considering prior cell output (O_t−1) and present cell input (I_t), the sigmoid activation layer (σ) is applied for each hidden unit, to get a value between 0 and 1. If the output is a “1”, then network will keep the information whicle if the output is “1”, network will discard the information.

The next gate (update gate) passes O_t−1 and I_t (the same input), to two activation layers: sigmoid (σ) and tanh (tanh). An element-wise multiplication is performed to two outputs of the respective layers and the result r is obtained. Finally, an elment wise addition is performed to output of “forget gate” and result r. The new indformation is used to update the state.

The third gate (output gate) has the responsible of deciding on the information that will be forwarded to the following state. The layer n’s output, works as the filter that decides the output O_t, that is to be passed to the next LSTM cell.

2.3.4 Implementation of Rice Distribution Planning Module

The main objective of rice distribution planning module was to optimize the paddy harvest distribution considering facts such as supply, demand, transportation cost, and storage cost. We have selected four districts of Sri Lanka (Anuradhapura, Polonnaruwa, Kurunegala, and Mathara) as the main rice supplying districts and Colombo, Gampaha, and Kalutara as three main consuming districts.

Traditionally, achieving an optimal plan or an optimal schedule is modeled as an optimization of ILP, but the process suffers from the inherent characteristic as a NP-complete problem. As such, obtaining solutions for limited, small scenarios might take long time [6]. Hence, for dynamic natured optimization problems, the ILP calculation is considered to be unfeasible, although we can expect an exact solution though an ILP formalization of the problem. With the dynamic nature of rice distribution process, use of an approximation approach (heuristic-based approach) would be better, when planning the rice distribution. Therefore, the distribution panning module is developed a heuristic-based optimization approaches: GP [6, 7].

The rice distribution planning module was formulated as a minimization problem, which tries to minimize profit loss due to surplus, profit loss due to deficit, transportation cost, storage cost, etc. Also, we have utilized constraints considering supply availabilities, storage capacities, preferences of consumer for different rice varieties, etc.

Optimum distribution plan can be derived as the optimization problem given in Equation (2.1):

Minimize:

(2.1)

Where,

2.3.4.1 Genetic Algorithm–Based Rice Distribution Planning

GAs are a subset of evolutionary computing algorithms, which were introduced as a computational analogy of adaptive systems [6, 7]. They are loosely modeled according to natural evolution principles via selection, while having a population of individuals that face the selection process in the presence of different variations, provoking genetic operators: mutation and crossover. Each individual is evaluated using a fitness function, and reproductive success depends on the fitness.

GAs follows five key steps [6, 7]:

1. Generation of the initial population P(0) with x solutions.
2. Computation of fitness g(p) for every individual solution p in the present population P(t).
3. Generation of the following population P(t+1), by selecting y fittest solutions from P(t).
4. Application of the genetic operators such as mutation/crossover to population P(t + 1) to produce of offspring.
5. Repetition from Step 2 until a termination condition is satisfied.

As shown in Figure 2.6, a solution (optimal rice distribution plan), which is also known as the chromosome is encoded. The n number of chromosomes: CA1, CA2, CA3, and CA4, constructs the solution pool: population.

In a chromosome, each gene denotes two main informations: (i) gene index number represents the consumer district and (ii) value in the gene indicates the allocated rice harvest from each supply district. The value of the first gene in first chromosome (CA1) is {100, 200, 300}, demonstrating rice units 100, 200, and 300 have been allocated to first consumer district from first, second, and third supply districts, respectively.

Each of the chromosome in the population is assessed according to a derived fitness function: Equation (2.2).

(2.2)

We have executed mutation and crossover: two genetic operators that produce offspring. The crossover operator is expected to bring in the population to a local min/max and considered as a convergence operation. The mutation operator is expected to pull off one (or multiple) solutions of a population from a local min/max occasionally and explore an improved space. Therefore, the mutation is considered as a divergence operation. Each i^th generation of GA approach follows mutations/crossovers.

In the crossover process, as demonstrated in Figure 2.2, we employed the single-point crossover technique to generate an offspring. Initially, we have selected a random point from two chromosomes (CA1 and CA2) and divide each chromosome into two parts. The offspring (CA3) is then created by copying the first part from chromosome 1, and second part from chromosome 2. As demonstrated in Figure 2.7, gene 7 (100, 200, 300) and 2 (400, 600,800) are copied from chromosome 1 (CA1). The gene 3 (400, 900, 140) is copied from chromosome 2 (CA2) to generate the offspring (CA3).

As the mutation technique, we have used is the swap method, and it shown in the Figure 2.8. Initially, we have selected two genes from a randomly selected chromosome (CA1). The offspring (CA4) is then created by swapping the values of these two genes. The newly generated offspring (CA4) consists of the swapped genes values.

The freshly generated solutions are assessed according to the derived fitness function. The generation-wise process is repeated until a condition is satisfied: either n generations are executed, or the optimal solution (with the minimum fitness value) is found.

2.3.5 Front-End Implementation

As shown in Figure 2.9, the front-end of the “isRice” system is implemented in a user-friendly manner, as a combination of an android platform and a web platform. The android application is for farmers, rice mill owners and consumers. The web platform is for Paddy Market Board and Rice Research and Development Center.

Figures 2.10 to 2.12 show main interfaces of the “isRice” system.

2.4.1 Paddy Harvest Prediction Function

The summary for the paddy harvest predicting function is as follows:

• Harvest prediction LSTM network: 200 iterations.
• Performance evaluation: Mean Squared Error (MSE) and RMSE for both training score as well as test score of the models.
• 78% for train score with MSE = 0.04.
• 75% test score with MSE = 0.11.

The MAE for train set as well as test set for paddy harvest prediction function is shown in Figure 2.13.

2.4.2 Rice Demand Prediction Function

The summary for rice demand predicting results is as follows:

• Harvest prediction LSTM network: 90 iterations.
• Performance evaluation: MSE and RMSE for both training score as well as test score of the models.
• 79% for train score with MSE = 0.17.
• 74% test score with MSE = 0.37.

The MAE for train set as well as test set for harvest prediction module is shown in Figure 2.14.

2.4.3 Rice Distribution Planning Module

As the evaluation of rice distribution planning module, we are summarizing the GA approach results, which was used to generate an optimal plan for rice distribution process. As the genetic operators, single-point crossover with two-points swapping mutation were used.

The key terms used for our experiments are as follows: g (number of generations), Fbest (best fitness value calculated over g generations), AVG (average Fbest), MIN (minimum Fbest), and STD (standard deviation of Fbest), Pm (mutation rate), and P (population size).

As explained in previous sections, the genetic operators: mutation and crossover are utilized to enhance the given initial solution, over n generations. To observe the improvements granted by genetic operators, we performed 30 experiment rounds, where we started with a set of initial solutions, then enhance them by applying GA steps. Since we formulated our problem as a minimization problem, the fitness value needed to be decreased for the solution to be improved. As demonstrated in Figure 2.15, the majority of the enhancements in the fitness function (decrease of the fitness value) were observed during first 200 generations: very early stages of the process.

To identify how the proposed algorithm scales with the increase in number of consumer districts, experiment rounds were performed. We increased the number of consumer districts, to simulate different scenarios and executed the GA with a population with the size of 20 over 200 of generations. Figure 2.16 demonstrates the computational time for different scenarios, as number of consumer districts increases.

We have evaluated our proposed GA to explore an effective mutation rate, to obtain the optimal solutions fast. We have assumed a population with the size of 20 and number of consumer districts are 3. The algorithm was executed for 200 generations. Mutation probability was kept as 0.15 (constant) while the mutation rate was varied. As summarized in Table 2.6, improved solutions are obtained when the mutation rate is low: 0.002: the lowest mutation rate, produced the best fitness value, within a shorter time.

Next, we have evaluated our proposed GA to explore an effective mutation probability, to obtain the optimal solutions fast. We have assumed a population with the size of 20 and 3 consumer districts. The GA is executed for 200 generations. Mutation rate was kept as 0.002 (constant) while the mutation probability was varied. As summarized in Table 2.7, 0.1 mutation probability produced the best average fitness value.