From (cyber)space to ground: new technologies for smart farming

Despite growing slower than in the recent past, the world population is projected to increase by more than one billion people within the next 15 years, reaching 8.5 billion in 2030, and to increase a further 9.7 billion in 2050 and 11.2 billion by 2100 (United Nations Department of Economic and Social Affairs Population Division 2015). Growth in population and income will imply a substantial increase in demand for water and food – not only due to a higher number of people, but also to trends towards more water demanding lifestyles and diets. The agricultural sector is going to face enormous challenges in order to sustain food production, which is required to increase by 70% by 2050.

Additional factors, such as climate change, will further contribute to affect water availability. Changes of average precipitation will not be uniform, with some regions experiencing increases, and others decreases, or not much change at all (Ravazzani et al. 2014a). According to climate projections, the Mediterranean area should be affected by a decrease of total precipitation with the exception of the Alps in winter (Coppola & Giorgi 2010). With earlier snow melting and rainfall variation, inter-annual run-off is changing towards less water during summer but more water during winter (Dedieu et al. 2014; Gaudard et al. 2014). This would negatively alter the current seasonal cycle of runoff, even in those areas where mean annual precipitation is expected to remain steady with negative impacts on agricultural production.

The increase in consumption of water resources, combined with climate change impacts, calls for new sources of water supply (Ravazzani et al. 2011) and/or different managements of available resources in agriculture. One way to increase the quality and quantity of agricultural production is using modern technology to make farms more ‘intelligent’, the so-called ‘precision agriculture’ also known as ‘smart farming’.

The scientific literature provides some studies focused on ‘smart farming’ by coupling meteorological and hydrological models (Gowing & Ejieji 2001; Cai et al. 2007). Ceppi et al. (2014) demonstrated that in-advance prediction of soil moisture (SM) and crop water requirement allows a precise irrigation scheduling with benefits on farmer income in terms of reduction of water consumption and increase of crop yield. However, their investigation was funded on local analysis in one single cultivated site where ground measurements of meteorological and hydrological variables were acquired hourly. Moreover, they used only temperature forecast to predict evaporation by applying an empirical model (Ravazzani et al. 2012). Open issues still remain about how to extend application to larger areas, and how physically based methods that are fed the complete set of meteorological variables can improve SM forecast.

Spatially distributed, physically based hydrological models, with their ability to estimate energy and water fluxes at the agricultural district scale, are invaluable tools for water resources management for agricultural water use (Corbari et al. 2015). Satellite data, for their intrinsic raster structure, can be effectively used for the internal calibration/validation of distributed hydrological models in each pixel of the domain. This can be achieved with those models based on energy and water balance algorithms in combination with remotely sensed data, in particular of land surface temperature (LST) (Corbari & Mancini 2014).

The information content of satellite images may be useful not only in providing a temporal dataset to calibrate and validate hydrological models, but even for assessment of biophysical attributes, such as leaf area index (LAI) (Colombo et al. 2003) and surface albedo (Corbari et al. 2014), and their temporal variation (Mattar et al. 2014). Thus, the combined use of physically raster-based hydrological models and satellite data may be an answer to the question about extending prediction to larger ungauged areas.

However, not all necessary information can be derived from satellite-based Earth observation. For example, vegetation height is an important piece of information, but it is rarely used due to challenges in its extraction. Therefore, crowdsourcing becomes a valid, integrative source of information, leveraging on the popularity of smartphones and tablets. Examples of applied crowdsourcing can be found in different topics, from fire mapping (Goodchild & Glennon 2010) to risk management purposes (Bevington et al. 2012).

Sophisticated physically based hydrological models need more meteorological variables to compute water and energy fluxes. Besides the fact that full meteorological observations are not always available with sufficient spatial density, questions arise about reliability of meteorological prediction by weather forecast models that are needed for SM and crop water requirement forecast (Ceppi et al. 2013). Many studies have been devoted to analyze the accuracy of precipitation forecast and performance of hydro-meteorological coupled systems, mainly for the purpose of flood forecasting (Amengual et al. 2008; Rabuffetti et al. 2008; Ceppi et al. 2010; Pianosi & Ravazzani 2010; Senatore et al. 2015; Arnault et al. 2016; Larsen et al. 2016). However, accuracy of the forecast of other meteorological variables except precipitation and performance of meteo-hydrological systems for SM forecast still need in-depth investigation.

The aim of this paper is to assess how mathematical models for weather and hydrological simulations, together with new technologies in the field of Earth observation from space and technologies for getting information from cyberspace (crowdsourcing), can help managing irrigation scheduling in a rich cultivated area in northern Italy. This work was part of the SEGUICI project, an Italian acronym that stands for smart technologies for water resources management for civil consumption and irrigation.

The Muzza canal (about 40 km long) derives water from the Adda river at Cassano d'Adda and it flows back into the Adda river close to Castiglione d'Adda. Along the canal there are 38 intakes and many more hydraulic nodes; the entire Muzza network is composed by open earth canals. The Muzza is both the largest irrigation canal by capacity and the first artificial canal built in northern Italy.

Average annual rainfall in the MBL consortium ranges between 800 (southern area) and 1,000 mm (northern area) with two peaks in spring and autumn (Ceriani & Carelli 2000). Winter is generally cold with a mean monthly temperature of 2 °C in January and summer is hot and humid with a mean monthly temperature of 23.4 °C in July (ERSAF 2004). Evapotranspiration (ET) amounts can reach up to 500 mm during the summer season, therefore, most of the water supply for agriculture comes from the irrigation network.

In the period 2010–2012 one test-site of the Pre.G.I. project (Prediction and Guiding Irrigation) (Ceppi et al. 2014) was located in the central area of the basin at Cascina Nuova farm, in Livraga town. Here, one meteorological and one eddy-covariance station and time-domain reflectometry (TDR) probes were installed to measure mass and energy exchanges between soil, plant and atmosphere (Masseroni et al. 2012, 2013; Corbari et al. 2013).

In 2015, the monitoring station was moved from Livraga to Secugnago site (Figure 1). In both the monitored fields, farmers cultivated corn and flood irrigation was scheduled by the MBL consortium according to planned water allotments that were determined in advance. Landowners cannot irrigate their fields on days other than the scheduled ones (the Italian name of this irrigation scheduling method is turno irriguo). On average, farmers can irrigate fields once every 2 weeks.

Specific field campaigns were performed in order to characterize soil properties. Soil water retention curve parameters for Livraga and Secugnago are reported in Table 1. Sampling points were selected randomly within both study sites. Samples were collected from different depths. The parameters presented in Table 1 are average values. Particle size distribution for each soil sample was determined by wet sieving and hydrometer method. Sand, silt and clay contents together with soil texture were identified according to the United States Department of Agriculture system of soil classification. Undisturbed soil samples were used to measure the saturated hydraulic conductivity following the falling-head method (Lee et al. 1985). Soil water retention curve parameters were defined through evaporation method experiments (Wendroth et al. 1993) using the Hydraulic Property Analyzer device. Results were afterwards fitted to the Brooks & Corey (1964) parametric equation.

In order to obtain a spatial estimation of some biophysical parameters (normalized difference vegetation index, NDVI; LAI; fractional cover, FC; albedo) and of the hydrological model variable LST over the Muzza basin, remote sensing data acquired from the moderate resolution imaging spectroradiometer (MODIS) were chosen, and in particular, two types of surface reflectance data (from Collection-5 MODIS/Terra Land Products) used, namely, MO09GQ and MOD09GA, both already atmospherically corrected for vegetation parameters' retrieval. Data from the MODIS near real-time (NRT) context were used.

As regards MOD09GQ, surface reflectance data in Bands 1 (red spectral range) and 2 (near infrared spectral range) were used, daily provided at a 250-m spatial resolution. As regards MOD09GA, surface reflectance data from Bands 1 to 7 (from visible to infrared spectral range), daily provided at a 500-m spatial resolution, and a specific dataset of reflectance data state quality assurance (to generate cloud-cover masks), daily provided at a 1,000-m spatial resolution, were used.

First, by means of a binary mask (0-1) identifying the study area and a land-cover map of Lombardy region, a time-invariant mask referred to Muzza basin was created, wherein a numerical value from 1 to 3 was assigned to every pixel, thus carrying information about the corresponding land-use class (i.e., crops, grasslands and agro-foresty areas; pixel not falling under these three classes were set to NaN).

MOD09GQ and MOD09GA products were initially converted from their original sinusoidal projection to UTM Zone 32N WGS-84, with a pixel size of 250 m, by using MODIS Reprojection Tool in batch mode.

In order to improve parameter estimation quality, a time-variant cloud-cover mask was daily created from the reflectance data state quality assurance dataset. Then, spatial maps of NDVI, FC, LAI and albedo were created for that day-of-year (DOY) throughout the study area.

The LST variable was also derived from NRT satellite imagery, for which MOD11_L2 product was used with a 1,000-m spatial resolution.

The model was set up using single-moment 6-class microphysics scheme (WSM6) containing ice, snow and graupel processes (Hong & Lim 2006), the Noah land surface model scheme (Tewari et al. 2004), the PBL Yonsei University (YSU) scheme (Hong et al. 2006) and the CAM scheme for radiation (Collins et al. 2004).

To improve the estimation of the initial values, observations were assimilated in the model using WRFDA system (Barker et al. 2012) with 3DVAR techniques (Barker et al. 2004). Data taken in were derived from NCEP database and include: satellite radiances in BUFR format and conventional observations from land, ocean and upper-air platforms in PREPBUFR format.

Other weather data (temperature, wind speed, wind direction and pressure) were taken from meteorological stations of the Meteonetwork database.

Finally, albedo and LAI data derived from satellite observations of land cover were used to replace standard values in WRF simulations for the higher resolution domain.

Two distributed hydrological models were used for simulating the water balance components: the flash-flood event-based spatially distributed rainfall–runoff transformation, including water balance (FEST-WB) (Rabuffetti et al. (2008)) and the flash-flood event-based spatially distributed rainfall–runoff transformation, including energy and water balance (FEST-EWB) (Corbari et al. (2011)). The primary difference between them is in the computation of ET. The FEST-WB model derives the actual ET by rescaling the potential ET using a simple empirical approximation, where the potential ET is computed based only on air temperature measurements (Ravazzani et al. 2012, 2014a). In contrast, the FEST-EWB model computes the actual ET by solving the system of water mass and energy balance equations (Ravazzani et al. 2014a). The differences in the input parameters and meteorological forcings are listed in Table 2.

Both models discretize the computation domain with a mesh of regular square cells (200 × 200 m in this study), in each of which water fluxes are calculated at hourly time step.

The idea is to let everyone collect useful information, even contributors without a specific background; therefore, pictures of the most commonly found vegetation species are included as examples, to guide the end-users in deciding what species they have just taken a picture of. The mobile app allows including the height of vegetation, directly related to the stage of growth, and if the field is flooded or not, useful information to know whether the farmer is irrigating the field.

Every collected report, which includes a geocoded and oriented picture and answers to the above-mentioned group of questions, is automatically uploaded and stored in a remote database (Galeazzo et al. 2015). To avoid weighing on the contributor's mobile data quota, an option can be activated to store reports on the hand-held device and upload them only when a WiFi connection becomes available. A webgis interface is used to display data on an OpenStreetMap-based map (Figure 3(b)). Within the server, an algorithm can automatically associate the geo-localized reports with polygons related to each single field using Global Positioning System (GPS) position and compass direction (Figure 3(c)). Cooperation is in progress with the Research Support Service of the European Space Agency to share the collected data for their possible use in validation of land cover/use information derived from Earth observation satellite datasets.

The FEST-EWB and the FEST-WB models were calibrated and validated following different procedures. In fact, the FEST-EWB model was calibrated distributed by comparison of simulated LST with the observed ones and validated against local SM and ET, whereas the FEST-WB model was calibrated locally and SM and ET were measured.

The 2010–2011 period was used to calibrate and the 2012 to validate the FEST-WB model against SM and ET observations acquired at Cascina Nuova field in Livraga. Only values of the parameters of the cell corresponding to the station site could be calibrated as there were no other stations with similar capabilities available in the consortium. Figure 7 shows the comparison between observed and simulated SM and ET, along with rainfall and irrigation amount, during the three growing seasons of 2010, 2011 and 2012.

In general, satisfactory results are found in terms both of SM and ET during calibration and validation periods. More details and comments can be found in Ceppi et al. (2014).

The FEST-EWB model was calibrated during the period 2010–2012 against observed MODIS LST. Hence, soil hydraulic and vegetation parameters were calibrated in each single pixel minimizing the difference between the observed and simulated land surface temperatures, following the procedure developed by Corbari & Mancini (2014) and Corbari et al. (2015). For the entire dataset of 166 images, statistical parameters between LST from calibrated FEST-EWB and LST from MODIS were computed: mean absolute error (MAE) is equal to 0.2 °C, root mean square error to 1.8 °C, relative error (RE) to 4.2% and the Nash & Sutcliffe (1970) index to 0.73. Cities areas were discarded from the comparison. In Figure 5, as an example, for 27 August 2012 at 13:00, MODIS LST and FEST-EWB LST images before and after the calibration, are shown.

SM and ET estimates from FEST-EWB and FEST-WB were then compared at local and basin scales for 2015. The simulations of both models for the 2015 growing season were performed without a local calibration, but using their own parameters previously calibrated for 2010. Hence, FEST-WB soil parameters were only locally (e.g., Livraga) calibrated, while FEST-EWB ones were calibrated in a distributed way for each pixel of the analysed domain.

Observed ET at Secugnago site was available concurrently to model simulations only from Day 154 (3 June) to Day 171 (21 June) due to station malfunctioning. Cumulated ET from FEST-EWB and from FEST-WB were then compared with observed values until 21 June and a RE equal to 0.69% and to −7% was obtained, respectively (Figure 6). The difference between the two models in computing ET over the whole growing season was equal to 42.7 mm.

The same comparison was then performed for ET maps. In Figure 7, as an example, the FEST-EWB and FEST-WB maps are reported for 30 September 2015 at 12:00. ET spatial mean and standard deviation are equal to 0.13 mm and 0.05 mm for FEST-EWB, while for FEST-WB they are equal to 0.037 mm and 0.01 mm.

When the entire simulation period is considered, the mean temporal differences of the spatial mean is equal to 1.1 mm. These differences are due to different modelling schemes on ET and calibration procedures; in particular, the FEST-WB was calibrated at local scale only, while the FEST-EWB was calibrated pixel by pixel at basin scale.

In order to assess how crowdsourcing data may affect accuracy of water balance, SM and ET were simulated with FEST-EWB at the Secugnago site according to three scenarios:

• 1. Vegetation weight was changed according to crowdsourcing information acquired with the smartphone application, LAI and albedo were retrieved from remotely sensed images, and crop minimum stomatal resistance (rsmin) was set to 150 s/m. This is the reference scenario whose results were presented in previous sections.

• 2. We assume the field was grass cultivated, and all vegetation parameters were assigned for a grass crop: height = 0.12 m, LAI = 1, rsmin = 70 s/m.

• 3. We assume the field was grass cultivated, height = 0.12, rsmin = 70 s/m, but LAI and albedo were taken from remotely sensed images.

The WRF meteorological model was daily launched from 3 June 2015 to 30 September 2015 in order to obtain weather forecasts over the two areas of study during the 2015 growing season. The main meteorological fields available to feed the FEST hydrological models were: air temperature and relative humidity, incoming shortwave solar radiation, precipitation and wind speed.

Table 3 highlights the performance of the WRF model forecasts in comparison with observed data for the entire forecast horizon of 9 days over the Secugnago site. The forecast of the day +0, i.e., the forecast of the same day of the model initialization, is here omitted, since it would not be exploitable for irrigation scheduling management.

Satisfactory results were found over the Secugnago test-bed during the 4 months of simulations. In fact, air temperature forecasts maintained a bias of about 2 °C for the entire forecast horizon; in particular, the WRF model tended to underestimate temperatures and even relative humidity forecasts were 6–8% below the observed data; in contrast, the incoming solar radiation, wind speed and daily precipitation were overestimated by the WRF model at about 80–90 W/m², 1.6–1.8 m/s and 2–7 mm, respectively. In general, no outliers were found during the analysed period and no significant decrement of the WRF performance at increasing of lead-time was present.

Weather forecasts were afterwards used to drive the FEST hydrological model simulations using the two schemes (ET computed with Hargreaves equation, FEST-WB; and ET computed by solving the energy balance, FEST-EWB) for calculating SM at Livraga and Secugnago sites. Goodness of forecast was assessed by computing literature fit indexes comparing SM simulated by FEST-WB and FEST-EWB fed with observed meteorological forcings and SM obtained with the same hydrological models fed with meteorological forecasts.

As shown in Tables 4 and 5, a better correlation (R²) was found using the energy-balance model in both the two sites, Livraga and Secugnago, respectively; however, the MAE for SM shows fairly good results using both the Hargreaves and energy-balance equations during the entire forecast horizon, also due to a good performance of weather forecasts previously described; acceptable values, in fact, were found between 0.01 and 0.03 from day +0 to day +8, respectively.

The benefit of having a good coupled hydro-meteorological system many days in advance can be summarized in the following picture where a reanalysis for the period 22 June 2015 to 15 July 2015 is shown. In this time span, according to the MBL consortium regulation, irrigation was scheduled on 30 June 2015 and 14 July 2015.

This work was part of the SEGUICI project, the aim of which was to develop and integrate smart technologies for water resources management for civil consumption and irrigation.

The aim of this paper was to assess how mathematical models for weather and hydrological simulations, together with remotely sensed images and crowdsourcing, can help in managing irrigation scheduling, by forecasting SM and crop water requirement. The test beds of the project were two maize fields at Livraga (2010–2012) and Secugnago (2015) in the MBL consortium, about 50 km south-east of Milan in northern Italy.

The SM forecast was accomplished by coupling a meteorological model with the FEST hydrological model. Numerical weather predictions were provided by the WRF-ARW meteorological model with a 10-day lead-time. Two configurations of the FEST distributed hydrological model were tested: the FEST-WB scheme that computes ET with the Hargreaves equation, and the FEST-EWB that solves the energy balance equation.

The FEST-WB model was calibrated against SM and actual ET measured at Livraga station during the 2010–2012 campaigns. Only parameters of the cells surrounding the Livraga station could be calibrated as no other measurements were available in the MBL area. The FEST-EWB model was calibrated during the period 2010–2012 against observed MODIS LST. The two models were further validated against SM measured in the 2015 campaign at Secugnago. Comparisons with observations show that, while FEST-EWB was able to properly simulate SM and ET, FEST-WB, which was not calibrated at Secugnago, showed greater error. Moreover, the comparison of spatial distribution of SM and ET computed by FEST-WB and FEST-EWB showed significant differences due to different methods used for their calibration. Calibration using remotely sensed images is an effective alternative to ground-based observations and provides spatially distributed information impossible to acquire with conventional technologies.

Crowdsourcing resulted in a fundamental source of information that could increase the accuracy of water balance simulation, with maximum advantage occurring when combined with remotely sensed information.

The performances of numerical weather predictions were assessed against air temperature and relative humidity, incoming shortwave solar radiation, precipitation and wind speed observations at Secugnago. Air temperature forecasts maintained a bias of about 2 °C for the entire forecast horizon; in particular, the WRF model tended to underestimate temperatures and even relative humidity forecasts were 6–8% below the observed data. In contrast, the incoming solar radiation, wind speed and daily precipitation were overestimated by the WRF model at about 80–90 W/m², 1.6–1.8 m/s and 2–7 mm, respectively.

Weather forecasts were afterwards used to drive the FEST-WB and FEST-EWB models for forecasting SM at Livraga and Secugnago in the 2015 campaign. Goodness of forecast was assessed by computing literature fit indexes comparing SM simulated by FEST-WB and FEST-EWB fed with observed meteorological forcings and SM obtained with the same hydrological models fed with meteorological forecasts. SM forecast was reasonably satisfactory no matter whether the FEST-WB or the FEST-EWB was used. Moreover, results showed how combing meteorological and hydrological model that were correctly calibrated, it was possible to get reliable SM forecasts for up to 1 week, and this helped farmers to properly decide irrigation scheduling.

This work was sponsored by the Lombardy region in the framework of the SEGUICI project. We thank ARPA Lombardia (http://www.arpalombardia.it) and the Meteonetwork Association (http://www.meteonetwork.it) for providing meteorological observations from automatic stations. The editor, Prof. Attilio Castellarin, and three anonymous reviewers are gratefully acknowledged for their efforts to improve the quality and contents of this manuscript.