The study was carried out in the Northwest of Cruz del Eje department (Figure 1), located approximately 150 km north from Cordoba city. The area has a warm and humid climate, with more than 300 days of sunshine per year; with a rainy season that extends from October to March and a dry one between April and September. The summer is warm and humid, while the winter is dry and not very harsh. Thus, in the extreme south, at higher altitude, the average maximum temperature is 14 oC and the precipitations exceed 800 mm, while in the Salinas, the thermal values exceed 27 oC and those of precipitation are less than 400 mm 9.

A dike supplies water for irrigation (Lat. -30° 64' and Long. -64° 30'). The maximum height of the reservoir is 111,985 Hm3. Potential irrigation is 12,050 has. The average annual rainfall of the basin is around 580 mm, the surface of the basin of contribution is 1,700 km². The system has two margins of Cruz del Eje River and 16 Channels 10. The soils have a frank-silty structure. They have a little evolved structure and the content of organic matter oscillates between 1.5-3% that decreases considerably in the inferior horizons. The existing water deficit in the region causes a huge evapotranspiration and low effective humidity in the soil profile 9, 11.

Throughout the year, irrigated crops have different characteristics among each other and from rain feed ones, either because of its phenological growth, or because of the management they are subject to. Based on this idea, Landsat OLI 8 satellite images from three dates (each agricultural year) were selected to explore the annual cycle according to the management and phenology. A winter date (June-July), and two spring-summer dates (October and December-January). It was done for 3 agricultural cycles including then the following image data set: 2013-07-28, 2014-01-04, 2014-06-29, 2014-08-16, 2014-12-06, 2015-03-28, 2015-06-16, 2015-08-03, 2015-09-04, 2015-10-22, 2015-12-08, 2016-01-18, 2016-05-01.

All these products were calibrated to reflectance. For each image, vegetation spectral indexes (NDVI, NDWI and SAVI) were calculated in order to observe which of the indexes is most useful for the identification of irrigation area. Finally, only the NDVI (Normalized Difference Vegetation Index) 12 was used because it is a good estimator of the photosynthetic activity of the canopy of vegetation and, therefore, of primary productivity. Many studies have used the NDVI to identify types of crops in areas under irrigation 13.

Specifically, to detect the presence of crops under irrigation, a map of agricultural plots (2,523 polygons) was built (e.g. 14,15,16,17) through visual digitalization at 1:25,000 scale using Google Earth® images. For each digitized plot, the average value of NDVI calculated on the dates analyzed (average of pixels within the polygon) was assigned. An example of the product obtained following this procedure is presented in Figure 2. This methodology allowed us to identify the distribution of agricultural parcels with different levels of primary productivity and their temporal behavior.

Then, based on the NDVI values assigned to each polygon, those that presented their annual maximum (constructed from 3 annual images) lower than the minimum NDVI value of the polygons on the ground truth irrigated fields (information provided by rural extension agency of Cruz del Eje) were filtered. For validation, another 14 horticultural plots were taken as a test set, yielding 100% accuracy. (Figure 3).

For the analysis and typological classification, we adopted the definition generated by INTA in their project entitled Water management and irrigation for sustainable development of the territory (Gestion del agua y el riego para el desarrollo sostenible del territorio, PE 1133044) on actual and potentially irrigated areas.

A. Irrigation region: This term refers to a geographical space where the socio-economic development results from the management of its available water resources, covering aspects of production, distribution, protection, in balance with the social, historical, cultural and economic issues. It is represented by a surrounding polygon that delimits land with agricultural / livestock use and within which the dominant class is irrigated crops.

B. Irrigable area: It corresponds to all that surface within an irrigation region that has potential to be irrigated. To define irrigable areas, the following criteria were considered: i) access to water, ii) irrigation system in conditions to be used, iii) irrigation aptitude. In this case the "cadastral domain" is stipulated as a unit of analysis and work.

C. Actually irrigated area: Corresponds to the lands that are being irrigated at a specific time. Thus, the area irrigated in a territory during spring is different from the area irrigated during summer, since we find different crops with significant variations of phenological cycle and agricultural management. These areas will be within the irrigable area and may coincide with it or be smaller.

The actual field data of horticultural farmers (polygons/plots) is obtained by grouping three different data sources:

Four classifiers were explored on a Sentinel 2A image, pre-processed to surface reflectance. We worked with the bands blue B2, green B3, red B4, three red edge bands B5, B6, B7, near infrared (NIR) B8 and two SWIR bands B11 and B12. All of them were resampled to a spatial resolution of 10 m. The image corresponds to June 7, 2016 presenting cloudiness percentage of less than 40%. The campaign to obtain ground truth points was carried out on July 19, 2016.

We classified the image in the software ENVI 4.8 © (CONAE license). Two unsupervised classifications were used (ISODATA and K-MEANS) instantiated with default values, assigning 5 to 15 classes in the search with 400 iterations (ISODATA) and 10 classes in the same conditions (K-MEANS). The urban area of Cruz del Eje was masked out. The unsupervised classifications are those in which the classifying algorithm does not need more information than the image and some parameters that limit the number of classes. These classification mechanisms search for classes with sufficient spectral separability to differentiate some elements from others 19,20,21. These classifications were used as a first approach to the spectral characteristics that allow to differentiate coverages.

In terms of supervised classification, the Support Vector Machine (SVM) algorithm of ENVI 4.8 © was used, with a Radial Basis Function type kernel. The algorithm was trained with field data taken with GPS on June 19, 2016; with 1089 pixels (57 polygons) to determine 5 types of coverage: Water and artificial cover (construction); Seminatural (pastures, shrubberies, buffel grass not irrigated, stubble-corn and cotton); Bare soil (bare soil and plowed earth); Irrigated herbaceous (horticultural, alfalfa and winter green); Irrigated Trees (pomegranates, olive trees and water boundaries trees).

In order to determine the accuracy of the resulting maps, we computed the confusion matrices, using 30% of the ground truth points (728 pixels in 25 polygons) as a validation data set. The Global Accuracy and Kappa Statistic (K) 21 were calculated. The kappa coefficient computes the agreement between the classified image and the ground truth, due solely to the accuracy of the classification, suppressing the agreement that could be expected simply by chance 22. In 23 a scale was proposed for the interpretation of the K where values lower than 0.40 are considered a poor classification, 0.40-0.55 sufficient, 0.55-0.70 good, 0.70-0.85 very good and greater than 0.85 excellent. We used the caret library 24 from R software 25 ensure that the distribution of classes is the same in training and test data sets.

The other supervised classification model applied is Random Forest (RF), an automatic algorithm in which decision tree models are iteratively adjusted to random subsets of the input data and use the combined result for prediction 26. With the processed image and the set of real field polygons; a RF model was applied. We used the R random Forest package 27 that implements the classic Breiman algorithm 28. (Figure 5)

On the base of the available data sources and the geographic link between them, we can obtain 3 GIS layers: i) irrigation region delimited by the official province administration presented in Figure 3, ii) irrigable area that is the union of census and registration forms (Layer A + Layer B) presented in Figure 4, and iii) actually irrigated area, presented in Figure 5, which is generated by the intersection between digitized agricultural parcels and irrigable area.

Next, the adopted criteria for the delimitation of the proposed products are described with more detail:

In Figure 9 we present a subset of the classified image using non supervised algorithms (ISODATA y K-Means) and supervised ones (SVM and RF), and in addition the real color image of the same subset. (Figure 9).

In the figure we can see how ISODATA and K-Means were able to detect different coverages within the plots. In the case of K-Means, it may be due to the greater number of classes, but for the ISODATA and SVM comparison, with the same number of classes, the first one is better. For a first approximation to the land cover, the classification with ISODATA was useful and allows us to define distinct classes.

The results from the confusion matrix of SVM seem promising and they represent an operational method to identify the amount of irrigated area, even with few field data. This result of global precision would improve if we increase the ground truth data; and thereby define other more specific interest classes. The final classification for the whole region obtained with RF is presented in Figure 10.

(Table 2) shows the values of the Kappa index and the overall accuracy (in percentage) obtained during the validation of supervised SVM and RF classifications.

The operational advantage of the SVM against RF (which showed better results) is something to be evaluated by the operator. If an algorithm is implemented as an R script in the QGIS environment, it can remedy the accessibility of this powerful classifier, while the processing cost is greater in RF compared to SVM.

Within the established irrigation region, the irrigated herbaceous (horticultural, alfalfa and winter green) area is 4,948.8 ha (Table 3). Another important area for the irrigation system are the Trees, that in this classification is built by fruit trees and water boundaries trees (those that are on canals, and in the river boundaries), this area is approximately 6,535.61 ha (Table 3).

Using a characterization including the official datasheets and remote sensing we can present a simple approach with which we can say that the actually irrigated area in the Cruz del Eje System is approximately 4,717 ha in contrast with the 11,450 historically reported as irrigable area, and the 2,680 ha that is officially reported for year 2016 with the proposed method 1.

This study based in an enrichment of that methodology propose for INTA 1, the identification of the irrigated area from non-irrigated, is based on the spectral behavior of the result of irrigating the land, which is the green vegetation 30. The proposed methodology using basic image processing procedures validated the choice of dates for the seasonal and yearly characterization of the area. It also confirmed the efficiency of NDVI as a descriptive index of great utility in this type of studies for irrigated crops detection. Like the conclusions in 31 this work indicated that NDVI can provide robust field-specific and regional estimates of phenology and water use for horticultural crops with minimal requirement for supporting information.

The improvement offered by the manual segmentation 14 of plots to the identification of actually irrigated areas is promising and can be incorporated into this method automatically using plugins in QGIS, GRASS, R and other open source software.

The results of the surveying, characterization and mapping of irrigated and / or potentially irrigable areas can provide products that allow the elaboration of analysis maps referring to the current use of water for irrigation purposes as a descriptive memory and management tool.

This work demonstrates that even without all the institutional field database needed by the methods 1, it is possible to provide an estimation of actually irrigated area with free satellite imagery and processing, with a supervised classification of Sentinel 2, in open source programs.

The importance of this study is to increase the knowledge and propose a constant update of the amount of hectares irrigated. Ina system such as Cruz del Eje, which manages limited water flows and that, its inefficiency results in an environmental and economic cost that affects the sustainability of food production.

This method is easy to adopt, given that the processing tools and source data are freely available and we currently have products processed almost immediately upon acquisition. For example, Sentinel 2 is available for free as calibrated surface reflectance. The processing validated by this work sets precedents for a systematic follow-up of the amount of hectares with just one campaign of ground truth per field to be classified and a script that is available in GitHub for R 32 and that can be operationalized from the QGIS environment.

Different institutions, the General Directorate of Irrigation (DGI), the National Institute of Agricultural Technology (INTA); as well as the community management of water resources for irrigation, find in this methodology a response to their demands for information that is fundamental for water management and technical support.

Once the areas have been established according to the method proposed by 1, the area that must be updated at least annually is the "actually irrigated area". This paper provides a simple method, validated by data intercrossing, remote sensing and multitemporal analysis. RF is mainly characterized by its ability to estimate the importance of each predictive variable in the modeling of the response variable. Works such as Strobl et al. (2007, 2008) 29 found that the random forest package produces poor estimates in certain scenarios such as those analyzed here. The package RF provides a solution that uses conditional inference trees and estimates of importance, which makes it an attractive alternative to others.