Modeling Artificial Neural Network of Insect’s Proliferation During Cocoa Beans Storage

The preservation of grain is an important element for storage organizations to ensure the permanent supply of markets. Thus, maintaining full quality during storage remains a major challenge because the grains to be marketed must comply with the requirements and standards that govern these markets. Compliance with these standards is of great interest to the scientific community [1]. This challenge is made more complex by climate, environment and/or biotic factors that can jeopardize conservation. There is a degradation of the initial quality due in part to attacks by insects. These seed-eating insects are the main causes of health and economic damage [2]. This damage can reach 30% of gross production if no action is taken [3, 4]. Therefore, it remains essential to improve conservation techniques, which makes it possible to reduce post-harvest losses [5]. According to Siaho et al. [6] the conservation of cocoa beans is no exception; it also faces these same difficulties. To overcome the phenomenon of infestation, several works have been proposed. Concerning cocoa, active studies on controlled and modified atmospheres as well as the effects of packaging on beans have been carried out to extend their shelf life. However, these methods are necessary and difficult to implement. Moreover, in the field of cocoa, no model of insect dynamics has yet been proposed to our knowledge. This contribution aims to propose a new approach for predicting the evolution of insects based on artificial neural networks.

1.1 Insect pests of stored

The presence of live insects in stocks represents an obstacle to the marketing of the grains because they damage and compromise the initial quality of the stored product. 50% of damage to cocoa beans occurs following attacks by insects such as the coleoptera Tribolium castaneum (tenebrionidae), causing economic losses estimated at 3.16 billion US dollars [7]. Thus, the provision of improved storage systems makes it possible in most cases to better protect the stored product and therefore limit losses [8]. To stay alive, insects need food, air, and water. Stored beans are favorable places for the residence and development of insects [8]. The bean includes a set of nutrients such as sugars, free amino acids, or water proteins [9, 10]. It should be noted that poorly preserved grain will spontaneously heat up, which initially favors the development of insects. This is the mechanism of infestation. During this mechanism, the insects consume the grain and cause an increase in temperature due to their metabolic activities which accelerate the heating of the beans. When this warming reaches a certain threshold, we see the appearance, in the presence of humidity, of the mechanism of mold growth which is also harmful to the quality of the cocoa beans [6, 11]. There are two types of storage insects according to their impacts on the grains (beans): primary insects and secondary insects. The primary insects are composed of very small beetles (weevils, beetles, bruchids); certain Lepidoptera and butterflies (Ephestia cautella, Araecerus faciculatus, Lasioderma serricorne), whose biological activity depreciates the food and whose larvae are harmful because they eat the inside of the grains [12]. Most species develop at temperatures of 18℃ <T° <40℃ [5]. This biological activity not only compromises the quality and commercial value of the stored beans, but also promotes the development of microorganisms. As for the secondary insects, they develop outside the grains, their larvae and pupae are therefore visible in an infested stock. Also referred to as free-form insects only. Thus, the morphology and development cycle of certain insects make their detection and control complex [13]. Therefore, a good understanding of the grain ecosystem as well as insect dynamics is of great importance to guarantee the volume and quality of storage. Several studies have been conducted to eradicate or to predict the abundance of insects during conservation.

1.2 Methods of insect control

Plant protection products are commonly used to control insects. However, in a few years, the number of these products has been considerably reduced [14], and increasingly oriented towards the exclusion of their use. The use of alternative methods then becomes an obligation because the exclusive use of insecticide can lead to the appearance of resistant insects [15]. Another method is the placement of insect traps above the stocks. Regular monitoring of these traps makes it possible to detect the presence of insects early and to implement curative methods if the infestation is minimal. However, these curative control methods still consist in most cases of treating the grains with a contact insecticide.

Fumigation is also an interesting technique for implementing an insecticide gas in a cell or hutch, but it has human and environmental disadvantages [16]. Environmental damage, human health problems and the emergence of resistance are pushing scientists to seek alternatives to synthetic insecticides. This is a first step towards developing a new model for alternative stored product pest management strategies [17]. Despite the large number of possible control techniques and the diversity of products used, the range of measures applicable by farmers and storage organizations is still limited. Also given the fact that no single technique gives on its own, complete satisfaction, we are moving more and more towards their judicious and reasoned combination in order to achieve an increasingly effective fight [18]. This search for alternative solutions to these treatments has been productive with several integrated protection approaches. The components of integrated protection start from an identification of insect species to storage forecasting methods and monitoring tools considering the concept of harmfulness threshold to develop technical storage itineraries and alternative methods.

Several integrated methods approaches have been proposed and validated by projects such as the QualiGrain, EcopréservGrain and EcoprotectGrain Qualigrain projects [13]. Among these current practices, Integrated Pest Protection (PAI). Although the PAI seems to be the most demanding among the others and applied to cereals according to the principles of organic farming and the wheat charter. However, the PAI is slow to be put into common practice for the protection of conventional culture stocks, those intended to supply the mass export markets [19]. Similarly, insect prediction models are empirical models based on the history (development time, number of eggs laid) of insect development as a function of several parameters, mainly temperature and grain moisture. These experimental models can be presented as mathematical equations resulting from nonlinear regressions, relating the parameters influencing the temporal dynamics of insect development [11]. However, these models are associated with very expensive infestation detection techniques [20, 21].

Recently, some authors such as Banga et al. [22] have proposed using techniques from artificial intelligence and acoustics, in this case artificial neural networks, for the prediction of the abundance of insects in chickpeas. According to them, the problem of predicting insects could be linked to a lack of appropriate method of detecting insects, preferably at an early stage of their growth cycle, to take preventive measures at the right time. These researchers have shown that the use of this technique could improve prediction results compared to regression methods by relying on the performance of detection methods. Some work like that done [22] has shown that modern engineering and acoustics have been incorporated into decision support systems for stored product insect management, but with somewhat limited due to device costs [23] and the skills needed to interpret the data collected. Thus, Mankin et al. [22] tested machine learning (ML) to describe the movement of adults and larvae of Sitophilus oryzae (Coleoptera: Curculionidae) and their feeding in grains,

Lutz and Coradi [24] proposed to improve this method by evaluating ongoing studies on the use of sensors in stored grains, the precepts of the Internet of Things (IoT) and technologies for machine learning (ML) in post-harvest to identify resources and new opportunities for technological development that can help in the quality control of stored grain. It shows that the results collected in real time allow better decision-making in situations where there is a risk of loss and deterioration of the stored grains due to variations in factors such as water content, temperature of the mass of the grains and the relative humidity of the intergranular air. Thus, the quality of stored grain can be predicted by monitoring moisture content. However, this work only considered the contribution of grain moisture content and temperature or moisture content and CO₂. To extend the effectiveness of the prediction, it would be necessary to associate it with the sugar content which, according to the study [25], showed that the sugar content is one of the key factors to ensure the functions of maintenance and reproduction. Because sugars play an essential role as building blocks for insects, but also as a source of energy.

1.3 Neural network

Neural networks (ANN) are a model test to approximate the functioning of neurons in the human brain. An artificial neuron is a mathematical operator, whose value depends on parameters called coefficients or weights [26]. The variables of this function are usually called the "inputs" of the neuron and the value of the function is called its "output". The neuron generates the output by applying the previously chosen activation function to a linear combination of the inputs. More concretely, a neuron is composed of an input layer accepting input variables denoted x with a bias, weight w, an activation function denoted here σ and one or more values output y. We thus have the following formula [27]:

$y=\sigma\left(w_0+\sum_{j=1}^p w_j x_j\right)$                  (1)

Neural networks are involved in several fields relating to optimization [28, 29], prediction [30, 31], image recognition [32], signal processing [33, 34], diagnosis [35]. The structure of neural networks, adapted to the non-linear problem, promotes the success of its use. Integrating new technologies such as ANN will allow an ever-increasing development of the agricultural sector, especially post-harvest handling [21, 22, 24, 36]. To this end, according to the study [37], given the complexity of the ecosystem of stored products, ANNs are of considerable contribution in the prediction of phenomena related to this environment. In post-harvest handling, different types of neural networks have been used to make predictions [38]. There are several types of neural networks such as multilayer, recurrent, and deep networks. Among these, feed-forward (FFN) and recurrent network (RNN) RNAs have been the most used to describe drying kinetics in cocoa. Thus in 2015, the authors Nogbou et al. showed that recurrent neural networks have an ability to predict water content during microwave drying. The performance was possible with an optimal 3-10-1 neural architecture (3 neurons on the input layer, 10 neurons on the hidden layer and one neuron on the output layer) [39].

The difference between FFNs and RNNs lies in the structure of the network. There is at least one cycle in the structure of RNNs unlike FFNs whose interconnection is from input to output. The network training techniques are the same as for conventional networks (gradient backpropagation). RNNs make it possible to predict the dynamics of a population. However, an inherent problem in the use of RNNs is the difficulty of retaining contextual information over large time intervals. This phenomenon is due according to Morillot et al. to the fading of the gradient [40]. A model of the evolution of the water content of cocoa beans was made using a multilayer perceptron in the context of drying [27]. To better understand solar drying, a modeling was carried out. Thus, the 2-2-1 architecture (two neurons on the input layer, two neurons on the hidden layer and one neuron on the output layer) was best suited to make this prediction.

Therefore, ANN could make a significantly effective prediction on insect density in cocoa beans to reduce post-harvest losses. The literature presents very few studies on the prediction of insects concerning stored cocoa beans. Also, the different methods for predicting bio aggressors have not taken advantage of the sugar content parameter as indicated [6]. Especially since the sugar content plays an essential role in the maintenance and reproduction functions of insects [25]. Thus, in this work, we propose a neural network-based approach that merges not only water content with integration of sugar content, but also temperature, to improve the prediction of insect density.

3.1 Selected topology

The configuration of our topology depends on both the number of hidden layers and the number of neurons on each layer.

Choice of the number of hidden layers: The ability of the network to translate coexisting relationships depends on its architecture [31]; moreover, when constructing the neural network, it is essential to determine an appropriate network topology. According to the study [31], the architecture of networks containing a structure with a hidden layer seems to perform better for nonlinear regression models of biological systems [52]. In this work, we then adopt a neural network with a single hidden layer for modeling the evolution of insect density in stored cocoa bean ecosystems.

Choice of the number of neurons on the hidden layer (NNCC): To obtain an optimal architecture, we varied the NNCC from 1 to 15. To obtain the value of which provides the good architecture, we will proceed in two stages. The first step was made based on the correlation coefficient obtained for each NNCC and the second step was made based on the AIC. Here, the order is important because the AIC does not have the ability to certify that a model is good or not. Therefore, it is necessary to first confirm the validity and reliability of the model starting with the correlation coefficient.

In Table 1, the correlation coefficient from rank k≥3 to order 2 provides significantly better results. It therefore seems appropriate to us to choose the k from rank 3. However, the choice of k one with large values necessarily implies an increase in the calculation time of the ANN model. About Table 1, from rank 5 to order 3 the values of the correlation coefficient are practically identical. Thus, one can choose k$\in$[3;5] without however losing the idea of the choice of the optimal architecture. However, a problem arises for the choice of the value of k.

Table 2. Values of the correlation coefficient during training

The second step was done using AIC. Table 2 presents the calculated AIC values of each model relative to the number of neurons in the hidden layer. For each of the architectures, we evaluate the AIC using parameters such as the number of weights and number of observations (n=266) in our dataset. The architecture with the minimum AIC value is selected to model the evolution of insect density in cocoa beans. From Table 3, we can observe that, the architecture with neurons in the hidden layer has the very smallest value of AIC (-795.451903). This explains that with this architecture, the loss of information is much less compared to other architectures relating to each AIC values. According to the study [51] the less information a model loses, the higher the quality of this model.

Table 3. AIC values

Thus, the 4-5-1 architecture could best fit our dataset. The 4-5-1 architecture, four (04) inputs, five (05) neurons in the hidden layer and one (01) output.

The neural network topology diagram used to model insect density (see in Figure 1).

The computational capabilities of a neural network reside in the connections that exist between neurons. Each connection is assigned a numerical value called weight (W). Thus, the weights of the connections between the neurons of the hidden layer and those of the input layer are represented in Table 4 and those between the neurons of the hidden layer and the output are in Table 5. The biases (b) also are represented relative to the neurons of the hidden layer and that of the output in Tables 4 and 5.

1.png

Figure 1. Architecture of the optimal neural network

The bias of a neuron represents the error committed by the neuron during the calculation of the output value. The different outputs of neurons in the hidden layer are expressed by Eqns. (9), (10), (11), (12) and (13) respectively.

$I_1^{\prime}=\operatorname{TanH}\left(w_{11} t^{\prime}+w_{21} T_{s u}^{\prime}+w_{31} T_e^{\prime}+w_{41} T^{\circ \prime}+b_1\right)$                   (9)

$I_2^{\prime}=\operatorname{TanH}\left(w_{12} t^{\prime}+w_{22} T_{s u}^{\prime}+w_{32} T_e^{\prime}+w_{42} T^{\circ \prime}+b_2\right)$                  (10)

$I_3^{\prime}=\operatorname{TanH}\left(w_{13} t^{\prime}+w_{23} T_{s u}^{\prime}+w_{33} T_e^{\prime}+w_{43} T^{\circ^{\prime}}+b_3\right)$                   (11)

$I_4^{\prime}=\operatorname{Tan} H\left(w_{14} t^{\prime}+w_{24} T_{s u}^{\prime}+w_{34} T_e^{\prime}+w_{44} T^{\circ \prime}+b_4\right)$                   (12)

$I_5^{\prime}=\operatorname{TanH}\left(w_{15} t^{\prime}+w_{25} T_{s u}^{\prime}+w_{35} T_e^{\prime}+w_{45} T^{0^{\prime}}+b_5\right)$                  (13)

with:

I’₁: Value at the output of the first neuron of the hidden layer;

I’₂: Value at the output of the second neuron of the hidden layer;

I’₃: Value at the output of the third neuron of the hidden layer;

I’₄: Value at the output of the fourth neuron of the hidden layer;

I’₅: Value at the output of the fifth neuron of the hidden layer;

t', T_Su', T'_e, T^°' and are the respective normalized values of time, sugar content, water content and temperature.

Our regression model is obtained by the normalized value of the density of insects in the cocoa beans at the outlet of the network. This model is expressed by Eq. (14):

$I^{\prime}=w_1 I_1^{\prime}+w_2 I_2^{\prime}+w_3 I_3^{\prime}+w_4 I_4^{\prime}+w_5 I_5^{\prime}+b_6$                       (14)

with I’ the value of the network output.

Table 4. Values of the weights and biases of the inputs

Table 5. Values of the weights and biases of the output

3.2 Validation of optimal neural architecture

Validation of the 4-5-1 topology neural model was performed by the coefficient of determination between the values predicted by the network and the experimental ones of the remaining measurements from the database (Figure 2). This result with a coefficient of determination equal to 0.9982, showing that our model is close to reality by 99.82%.

Also Figure 3 confirms the performance of our neural network with an R-value very close to 1.

2.png

Figure 2. Experimental and simulated curve of the evolution of insect density

3.png

Figure 3. Evaluation of model for prediction of insect density

The main technique used in this study is the AIC for the choice of the optimal values of hyperparameters of the neuronal architecture. Thus, the experimental results with an equal to 0.9982 confirmed the relevance of our choice of the 4-5-1 architecture, after tests on the data set. This high value of R², indicates that a good model has been built for the prediction of insect density. In the same way, a high R² was obtained in the prediction of the infestation caused by C.chinensis and C.maculatus in the chickpeas stored without and in the isolated conditions with an ANN architecture comprised a hidden layer with a input layer (variable X) and an output layer (variable Y) [21]. Similar studies not with ANNs but with multiple linear regression (MLR) models predicted uric acid and protein content as indices of Sitophilus granarius infestation in wheat with as high (0.958; 0.978) respectively. In the same order, with ANN, the backpropagation MLP approach was used to classify Rhyzopertha dominica infestation in stored and wheat found 98.9% accuracy [53]. In most predictions of insects in grains in general, we observe R² greater than 0.90, also showing the robustness of ANNs.

The results reinforced the hypothesis that ANNs could successfully predict the evolution of insect density in cocoa beans. This prediction is of importance in terms of reducing post-harvest fumigation treatments and related losses. The ANN technique can be adopted to assess insect population growth in cocoa bean storage as a decision support tool in any cocoa bean storage facility.