IntroductionTop
Nowadays lowering the resources use with practices not affecting the production rates is becoming more and more crucial due to their scarcity and the increasing competitiveness together with the raising awareness of the pressure agriculture has on the environment. An example is given by the introduction of precision farming technologies into conventional farm activities that has provided operators the opportunity to cope with in-field variability and to handle and manage the resources and information efficiently (Fountas et al., 2015). Similarly, choosing agricultural machinery of the appropriate size, and being able to predict in advance the operational costs, can significantly contribute to optimize the resources use. The estimation of operating costs of agricultural and forestry (Verani et al., 2015) machinery and the definition of economic competitiveness gap (conditionality standards on agricultural farms and short- and medium-term business planning of the agricultural farm activities) are a key factor in planning policies of Rural Development Programs (RDPs) both at national and European level.
Several authors developed different approaches and methods for cost estimation and calculation of mechanized farming operations in relation to their specific needs. Lazzari & Mazzetto (1996) developed the Computed Farm Machinery System model (ComFARMS) to analyse arable farm mechanisation problems from strategic or management standpoints. In this context, the purpose is to enable the user to carry out sensitivity analysis by modifying input off-farm and on-farm data and to perform multi-criteria methods that include subjective and non-numeric information. Moreover, Søgaard & Sørensen (2004) developed a non-linear programming model implemented by using the programming software suite General Algebraic Modelling System (GAMS) to support the process of choosing the optimal level of farm mechanisation in terms of technical capability. It is based on a least-cost concept, which involves all expected fixed and variable costs for a particular farm size and crop plan Camarena et al. (2004) developed an integrated approach called MULTIPREDIO created mixed integer linear programming linked to several databases contained in spreadsheets to select agricultural machinery for a multifarm system. Their approach aims to select the optimal machinery set for each farm, which corresponds to the lowest annual mechanisation cost of the multifarm system through the time. Another method has been developed by Gunnarsson & Hansson (2004) to examine different harvesting systems, estimate timeliness costs and to present conclusions on harvesting machinery selection. In the study, the direct machine costs (including both, fixed and variable) were calculated with conventional methods using parameters from the ASAE Standards (ASAE D497.4, 2000). Bochtis et al. (2010) developed a targeted approach for the estimation of the operational machinery costs on an annual basis in controlled traffic farming (CTF) systems. Their approach combines four sub-models based on specific algorithmic approaches in order to evaluate the consequences in terms of machinery performance (following different driving directions) when establishing tramlines in a CTF system. The work used specific equation given in the Agricultural Machinery Management Data ASAE Standard (ASAE D497.6, 2009).
Spinelli et al. (2011) carried out a study in order to calculate the economic life, the annual use and the residual value of two harvesting machines of forestry products. This was done starting from a large database (coming from Europe and North America) of second-hand machine sale offers containing over 1000 records. The information contained in the study are crucial for machine rate calculation often based on rule-of-thumb assumptions, following the absence of empirical data. In the study of De Toro et al. (2012) a model capable to predict the moisture content of wheat using historical weather data, was used to assess the effects of weather on cereal harvesting costs (machine, labour, timeliness, drying). The specific machine costs (i.e., for the combine harvester) were estimated using ASAE standard methods (ASAE D497.5, 2006; ASAE EP496.3, 2006) with the following parameters: depreciation, residual value, annual use and the economic life.
In spite of the numerous elements of randomness illustrated above, traditional estimation of operating costs of agricultural machinery is largely based on deterministic models, which perform the same way for a given set of initial conditions (Abramo et al., 2015). Conversely, in the statistical model with respect to the deterministic ones, randomness is present and variable states are not described by unique values, but rather by probability distributions.
In this study, a Partial Least Squares (PLS) multivariate statistical model was adopted to predict the fuel consumption and costs of six agricultural operations such as: ploughing, harrowing, fertilization, sowing, weed control and shredding. This prediction has been done starting from some selected parameters, which are: soil workability, non-working distance travelled, time per surface-area unit, maximum engine power, purchase price of the tractor and purchase price of the operating machinery in order to optimize crucial agricultural operations and thus increase the farm performance.
Material and methodsTop
Data collection
Data were collected from 2011 to 2014 in several experimental fields of CREA located in different parts of the Italian country (Fig. 1). The soils of the various experimental fields have different characteristics regarding slope, texture, shape, surface and crop grown (Table 1).
|
|
Table 1. Soil characteristics of experimental fields regarding: number of sites, slope (%), texture (%), shape, surface (m2) and crop grown.
The different agricultural operations are in details: ploughing (54 observations), harrowing (70 observations), fertilization (65 observations), sowing (43 observations), weed control (22 observations) and shredding (19 observations). The agricultural operations were carried out using tractors with different brand and power and operating machineries with different life time, annual machine use, repair and maintenance factor and accumulated use hours of machine (Fedrizzi et al., 2015). In Table 2, brand and model, engine power of the tractors for different operations are reported.
Table 2. Model and engine power of the machines used in the different operations.
The work times, at each agricultural operation, were recorded following the recommendations of the Italian Rural Engineering Association (AIGR), which considers the official methodology of the “Commission Internationale de l’Organisation Scientifique du Travall en Agriculture” (CIOSTA) (Manfredi, 1971; Biondi, 1999). In this study, in order to significantly reduce the influence due to the different characteristics of the farms, in terms of type of machinery used, distance of the fields and so on, only some items from the CIOSTA official methodology were considered.
The plots of land of the agricultural holdings examined differ in shape and size, geomorphology, soil composition and geographic location, as well as agronomical and administrative management. In order to significantly reduce the influence induced by this vast variability of agricultural holding characteristics, only several items of the CIOSTA method were taken into account.
Since the small area of experimental fields (about 0.5 ha) for the collection of work times (as reported by CIOSTA), only those related to the effective work time (TE) and to the tractor turn-around time (TAV) (which together represent the net time, TN) were considered. Considering the TN, the hourly operating cost of each tractor and piece of machinery used, was determined by means of specific analytical methods, and successively, the cost per surface-area unit was determined for each agricultural operation.
The operating costs, for each agricultural operation, of the tractors and operating machineries, were identified by considering two main parameters: fixed and variable costs. The former involve the reintegration of the invested capital, the cost of capital using, and the various expenses (insurance, storage and taxes). The variable costs were related to the use of the agricultural machinery and include the expenses incurred for repairs and maintenance, fuel, lubricants and labour. Relatively to the life time and annual use for all tractors, a standard value of 15 years and 1067 hours per year respectively are considered.
The methods proposed in the bibliography are substantially similar in relation to the calculation of the fixed costs, whereas they differ in the formulas and coefficients adopted in calculating the variable costs. As far as this last item is concerned, reference has been made to a specific method (Biondi, 1999). In this study, the method proposed by Biondi (1999) has been updated with precise references to the technical standards that indicate the technical and economic coefficients to be used in the calculations (ASAE D497.4, 2003; ASAE EP496.2, 2003).
The maximum engine power (P; kW), and the purchase price of the tractor (€) were obtained from the publication “Buyers’ guide 2013” (Guida all’acquisto 2013), edited by the Italian magazine “L’informatore Agrario” (http://www.informatoreagrario.it/ita/riviste/infoagri/13Ia19/sommario.asp).
Furthermore, it was also necessary to carry out an economic assessment of all the operating machinery (ploughs, harrows, seeders, fertiliser spreaders, etc.) used in the cultivation activities. The purchase prices of the various machines was determined when possible form the producers price list otherwise contacting specific retailers through personal communications.
The fuel consumption per hectare (modified from Biondi, 1999) was calculated considering, first of all, the hourly fuel consumption (Fch; kg/h) per each agricultural operation using the following formula:
where, Sc = specific fuel consumption in kg/kWh; P = maximum engine power in kW; and d = power utilisation factor in %.
Then, d was described considering two different conditions: dte is the power used during effective operation and dtav is the power used during the turn-around operations and manoeuvres. By this way Eq. [1] becomes:
To calculate the fuel consumption per hectare (Fcha; kg/ha) we modified [2] as follows:
where Te = effective time consumed during the operation in hours; and Ttav = time consumed during the turn-around operations and manoeuvres in hours.
The methodology used for calculating tractors and machineries operational costs, is referred to the one used by Fedrizzi et al. (2015).
The values used for the specific fuel consumption (Sc), and the power utilisation factor (d), related to different examined operations are reported in Table 3. The value ranges are derived according to Biondi (1999) methodology. In the development of the exposed calculations, it refers exclusively to the aspects relating to costs directly connected with the use of machines and equipment excluding the costs of other inputs such as fertilizer, seed and pesticides.
Table 3. Factors used for the fuel consumption calculation in relation to the different operations extracted from Biondi´s (1999) methodology.
Multivariate modelling
A multivariate modelling approach was adopted to predict six agricultural operations, fuel consumption and costs. A two-step approach was applied.
In the first step, the fuel consumption for each agricultural operation was predicted from the first four variables: time per surface-area unit (h/ha), maximum engine power (kW), purchase price of the tractor (€) and purchase price of the operating machinery (€). Only for the ploughing fuel consumption two additional dummy variables were considered: soil workability (high = 1, low = 0); and minimization of the tractor non-working distance travelled (optimized = 0; not optimized = 1, where not optimized regards outward with ploughing and return without ploughing).
In the second step, the costs for each agricultural operation were predicted from the four above mentioned variables and the fuel consumption predicted as a result of the first step. A PLS regression approach was applied (e.g., Wold et al., 2001; Costa et al., 2013; Infantino et al., 2015; Cutini et al., 2016) on the above mentioned datasets. The PLS is a particular type of multivariate regression which uses a two-block predictive PLS model. The regression analysis objective is achieved by using the equation that minimizes the residual mean square error, or maximizes the coefficient of multiple determination r2, which is the most commonly used statistic to measure the forecasting potential of a multiple regression equation. The predictive ability of the model also depends on the number of latent vectors (LV) used. Generally, a good predictive model should have high values of the Pearson correlation coefficient (r) and low values for the root mean square error in calibration (RMSEC). The procedure calculated the ratio of percentage deviation (RPD), which is the ratio of the standard deviation of the measured data to the RMSE (Williams, 1987). The model chosen was for the number of LV that yielded the highest r, minimum standard error of prevision (SEP) for predicted and known Y-block and maximum RPD. The PLS models were developed starting from a calibration set (training/evaluation set; Forina et al., 2008), consisting of 50% of the sample. The PLS model (cross-validated) was then validated on a set on internal tests consisting of the remaining 50% of the samples. The partitioning was carried out using the sample set partitioning-based on joint X-Y distances (SPXY) algorithm (Harrop Galvao et al., 2005) that takes account of the variability in both X and Y. The independent variables (X-block) datasets were standardised using an autoscale algorithm (i.e., centers columns to zero mean and scales to unit variance). For each PLS model a summary of the relative importance of the independent variables (Y-block) to predict the dependent one is given by variable importance in the projection (VIP) (Febbi et al., 2015; Taiti et al., 2015). The VIP scores estimate the importance of each variable in the PLS-based models and were calculated according to Chong & Jun (2005). The PLS models were developed using a procedure written in the MATLAB 7.1 R14 environment.
The obtained PLS models were applied to a standard sized Italian farm: <10 hectares farm with two tractors with reduced power (118 and 59 kW). This in order to calculate costs per hectare and fuel consumption for the six operations examined.
ResultsTop
The performance of PLS models with different LVs in the determination of the fuel consumption for each agricultural operation is summarized in Table 4. The cumulated variance for the X-blocks ranged from 56.9% for weed control to 99.88% for fertilization. For the Y-block the cumulated variances ranged from 5.7% for sowing to 15.85% for fertilization. The RMSEC unbiased values had the same unit range of the original data and were always very low. Considering correlation coefficients in the calibration/validation set, these were always very high (r> 0.96); the same result could be observed for test set showing very high performances (> 0.94), except from fertilization (0.91).
Table 4. Descriptors and principal results of the Partial Least Squares (PLS) regression model in estimating fuel consumption at different agricultural operations.
Figure 2 shows the scatter plot of the observed versus predicted (biased) values by the PLS models in estimating fuel consumption at different agricultural operations. For both datasets, the biased observations are well distributed along the bisectrix, indicating a good performance in predicting the fuel consumption.
|
|
The VIP scores obtained by the PLS regressions to estimate fuel consumption of the different agricultural operations are showed in Table 5. It is possible to observe as, for all the agricultural operations, the time per surface-area unit (h/ha) is the most important variable in predicting fuel consumption. For ploughing, also the soil workability is an important variable in prediction.
Table 5. Variable importance in the projection (VIP) scores of the partial least squares (PLS) regression models in estimating fuel consumption at different agricultural operations.
The performance of the PLS models with different LVs in the determination of the costs for each agricultural operation, is summarized in Table 6. The cumulated variances for the X-blocks ranged from 96.67% for sowing to 100% for ploughing and shredding. For the Y-block the cumulated variances ranged from 9.82% for weed control to 20.29% for fertilization. The RMSEC unbiased values had the same unit range of the original data, and were very low. Considering the correlation coefficients in the calibration/validation set, these were always very high (r> 98); the same result could be observed for test set showing very high performances (> 98).
Table 6. Descriptors and principal results of the Partial Least Squares (PLS) regression model in estimating costs at different agricultural operations.
Figure 3 shows the scatter plot of the observed versus predicted (biased) values by the PLS models in estimating costs at different agricultural operations. For both datasets, the biased observations were well distributed along the bisectrix, indicating a good performance in predicting the costs.
|
|
The VIP scores obtained by the PLS regressions to estimate costs of the different agricultural operations are showed in Table 7. It is possible to observe as, for ploughing, harrowing, sowing and weed control the fuel consumption was the most important variable. Equally, the time per surface-area unit (h/ha) was the most important variable for fertilization and shredding. Sowing showed higher values for the purchase price of the operating machinery, differently from the others (Table 7).
Table 7. Variable importance in the projection (VIP) scores of the partial least squares (PLS) regression models in estimating costs at different agricultural operations.
An example of application of the proposed model for a standard sized Italian farm (<10 hectares farm with a reduced mechanization), to calculate costs per hectare and fuel consumption for the six operations examined, is shown in Table 8. The proposed model, differently from the results obtained according to Biondi (1999) methodology, allowed a more simpler calculation of consumption, based on the most simple and expeditious parameters already known by the farmer, as for example the tractor engine power, total work time, tractor and equipment price.
Table 8. Example of application of the prediction model of fuel consumption and unit costs for the six operations in a standard Italian farm.
DiscussionTop
The monitoring of agricultural operations, in particular the costs and fuel consumption of the machinery, is a large important portion for the economical farm balance. In order to obtain an uniform procedure for machinery cost analysis, this study adopt standard models to predict fuel consumption and costs of six agricultural operations, such as ploughing, harrowing, fertilization, sowing, weed control and shredding, on the base of six variables.
For the first time, a two-step approach was applied to predict six agricultural operations fuel consumption and, then, from esteem fuel consumption, the costs in agricultural engineering. The following variables were used to develop a predictive model and thus for the estimation of costs and fuel consumption: time per surface-area unit; maximum engine power; purchase price of the tractor and of the operating machinery; soil workability; and non-working distance travelled. This method on one side uses fewer variables with respect to the ASAE one (ASAE EP496.2, 2003) or other methodologies, and on the other side can be easily utilized by farmers. In fact, the farmer knows the surface-area of their plots, the technical and economic aspects of their tractors and operating machines, and the average time required to perform the various growing operations. Since this method requires as input only this information, its use is greatly simplified if compared to other systems and methodologies more articulated that require numerous other qualified input factors. In this scenario, a great number of authors have developed different modelling approaches for cost and fuel consumption estimation and calculation of mechanized farming operations.
The majority of these methods is based on quantitative agricultural analysis (e.g., income, prices, farm size, efficiency, factor allocation, production, welfare, etc.) at different levels of scale (Happe et al., 2006). In this study, the analysis is based on a multivariate statistical model, in which randomness is present and variable states are described by probability distributions, with respect to a deterministic one. As reported by Costa et al. (2012), a multivariate regression approach is particularly useful when predicting one or more dependent variables from a large set of independent variables, often collinear. In fact, this work adopted a multivariate modelling approach to predict agricultural operations, fuel consumption and costs applying two steps. In the first, the fuel consumption for each agricultural operation was predicted from the above mentioned variables meanwhile in the second, the costs for each agricultural operation were predicted from these variables and the fuel consumption predicted as a result of the first step. All the models constructed for the prediction of fuel consumption resulted to be highly performant.
Considering the correlation coefficients, the r values for test set were > 0.94, except for fertilization (0.91). Considering also the performances in the test set of the model constructed to predict costs, the r values were all > 0.96.
Using this approach, it is possible to observe also the importance of the variables in the prediction with the analysis of the VIP scores. For ploughing, harrowing, fertilization, weed control, sowing and shredding, the time per surface-area unit is the most important variable in predicting fuel consumption. For ploughing, also the soil workability is an important variable in prediction. In addition, for ploughing, harrowing, sowing and weed control the fuel consumption is the most important variable in predicting costs. Equally, the time per surface-area unit is the most important variable for fertilization and shredding. Sowing showed higher values for the purchase price of the operating machinery, differently from the others. This is due to the high initial cost of purchase (ranging from €500 to €18,000) and for the oversizing in terms of technical and economic aspects of the operating machineries.
The provisional models applied to a standard Italian farm (<10 hectares farm with a reduced mechanization), showed the possibility to apply such an approach to generate scenarios different purposes, from policy makers to agricultural operators to farm contractors.
In conclusion, the obtained models being based on an input dataset very heterogeneous (in terms of field: shape, dimensions, slope, texture, surface, crop grown etc. and in terms of machines/operators) resulted to be extremely efficient and so generalizable and robust. This represents a crucial characteristic needed to transfer the approach from theory to practice.
The advantages of the proposed predictive model are related to the simplicity for the farmers and policy makers to acquire the necessary information. In fact, to calculate consumption and cost per hectare, is sufficient to know the tractor engine power, total work time, tractor and equipment price. In this way, it is possible to get the desired results, without a priori knowledge of a great number of parameters and elements as required by the traditional methodologies.
This approach may results extremely useful for both farmers (in terms of economic advantages, e.g. while choosing correct machines dimension and power or operations scheduling and management) and at institutional level representing an innovative and efficient tool in order to define correctly the gasoline reimbursement or planning future Rural Development Programmes and the Common Agricultural Policy. In light of these advantages, the approach may be implemented on a web platform (including specific app) and made available to all the stakeholders.