Design of Groundwater Level Monitoring Networks for Maximum Data Acquisition at Minimum Travel Cost

: Groundwater monitoring networks represent the main source of information about water levels and water quality within aquifers. In this paper, a method is proposed for the optimal design of monitoring networks to obtain groundwater-level data of high spatial relevance at a low cost. It uses the estimate error variance reduction obtained with the static Kalman ﬁlter as optimization criteria, while simultaneously evaluating the optimal routes to follow through the traveling salesman problem. It was tested for a network of 49 wells in the Calera aquifer in Zacatecas, Mexico. The study area was divided into three zones, and one working day (8 h) was taken to visit each one, with an average speed of 40 km/h and a sampling time of 0.5 h. An optimal network of 26 wells was obtained with the proposal, while 21 wells should be monitored if the optimal routing is neglected. The average standard error using 49 wells of the original network was 35.01 m, an error of 38.35 m was obtained for 21 wells (without optimal routing) and 38.36 m with the 26 wells selected using the proposal. However, the latter produce estimates closer to those obtained with the 49 wells. Following the proposal, more ﬁeld data can be acquired, reducing costs.


Introduction
Groundwater represents 96.3% of the total freshwater on Earth. It is the principal source of water, and necessary for the survival and development of humanity. Around 69% of the extracted groundwater is used for agricultural activities, 19% is required by the industrial sector, and the remaining 12% is used for domestic purposes [1,2]. According to the National Water Commission (CONAGUA in Spanish) of the Mexican government, the water that is available to exploit without altering its natural balance is 451,585 hm 3 ; 39% of it corresponds to groundwater, while 61% corresponds to surface waters [3].
For adequate groundwater management, it is necessary to implement monitoring schemes in a set of piezometers or wells that allow for spatial and temporal information on the quality and levels of groundwater to be obtained by carrying out field campaigns. These networks represent the main source of information needed to make inferences about the hydrological behavior of water in the subsoil [4], which is of vital importance since the extraction of considerable volumes of groundwater has caused adverse effects on water levels, as well as quality [5].
The groundwater data collected in space and time are useful for the construction of numerical models that allow for a simulation of the evolution of groundwater levels and quality [6]. These models represent robust tools supporting water management policies. There are monitoring networks in several parts of the world at present, where the level and quality of groundwater and springs are measured at the regional level. The processing and representation of this information is carried out in Geographic Information Systems (GIS).
In many regions, the adequate management and assessment of water is difficult to achieve due to the limited monitoring the groundwater. To overcome this situation, the International Center for the Assessment of Groundwater Resources (IGRAC) belonging to the United Nations Educational, Scientific and Cultural Organization (UNESCO) founded the Global Groundwater Monitoring Network (GGMN). This network is based on a catalog of 166 parameters/variables included in a GIS, with the aim of facilitating access and updating information to gaining broader knowledge about the functioning of water in subsoil with the aim of its sustainable management [7].
In Mexico, a total of 1118 groundwater monitoring stations were reported in 2017 (1096 of them were wells). These were used for the surveillance of the 653 administrative aquifers defined by CONAGUA. The annual monitoring of groundwater quality has been carried out since 2005, especially in the central and northern regions, due to its relevant role in the country's economic activities. The data obtained from this monitoring program were used to evaluate compliance with the Official Mexican Standard (NOM-127-SSA1-1994), which establishes the permissible limits of groundwater quality parameters for human consumption and agricultural purposes [3]. These networks must be designed by following the criteria that allow for the greatest amount of information to be obtained at the lowest possible cost. Accordingly, several researchers have developed methods for the optimal design of monitoring networks by looking at one or several variables, in space or space-time [8,9]. Some designs aim to minimize the error in the estimation of the hydraulic head in space and over time with the lowest possible number of wells, to reduce the cost of exploration. The estimation method that is usually used is kriging, although other works propose the use of the Kalman filter [10,11]. Bierkens et al. [12] use an autoregressive exogenous (ARX) model to estimate changes in groundwater levels. Geostatistical techniques are used to calculate the ARX parameters at unmonitored sites. The regionalized model was incorporated into a spatiotemporal Kalman filter for the optimal prediction of water level variation.
Different methods for the optimal design of groundwater quality and groundwaterlevel monitoring networks used an extended, non-dominated classification genetic algorithm in the optimization [9,10]. Luo et al. [13] sought to optimize a long-term monitoring network of a contaminant concentration where there is a scarcity of hydraulic conductivity data through a method that incorporates total sampling costs and estimation errors for a mass contaminant. Several researchers designed optimal monitoring networks for the study and treatment of groundwater contamination [14][15][16][17][18].
Other works were developed for the estimation and the optimal design of groundwaterlevel monitoring networks using artificial neural networks [19], or multicriteria analysis implemented in a geographic information system (GIS) [20]. However, the most popular approach for the design of groundwater-level monitoring networks (GLMN) is the application of geostatistical interpolation techniques [4,[21][22][23][24].
On the other hand, few works have considered the routes that should be followed in sampling campaigns to obtain field data. One alternative is to solve the traveling salesman problem (TSP) to optimize the route and visit several sites of interest [25][26][27][28][29]. Nunes et al. [30] proposed optimal well subsets that will form part of a groundwater-level monitoring network, incorporating the TSP. They built an objective function (OF) that evaluates the reduction in the spatial variance, temporal redundancy and exploration costs (travel and monitoring) for a previously defined fixed number of wells. Simulated annealing (SA) was used to optimize the OF. The problem described is of a combinatorial type and involves a cost-benefit analysis, which guarantees quality information at low cost. With this method, the selected wells vary depending on the size of the network that was set prior to optimization.
The aim of this paper is to present a new method for the design of groundwater-level monitoring networks, to select the maximum number of wells that can be visited with a predefined budget, minimizing travel distances and maximizing the amount of information that can be provided. The objective function considers both the order of priority assigned to each well, according to the reduction in the estimate error variances of the groundwater levels, obtained by means of the static Kalman filter and the order of priority from an analysis of the optimal route, solving the TSP. This proposal was developed based on [30], which incorporates exploration times in the objective function, and [31], which employs a Kalman filter-based method to evaluate the reduction in the total estimate error variance.

Location of the Study Area
The method was tested to redesign the Calera aquifer monitoring network located in Zacatecas, Mexico, between parallels 22 •  The aim of this paper is to present a new method for the design of groundwater-lev monitoring networks, to select the maximum number of wells that can be visited with predefined budget, minimizing travel distances and maximizing the amount of info mation that can be provided. The objective function considers both the order of prior assigned to each well, according to the reduction in the estimate error variances of t groundwater levels, obtained by means of the static Kalman filter and the order of prior from an analysis of the optimal route, solving the TSP. This proposal was developed bas on [30], which incorporates exploration times in the objective function, and [31], whi employs a Kalman filter-based method to evaluate the reduction in the total estimate err variance.

Location of the Study Area
The method was tested to redesign the Calera aquifer monitoring network located Zacatecas, Mexico, between parallels 22°41′ and 23°24′ north latitude and meridia 102°33′ and 103°01′ west longitude, with an area of 2226 km 2 (Figure 1  According to the 2010 census in Mexico, within the Calera aquifer, there are 14 wells, from which a total volume of 176.5 hm 3 of water is extracted annually. Of this, 15 hm 3 is used in the agricultural sector, 11.1 hm 3 for drinking water supply, 1.1 hm 3 f According to the 2010 census in Mexico, within the Calera aquifer, there are 1417 wells, from which a total volume of 176.5 hm 3 of water is extracted annually. Of this, 159.2 hm 3 is used in the agricultural sector, 11.1 hm 3 for drinking water supply, 1.1 hm 3 for livestock and domestic use, and 5.1 hm 3 for industrial use. The Calera aquifer is of an unconfined, heterogeneous and anisotropic type. Its thickness is approximately 400 m, the hydraulic conductivity values vary from 0.22 to 2.2 m/day, and it has an average specific yield of 0.13. Geologically speaking, its upper layer largely consists of alluvial material, conglomerate and rhyolite-acid tuff. The underlying layers are composed of volcanic (acid tuffs, rhyolites and ignimbrites) and sedimentary (limestone, shale and sandstone) rocks. In the area of interest, a semidry climate was identified. The rainy season occurs more frequently in the summer. For the 1980-2009 period, the mean annual values for rainfall, temperature, and potential evaporation were 425 mm, 16.3 • C, and 2263 mm, respectively [32]. The most important crops in the region are green chili, beans, corn for grain, onion, garlic and alfalfa, and they are produced with irrigated agriculture, which occupies four times less surface than rainfed agriculture [34].

Existing Groundwater-Level Monitoring Network
There are no piezometers built within the Calera aquifer, so the monitoring network is made up of abandoned wells enabled for this purpose; the existing monitoring network (MN) comprises a total of 49 wells.
The groundwater levels in these wells are measured annually with the use of water level meters. Additionally, water samples are obtained to analyze their quality in the laboratory. Since the monitoring is carried out in this way, a design that considers ways to reduce the costs in the acquisition of piezometric data is needed. For the redesign, groundwater-level data for the year 2017 were used, since this was the most sampled year in recent times. These data were provided by the Department of Groundwater of CONAGUA in Zacatecas (Table A1), which is shown in Appendix A.

Estimation Grid and Monitoring Zones
An estimation grid was defined in the study area, with nodes separated by 2 km in longitude and 2 km in latitude (291). The financial budget for monitoring was sufficient for three working days (8 h each). It was decided to divide the study area into three zones, investing one working day for each of them. A total of 84 nodes of the estimation grid corresponded to Zone C1, 140 to Zone C2, and 143 to Zone C3 ( Figure 2).
The matrix of route distances between all the sites (including the base and all the monitoring wells) was built using the measurement tool incorporated in the digital map of the National Institute of Statistics and Geography (INEGI in Spanish), which includes the roads and paths that connect the different communities of Mexico [35].

Geostatistical Analysis
Geostatistics is composed of techniques to obtain minimum variance estimate values at unsampled sites or times [36].
A classical geostatistical analysis consists of three steps: an exploratory analysis (identification of outliers and evaluation of the data distribution function), the structural analysis (calculation of the sample variogram and the adjustment of a valid variogram model), and the kriging estimation. For the proposed method, the exploratory and structural analyses of a geostatistical procedure were used to derive a variogram model and its corresponding covariance matrix, but the Kalman filter was selected for the optimization estimation method instead of kriging to reduce computational efforts.
To complete the exploratory analysis, the histogram of data must approximate a gauss bell. Within the structural analysis, a theoretical variogram model is fitted to the sample variogram.
To select the variogram model, a cross-validation is carried out using the procedure called leave one out, in which each data are individually removed, and the others are used to estimate or predict at that site or time. The entire geostatistical procedures were performed in the R free software [37].
Finally, the covariance matrix was constructed with the parameters of the variogram model [38].

Geostatistical Analysis
Geostatistics is composed of techniques to obtain minimum variance estimate values at unsampled sites or times [36].
A classical geostatistical analysis consists of three steps: an exploratory analysis (identification of outliers and evaluation of the data distribution function), the structural analysis (calculation of the sample variogram and the adjustment of a valid variogram model), and the kriging estimation. For the proposed method, the exploratory and structural analyses of a geostatistical procedure were used to derive a variogram model and its The MSE is the average value of the square differences between the measured and estimated values. The SMSE measures the consistency between the calculated variances and the estimate error variance (σ 2 i ). It is defined as the average of the square of the differences between the measured and estimated values over the estimate error variance, at the point of observation. A value close to one indicates that the model is adequate.
where n is the number of observations, Z(x i ) is the value of the property at point is the estimated value at point x i , and σ i is the estimated standard deviation.

Kalman Filter
The Kalman filter (KF) is an algorithm based on a set of mathematical expressions to obtain unbiased linear recursive estimates of minimum variance for a system with noise. The recursive term refers to the fact that the filter recalculates the solution each time a new value or measurement is included [31]. The static Kalman filter version used in this paper is implemented with the following equations: Measurement model: This links the measurement vector z n , which is presented in Equation (3), with the current state of the system C n through the measurement matrix H n and a Gaussian white noise v n with covariance matrix R n .
In the update phase, the Kalman gain (Equation (4)) is calculated to obtain the new state vectorĈ n+1 n+1 (Equation (5)) and its covariance matrix P n+1 n+1 (Equation (6)): where the subscripts with n are the initial arrays and vectors, the subscripts n + 1 are the arrays and vectors resulting from the first calculation, and the superscripts are the arrays of the current operation.

Traveling Salesman Problem (TSP)
Ouaarab et al. [25] define the TSP as the shortest journey for a seller visiting different cities, adding the distance between them and visiting each one, before returning to the place of departure.
Following Cárdenas-Montes [26], the TSP (Equation (7)) can be expressed as: x ij = 1 the route goes from city i to city j. 0 another route.
where x ij = 1 if city i is connected to city j; x ij = 0 otherwise. For i = 0, . . . , n, take c ij as the distance from city i to city j. Then, TSP can be represented as shown in Equation (8): The optimal monitoring route, obtained through the implementation of the TSP, is determined with the Branch and Bound method [39]. This method contributes to the design of monitoring networks where it is necessary to visit each monitoring network in situ for data acquisition.

Branch and Bound
This algorithm generates subsets of solutions (branching), which are pruned or discarded if they do not meet the values of the limits established to delimit the problem, until the optimal solution is found [39].
This can be used to solve an optimization problem c 0 (x) with an acceptable optimization domain X 0 , as seen in Equation (9).
where x represents a vector (x 1 , x 2 , . . . , x n ). The solution is said to be optimal if x satisfies all the constraints and belongs to X 0 and c 0 (x) is the minimum. However, if the problem c 0 (x) is very complicated, it can be broken down into simpler problems p j (branching) whose optimal solutions are x j . It is necessary to include bounded problems p k . The number of restrictions (m) changes according to the problem. This is represented as Equation (10), proposed by Lawler and Wood [40].

Objective Function for the Design of the Monitoring Network
The proposed objective function (OF) represented in Equation (11) considers two variables and their respective weight, which can vary depending on the importance given to each variable. These variables are the priorities of the monitoring wells based on their contribution to reducing the total estimate error variance for the groundwater levels (obtained with the KF) and the optimal sampling route (solving the TSP).
where w v and w r are the weights assigned to the variables PV i (the priority assigned to well i based on the reduction in the total estimate variance it produces) and PR i (priority given to well i based on the optimal distance to visit it and return to the starting point). The priority PV i assigned to each well was obtained with the static Kalman filter, following the method proposed by Júnez-Ferreira [31]: a PV i equal to one will be assigned to the well that produces the largest reduction in the total estimate error variance, while a N − n [MN] value (N is the total number of wells available for the MN design and n [MN] is the number of wells included in the optimal monitoring network (MN) will correspond to the last selected well). The TSP was used to obtain the optimal route to visit well i and returning to the starting point (the base). A PR i equal to one will be assigned to the well for which the minimum distance is required, while a priority equal to N − n [MN] will be assigned to the well with the longest route. The first well selected for the monitoring network will correspond to the position i for which the lowest OF value was obtained. For those cases where the sum is the same for two positions, then the well that provides more information in terms of the total estimate error variance reduction (with a lower PV i value) is selected. Each time a new well is included in the network, the covariance matrix is updated, and the method is applied to test the remaining positions (at this stage, the TSP includes visiting the previously selected wells and the one that is being tested).
The sampling time considered at each visited site was 0.5 h. An average speed of 40 km/h was also taken to travel through the communication routes. The wells to be monitored in a predefined region will correspond to those where the necessary time for monitoring (the sum of sampling and travel times) is lower than or equal to 8 h (a working day). If there is additional budget to invest in another working day, the method was repeated, starting with the final covariance matrix of the previous day, and a new routing was initiated. In cases where all the wells of the MN have already been integrated to the MN and the working day has not finished, the method stops when a priority is assigned to each of them. The TSP is solved with the Branch and bound technique [39], using the network modeling (NET) version 1 module of the WinQSB software [41] through a Windows XP virtual machine [42]. The OF was evaluated using Microsoft Excel Version 2021 [43].
The proposed method is explained in the flow diagram of Figure 3. To fully understand this, it is necessary to declare the following elements: i is the number assigned to each well that is tested (i = 1, . . . , N), n is the sum of the wells included in [MN] plus the i well with the minimum OF, r n is the distance of the optimal route to visit the wells that were already included in [MN] and the i well with the minimum OF, the vector [MN] includes the wells of the optimal monitoring network and [MN] f is the vector with the wells of the optimal monitoring network ordered following the optimal route determined with the TSP.
Process 1, included in Figure 3, consists of the following steps: 1.
For each i well that is not yet part of the [MN], the TSP is applied (including the previously selected wells) to compute the optimal route (r i ) from the starting point and back.

2.
A PR i is assigned to each well that is not part of the [MN], PR i = 1 for the i well with minimum r i , PR i = N − n [MN] for the i well with maximum r i .

3.
A PV i is assigned to each well that is not part of the [MN] following the Júnez-Ferreira method [31]; PV i = 1 for the i well that produces the largest reduction in the total estimate error variance, and PV i = N − n [MN] for the i well that was last selected.

Results and Discussions
The method was applied to the existing monitoring network of the study area, which consists of 49 wells. Evidently, a frequent visit to all the wells of the existing monitoring network would be time-consuming and would lead to high costs. The proposed method was designed to reduce the number of wells selected in monitoring schemes without losing valuable information and reducing costs.
Although the optimization was performed for separate zones, the variogram was constructed with all the available groundwater-level data for a better representation of the spatial correlation for the groundwater levels in the study zone. Finally, a global model is preferred to estimate the groundwater levels within the aquifer to maximize the contribution of all the data collected at the selected monitoring sites (not only those located within a zone). Furthermore, the calculation of separate variograms for each zone could lead to a poor representation of the spatial correlation due to the limited number of data available for each of them [44].
The method was applied by assigning a w v = 0.5 and w r = 0.5, which means that it was given the same importance to the reduction in the estimate error variance and the traveled distances.
As part of the geostatistical analysis, the available groundwater-level data were normalized. Statistics for these transformed data are shown in Table 1.  Figure 4 shows the values of the adjusted variogram model and its parameters (nugget, sill and range) that allowed for the lowest value of the MSE and an acceptable SMSE value (close to one) to be obtained. The large value in the range reflects the strong correlation between the normalized data.
x FOR PEER REVIEW 11 of 21 The value of the model range reflects that the correlation between the normalized data remains at large distances (almost reaching the extent of the estimation grid); how-   The value of the model range reflects that the correlation between the normalized data remains at large distances (almost reaching the extent of the estimation grid); however, the contribution of distant wells in the node estimation will be low, due to the screening effect.  The value of the model range reflects that the correlation between the normalized data remains at large distances (almost reaching the extent of the estimation grid); however, the contribution of distant wells in the node estimation will be low, due to the screening effect. Monitoring Schemes for Zones C1, C2 and C3 According to the proposed method, seven wells of Zone C1 were selected to be visited within an estimated time of 7.78 h. The selected wells following the optimal route are: 47, 6, 2, 4, 48, 9 and 8, as shown in Figure 6. For Zone C2, it is possible to visit nine wells with a calculated time of 7.77 h. The sequence for visiting these selected wells was 43, 18, 32, 13, 11, 12, 17, 15 and 16, as presented in Figure 7.

Monitoring Network Optimization
When applying the proposed method for Zone C3, nine wells were selected to be visited in 6.92 h; visiting 10 wells leads to a routing of 8.07 h, exceeding the restriction of 8 h (a working day). However, since the extra time corresponds to only 4 min, the latter case was accepted. The chosen wells following the optimal monitoring route are 40,31,28,29,26,27,25,24,23 and 49, as illustrated in Figure 8.
For the same case study, the Júnez-Ferreira method was applied (considering only the estimate error variance reduction), which resulted in the monitoring scheme presented in Figure 9a-c for zones C1, C2 and C3, respectively. Table 2 shows a comparison between the monitoring routes in the different zones, applying both methods. Following the proposed method, the travel distance was reduced by 28.95 km, 33.98 km and 18.96 km, in zones C1, C2 and C3, respectively, compared to the distances obtained with the Júnez-Ferreira method. An average reduction of 27.63 km was achieved; furthermore, more wells are visited when the proposed method is applied, which offers the advantage of collecting a higher number of field data with a lower cost.     Monitoring Schemes for Zones C1, C2 and C3 According to the proposed method, seven wells of Zone C1 were selected to be visited within an estimated time of 7.78 h. The selected wells following the optimal route are: 47, 6, 2, 4, 48, 9 and 8, as shown in Figure 6.
For Zone C2, it is possible to visit nine wells with a calculated time of 7.77 h. The sequence for visiting these selected wells was 43, 18, 32, 13, 11, 12, 17, 15 and 16, as presented in Figure 7.
When applying the proposed method for Zone C3, nine wells were selected to be visited in 6.92 h; visiting 10 wells leads to a routing of 8.07 h, exceeding the restriction of 8 h (a working day). However, since the extra time corresponds to only 4 min, the latter case was accepted. The chosen wells following the optimal monitoring route are 40,31,28,29,26,27,25,24,23 and 49, as illustrated in Figure 8.
For the same case study, the Júnez-Ferreira method was applied (considering only the estimate error variance reduction), which resulted in the monitoring scheme presented in Figure 9a-c for zones C1, C2 and C3, respectively. Table 2 shows a comparison between the monitoring routes in the different zones, applying both methods. Following the proposed method, the travel distance was reduced by 28.95 km, 33.98 km and 18.96 km, in zones C1, C2 and C3, respectively, compared to the distances obtained with the Júnez-Ferreira method. An average reduction of 27.63 km was achieved; furthermore, more wells are visited when the proposed method is applied, which offers the advantage of collecting a higher number of field data with a lower cost.
Considering a vehicle performance of 6 km/L and a fuel cost of 0.99 USD per liter, visiting the 26 wells selected with the proposed method represents a total cost in transportation of 70.15 USD. This increases up to 83.66 USD for visiting the 21 wells selected without optimal routing. The total travel distance for visiting the 21 wells in three working days was 507.07 km, whereas 425.18 km was necessary to visit the 26 wells selected with the proposal.    Table 2. Comparison between the different routes for the optimal Calera aquifer monitoring network, selected with the proposed and the Júnez-Ferreira methods.

Concept
Zone C1 Zone C2 Zone C3   Figure 10 shows the maps of groundwater-level estimates for the complete network (49 wells), the network obtained with the proposed method (26 wells), and the network obtained using the Júnez-Ferreira method (21 wells).  As expected, a more detailed description of the spatial configuration of groundwater levels within the study zone was obtained for an estimation using all the data collected with the complete network ( Figure 10a). It also provides the lower estimated error variances. The estimation using the groundwater-level data for the wells selected with the proposed method ( Figure 10b) reflects this flow pattern better than the estimates using the wells selected using the Júnez-Ferreira method (Figure 10c). Notably, the extended drawdown cone at the northeast area is not visible when the latter monitoring network is used. These differences are due to the different numbers and location of groundwater-level data used for each case; however, in the case of the northwest and southeast areas, the estimates produce similar spatial configurations for this variable.

Number of wells visi-
The estimate error variance maps for the complete network (49 wells), the network with the proposed method (26 wells), and the network with the Júnez-Ferreira method (21 wells), are displayed in Figure 11a-c. The variances obtained with the estimation using the wells for the three cases are very similar, which shows that the monitoring networks selected with the optimization methods are very effective. The estimate error variance maps for the complete network (49 wells), the network with the proposed method (26 wells), and the network with the Júnez-Ferreira method (21 wells), are displayed in Figure 11a-c. The variances obtained with the estimation using the wells for the three cases are very similar, which shows that the monitoring networks selected with the optimization methods are very effective.
The higher density of wells selected with the monitoring network obtained with the proposed method considerably reduces the estimation error variances in the valleys of the study zone (the middle strip that goes from south to north). These values are larger for the estimation using the data for the wells selected with the Júnez-Ferreira method. However, lower values are obtained around some wells located at peripheral areas of the estimation grid. These wells were discarded with the of the proposed method due to the increase in travel distance they produced.  Table 3 shows the statistics for the estimates using the selected monitoring networks for both approaches, and their comparisons with the estimates using the complete monitoring network. Taking the estimates using the complete monitoring network as a reference, the results show that the root of the total estimate error variance (the average standard error) increases up to 3.34 m when using the Júnez-Ferreira method and 3.35 m for the proposed method. The differences between both monitoring options for this value are negligible.
The mean difference, the mean square difference (MSD) and the root of the MSD values reflect the advantage of using a higher number of wells in the optimal monitoring network (26 vs. 21). Although, for the case of 26 wells, the maximum difference is larger The higher density of wells selected with the monitoring network obtained with the proposed method considerably reduces the estimation error variances in the valleys of the study zone (the middle strip that goes from south to north). These values are larger for the estimation using the data for the wells selected with the Júnez-Ferreira method. However, lower values are obtained around some wells located at peripheral areas of the estimation grid. These wells were discarded with the OF of the proposed method due to the increase in travel distance they produced. Table 3 shows the statistics for the estimates using the selected monitoring networks for both approaches, and their comparisons with the estimates using the complete monitoring network. Taking the estimates using the complete monitoring network as a reference, the results show that the root of the total estimate error variance (the average standard error) increases up to 3.34 m when using the Júnez-Ferreira method and 3.35 m for the proposed method. The differences between both monitoring options for this value are negligible.  The mean difference, the mean square difference (MSD) and the root of the MSD values reflect the advantage of using a higher number of wells in the optimal monitoring network (26 vs. 21). Although, for the case of 26 wells, the maximum difference is larger than that obtained for the other monitoring option, the minimum and maximum values are more balanced.
The maps of the differences between the groundwater-level estimates using the monitoring networks of 26 (proposed method) and 21 (neglecting optimal routes) wells, and the estimates obtained with the complete monitoring network of 49 wells are shown in Figure 12a, b, respectively. A larger area with low differences (between −10 and 10 m) was obtained for the monitoring network selected with the proposed method (26 wells).
For the monitoring network of 26 wells, the <−10 m differences cover a slightly larger area than that for the monitoring network of 21 wells. The extreme negative differences (<−19.83) that only exist for the monitoring network of 26 wells represent small portions of the southern, central and northwest. On the other hand, the areas for positive differences > 20 m are much larger when using the monitoring network of 21 wells. This result confirms the advantage of the proposed method in increasing the amount of collected data (five additional wells) to obtain a better spatial configuration of the groundwater levels.
These differences are associated with the coverage achieved by the wells selected for each monitoring scheme. Negative differences reflect that at least one well with a higher value than its neighbors was not included in the monitoring network (for example, well 1 is absent in Figure 12a). Conversely, positive values indicate that at least one well with a lower value than its neighbors was not included in the monitoring network (for example, well 31 in Figure 12b). each monitoring scheme. Negative differences reflect that at least one well with a higher value than its neighbors was not included in the monitoring network (for example, well 1 is absent in Figure 12a). Conversely, positive values indicate that at least one well with a lower value than its neighbors was not included in the monitoring network (for example, well 31 in Figure 12b).

Conclusions
The proposed method allows for the design of groundwater-level monitoring networks based simultaneously on spatial estimation and the routes to follow for data collection; it considers the contribution of the available wells in the reduction in the total estimate error variance over a study zone, while looking for the minimum travel distance, thus reducing monitoring costs. This method could be adapted to the design of groundwater quality monitoring networks by simultaneously using the covariance matrices of the groundwater quality parameters considered within the optimization procedure.
The obtained results show that a monitoring network designed with the proposed method considerably reduces the monitoring costs (visiting 26 wells in three working days) without a significant loss of the amount of information that is obtained with respect to the reference monitoring network (49 wells). Neglecting the travel distances in the optimization procedure leads to the selection of only 21 wells to be monitored over three working days, which implies that the proposal allows more field data to be obtained at the same cost. Furthermore, the higher number of sampling wells selected with the proposal produces a spatial configuration of groundwater levels that are more similar to those obtained with the reference monitoring network.
An important feature of the proposed method is its additive nature, which means that new wells can be incorporated into the monitoring network if the budget increases without eliminating the previously selected monitoring sites, favoring the continuity of the groundwater-level time series that is collected with them.
For the case study, it is assumed that the routes to follow correspond to the same type of communication route, in which a constant speed of the vehicle can be maintained during travel. Additionally, the same sampling time was considered for each visited well. Future works to improve the method could integrate specific aspects that produce variations in monitoring conditions, such as traveling on dirt roads, highways, ease of measuring, well depth, dead times, lunch time, etc. Another interesting option is the inclusion of the temporal component in the objective function, using time series information.
The monitoring network selected with the proposal very slightly increases the total estimate error variance compared to the case when the travel distances are neglected during the optimization, but produces estimates that are closer to the complete monitoring network. Additionally, the efficiency in the travel distance allows more wells to be visited in a working day, which implies a greater amount of direct (field) information at the same cost.
Future work should be oriented toward the development of a user-friendly software that integrates the different routines that are necessary for the optimization of the monitoring network, including a procedure to automatically generate the route matrix from a map with the different roads connecting the tested monitoring sites.