Modelling hatchability and mortality in muscovy ducks using automatic linear modelling and artificial neural network

This study was embarked upon to predict hatchability and mortality rate of Muscovy ducks in Nasarawa State, Nigeria. Data were obtained from a total of 119 duck farmers. The automatic linear modelling (ALM) and artificial neural network (ANN) models were employed. The average flock size was 9.84±0.60 per household. The predicted hatchability mean values using ALM (8.66) and ANN (8.65) were similar to the observed value (8.66). The predicted mortality mean values using ALM (2.95) and ANN (3.03) were also similar to the observed value of 2.95. Experience in duck rearing, the educational status of farmers, source of foundation stock and season were the variables of importance in the prediction of hatchability using ALM and ANN models. However, primary occupation, source of foundation stock, experience in duck rearing, land holding and management system were the important variables automatically selected for the prediction of mortality. Moderate coefficients of determination (R2 = 0.422 vs 0.376) and adjusted R2 (0.417 vs 0.371) estimates were obtained for hatchability and mortality using ALM. Different patterns were obtained under the ANN models as regards the prediction Predicting hatchability and mortality in Muscovy ducks (A. Yakubu et al.) 65 of hatchability (R2= 0.573 and adjusted R2= 0.569) and mortality (R2= 0.615 and adjusted R2= 0.612). The present information may aid management decisions towards better hatchability and mortality performance in Muscovy ducks. Kata kunci: Ducks, performance, neural network, regression, Nigeria

of hatchability (R 2 = 0.573 and adjusted R 2 = 0.569) and mortality (R 2 = 0.615 and adjusted R 2 = 0.612).The present information may aid management decisions towards better hatchability and mortality performance in Muscovy ducks.

INTRODUCTION
In developing countries such as Nigeria, poultry production is largely managed under extensive free range or scavenging system, particular at villages and peri urban areas.Majority of the birds are reared at the rural level especially the indigenous stock, providing reservoir for the genetic conservation of the indigenous population.Poultry provide enormous opportunity to the rural poor from the generation of family income to employment opportunity (Yakubu, 2010;Yakubu et al., 2011;Johari et al., 2013).Lack of understanding of village poultry production system will normally impede design and implementation of poultry bird advancement program that will impact positively on the rural poor.It is pertinent to understand production system and constraint at this level in other to fashion policies that will enhance productivity of this system, thereby guaranteeing sustainable agriculture (Gómez et al., 2016).
Ducks ranked third among the various poultry species in Nigeria (Hassan and Mohammed, 2003), with population put at approximately 11 million and distribution cutting across all the agro-ecological zones of Nigeria particularly in village settings (NBS, 2012).In a section of the country, most farmers were found keeping Muscovy ducks on extensive sheds (Etuk et al., 2006).The advent of commercial fastgrowing and egg-laying strains of chickens has relegated to the background the relevance and relative contribution of indigenous poultry species such as chicken, duck, guinea fowl and pigeon to the internal animal protein production in Nigeria.This trend has adversely impacted on duck production as exemplified in its remarkable reduced population and dearth of empirical studies directed towards management and genetic improvement of this waterfowl in Nigeria (Yakubu, 2013;Oguntunji, 2013;Oguntunji and Ayorinde, 2014).The dwindling reproductive performance and high mortality rates of Muscovy ducks is a major concern as farmers income and protein intake are drastically affected.This may in the long run negatively affect food security and livelihood of the farmers.Hence, the need to identify the factors influencing the performance of the birds at the village level with a view to mapping out appropriate strategies to boost production.
The artificial neural network (ANN) is an alternative to the traditional regression statistical technique and a potential tool in poultry production for the modelling of performance data.ANN is a non-linear parametric model that mimics the processing mechanism of the human brain.There is increasing use of this algorithm to predict hatchability (Bolzan et al., 2008), growth (Yakubu et al., 2018a) and egg production (Ahmad, 2011).It has also been used to model disease occurrence (Akil and Ahmad, 2016).
There is dearth of literature on the use of robust models to forecast reproductive and mortality performance in Muscovy ducks in Nigeria.Therefore, this study aimed at predicting the reproductive and mortality rates of Muscovy ducks from some social-economic factors of smallholder farmers and performance characteristics using different statistical algorithms.

Description of Study Area
The study was carried out in Nasarawa State, North Central Nigeria.It is located within the guinea savannah agro-ecological zone and lies on latitudes 7 ° 52′ N and 8 ° 56′ N and longitudes 7 ° 25′ E and 9 ° 37′ E, respectively (Lyam, 2007).The three Senatorial Zones of Nasarawa South, Nasarawa North and Nasarawa West were covered.

Sampling Techniques
A total of 120 Muscovy duck farmers (40 per zone) were randomly sampled in selected villages of the study area, but data from 119 farmers were eventually used for analysis.Only farmers who were willing to participate in the exercise were interviewed.

Data Collection Techniques
Structured questionnaires were administered to the duck farmers including face-to-face interview.Information sought included the socioeconomic characteristics of the respondents, livestock ownership, flock sizes and structure, productive and reproductive performance indices, mortality rate, knowledge on health and other management practices.

Statistical Analysis
The categorical (using Chi-square) and continuous variables (using Means±S.E.) were subjected to descriptive statistics.The relationship between the response variables (hatchability and mortality number; each handled singly) and predictor variables were established using Automatic Linear Modelling (ALM) and Artificial Neural Network (ANN) algorithms.The hatchability parameter fitted was number of eggs hatched while mortality was assessed in terms of number of birds that died.
Age of farmers, sex, marital status, educational background, primary occupation, experience in poultry keeping, management system, health management practices (veterinary access, veterinary category, use of herbs) season of highest hatchability, age at first lay, access to credit, personal savings in financial institution and land holding were the input predictor variables fitted into the ALM to estimate reproductive success.Similarly, mortality rate was predicted from age of farmers, sex, marital status, educational background, primary occupation, experience in poultry keeping, management system, health management practices (veterinary access, veterinary category, use of herbs) season of highest mortality, age at first lay, access to credit, personal savings in financial institution and land holding.In each case, all the variables that were nominal were assigned as factors while all variables that were continuous were treated as covariates.Every other step was as described by LaFaro et al. (2015) and adopted by Yakubu et al. (2018).
All the explanatory variables of importance under ALM were fitted into the ANN model to predict hatchability and mortality number, respectively as described by LaFaro et al. (2015).Multilayer Perception (MLP) with Back-Propagation network was used.The network was trained with 80% and tested (model validation) with 20% of the data set.Every other choice in the neural network was set to default (Yakubu et al., 2018a and b).SPSS (2015) was employed in both analyses.

RESULTS
The sex, marital status, education, primary occupation, access to credit and type of landholding varied significantly (P≤0.05;P≤ 0.01) among the duck farmers (Table 1).As regards the continuous variables, the average age of respondents, family size and experience in duck keeping (years) were 44.54, 8.49 and 5.02.
The average age of ducks at first lay (months), clutch number per year, egg number in a clutch, brooding length (weeks), egg number hatched in a clutch and mortality rate per annum were 5.51, 2.84, 9.87, 4.68, 8.66 and 2.95, respectively (Table 4).While the highest hatchability was recorded in the wet season (P ≤ 0.01), mortality rate was highest in the hot-dry season (P ≤ 0.01).
The summary statistics of observed and predicted hatchability and mortality rate of Muscovy ducks are shown in Table 5.The predicted hatchability mean values using ALM (8.66) and ANN (8.65) were similar to the observed value (8.66).The Standard deviations were 1.80 (ALM), 2.12 (ANN) and 2.78 (observed), respectively.As regards mortality, the predicted mean values using ALM (2.95) and ANN (3.03) were also similar to the observed value of 2.95.The respective standard deviations were 1.72, 2.14 and 2.80.
In the ALM model, experience in duck rearing and the educational status of farmers were the two significant variables in the prediction of hatchability out of the four important parameters Table 6).In the ALM model, primary occupation, source of foundation stock, experience in duck rearing, land holding and management system were the five significant variables automatically selected for the prediction of mortality (Table 7).
In ANN model, experience in duck rearing (0.387), source of foundation stock (0.320), educational status (0.148) and season of hatchability (0.144) were the four parameters of utmost importance in the prediction of hatchability (Table 8).As regards the prediction of mortality using ANN, experience in duck rearing (0.422), primary occupation (0.315), source of foundation stock (0.125), land holding (0.082) and management system (0.057) were the five parameters of utmost importance (Table 9).The association between the observed and the predicted hatchability and mortality in form of a linear regression using ALM is shown in Figures 1 and 2. The correlation coefficients (r = 0.649 vs 0.613) were fairly high, while moderate coefficients of determination (R 2 = 0.422 vs 0.376) and Adjusted R 2 (0.417 vs 0.371) estimates were obtained for hatchability and mortality.The root mean square errors (RMSE) of 2.12150 and 2.22431 and akaike's information criterion corrected (AICC) values of Different patterns were obtained under the ANN models as regards the prediction of hatchability and mortality, where r =0.757 R 2 = 0.573; Adjusted R 2 = 0.569 and RMSE was 1.82357 (hatchability) (Figure 3); r =0.784 R 2 = 0.615; Adjusted R 2 = 0.612 and RMSE was 1.75277 (mortality) (Figure 4).

DISCUSSION
Muscovy duck is one of the meat-producing livestock (Susanti and Purba, 2017).The preponderance of women farmers agrees with the general assertion that smallholder poultry is to a large extent under the control of the women folks.The flock size obtained in the present study is   Nwanta et al. (2006), Ola (2000) and Chia and Momoh (2012), respectively.The difference between their findings and that of the current study may largely be attributed to genetic factor, varying management systems and the periods records were taken.However, there is need for improved management practices by farmers in Nasarawa State to guarantee higher egg production.The low number of farmers that use herbs to treat their ducks in the present study is an indication of poor knowledge on the use of ethnoveterinary medicine.
Muscovy duck represents a suitable model for hypothesis testing in breeding biology of waterfowl under natural incubation; reproductive consequences of eggs laid can best be assessed through the number of eggs hatched.The current findings are congruous to the findings of Oguntunji and Ayorinde where majority (44.5%) of the respondents indicated that female ducks underwent two reproductive cycles in a year.Muscovy ducks are very good setters, capable of hatching 12-15 duck eggs.The hatchability value of the present study (about 88%) appears higher than the values reported for between normal (76%) and dump nests (77%) genetically unselected variety of Muscovy duck in Mozambique (Harun et al., 1998), 70.7% and 69.7% reported by Widiyaningrum et al. (2016) and 54.21% reported by Rashid et al. (2009).Similar hatching rate above 80% with that of the current study has been reported (Oguntunji and Ayorinde, 2015).
The higher hatchability recorded in the wet season is an indication of the degree of environmental comfort experienced by birds.This is in consonance with the report of Widiyaningrum et al. (2016) that environmental factors such as temperature and humidity are important for successful hatching.Our observation, however, is contrary to the report of Boonprong (2000), where hatchability was highest in winter followed by summer and rainy season, respectively.Harsh environmental factors (e.g.temperature, humidity, turning etc.) might be causes of higher mortality in the hot-dry season as observed in the current study.Heat stress made birds to pant and could result in heat stroke and mortality.It has been reported that extremes temperatures could be experienced in the hot-dry season in North Central Nigeria (Yakubu et al., 2018a), thereby making the birds uncomfortable.Such heat stressed birds could experience high rate of mortality and morbidity (Nidamanuri et al., 2017).According to Shittu et al. (2014), hotdry climatic environment is characterized by heat stress, inefficiency in the usage of feed and waning immunity, thereby leading to high mortality.
To the best of our knowledge, the present study appears as the first to predict hatchability and mortality rate of Muscovy ducks in Nigeria using robust algorithms such as ALM and ANN.Application of appropriate models to approximate the performance function warrants more precise prediction and helps to make the best decisions in the poultry industry.The better predictive ability of ANN in the present study could be as a result Therefore, it could serves as a veritable means of forecasting incubation performance in Muscovy ducks.This is in consideration of its robustness in tackling noisy input data, high tolerance to faults and dimensionality problem and generalization from the input data.According to Bolzan et al. (2008), ANN model outperformed its multiple linear counterpart in the prediction of hatched eggs.Mehri (2013) reported ANN-based model with a better accuracy (R 2 = 0.99) than that obtained in the present study.However, the difference might be attributed to the use of egg main physical characteristics as input variables in the earlier study as against socio-economic factors in the present study.Chamsaz et al. (2011) reported that the ANN produced more accurate predictions of hatchability than the linear regression equation (R 2 = 0.9984 versus 0.4003).It is, therefore, possible to elucidate the performance variables of birds using ANN as it facilitates scientific and objective decision making including the simulations (Salle et al., 2003) of the consequences related to such decisions.When the current knowledge is applied to the present study, it could guide management decisions and strategies geared towards boosting production duck production.In a related study in humans, high accuracy was obtained in the prediction of mortality using ANN model (Shi et al., 2012) while ANN has also been used to detect chicken growth anomaly from mortality rate and feed conversion ratio (Purnomo et al., 2018).

CONCLUSION
The predicted hatchability and mortality mean values using both ALM and ANN algorithms were similar to their respective observed values.Considering the moderate to high variation explained by ANN and ALM models in the prediction of hatchability and mortality rates, they appear to be reliable.Therefore, the two models could be recommended as veritable tools for the prediction of hatchability and mortality rates in ducks.Such prediction will aid management decisions to improve flock size and the associated profitability of the farm.

Figure 1 .
Figure 1.The Ccatter Plot of the Predicted and Observed Hatchability using ALM

Figure 3 .
Figure 3.The Scatter Plot of Observed and Predicted Hatchability using ANN

Table 2 .
Flock Structure of Muscovy Ducks Kept in Nasarawa State

Table 3 .
Management of Muscovy ducks Kept in Nasarawa State

Table 4 .
Productivity Indices of Muscovy Ducks Kept in Nasarawa State

Table 5 .
Descriptive Statistics of the Observed and Predicted Hatchability and Mortality Rates

Table 6 .
Fractional Importance of Some Variables to the Prediction of Hatchability using Automatic

Table 7 .
Fractional Importance of Some Variables to the Prediction of Mortality using Automatic Linear

Table 8 .
The Importance of Independent Variables in the Prediction of Hatchability using Artificial

Table 9 .
The Importance of Independent Variables in the Prediction of Hatchability using Artificial