Internet of Things for Hydrology:

Potential and Challenges

Andrea Zanella Sergio Zubelzu Mehdi Bennis
Department of Information Engineering Departamento de Ingenierı́a Agroforestal, University of Oulu
University of Padova Universidad Politécnica de Madrid Oulu, Finland
Padova, Italy Madrid, Spain Email: [Link]@[Link]
Email: [Link]@[Link] Email: [Link]@[Link]

Martina Capuzzo Paolo Tarolli

Department of Information Engineering Department of Land Environment Agriculture and Forestry
University of Padova University of Padova
Padova, Italy Padova, Italy
Email: [Link]@[Link] Email: [Link]@[Link]

Abstract—The management of water resources has always been from observations, thus lacking generality, adaptability and
important for the sustainability of our society and economy. This interpretability.
need has been further increased by climate change in recent years In this scenario, the Internet of Things (IoT) technologies,
that, among other effects, has led to an increase in extreme events,
such as prolonged droughts, severe storms, hurricanes, and so together with machine learning (ML) methodologies, are in-
on. It is therefore urgent and critical to develop new and more strumental in developing more accurate hydrological models
sophisticated tools and methodologies to observe and possibly and more effective water management strategies and interven-
predict fundamental water processes. Internet of Things and tions. More specifically, IoT and ML can bring innovations
machine learning can provide a significant contribution to this in three main directions, namely: (i) data collection, i.e.,
end, which requires bridging the gap that still exists between the
communities of hydrologists, data scientists, and communications the measurement and gathering of relevant environmental and
engineers. This article aims to help fill such a gap by introducing weather data; (ii) data processing, i.e., the analysis of the col-
engineers to the challenges of hydrology, and reviewing existing lected data to refine hydrological models and develop accurate
solutions proposed in the literature to such challenges. Some predictions of extreme events; and (iii) data visualization, i.e.,
results obtained from empirical data sets are used to illustrate methodologies to ease the understanding of the models and
the main concepts and corroborate the theoretical discussion with
some practical examples. Finally, open problems and possible related inferences by means of graphical tools.
avenues for future research are discussed. In this paper, we discuss the state of the art in these three
fundamental areas and the many open challenges that still need
I. I NTRODUCTION to be addressed to improve current hydrology models. We
Recently, many European countries have experienced the focus on the problem of determining the behavior of the water
effects of climate change in the form of a scarcity of drinking level along the main discharge river of a catchment during
water resources, prolonged periods of drought, and extremely and after a storm. This turns out to be a critical aspect for
heavy rainfall, with unprecedented dramatic environmental, the control of water locks in order to manage tributaries and
economic, and social costs. Therefore, understanding, mod- the main channel and avoid, or limit, flood events. Thus, a
eling, and predicting the movement and distribution of water proper modeling and prediction of this process is necessary.
on Earth and effectively managing water resources is more We stress out that this is a problem more complex than just
than ever vital for agriculture, industry, and society at large. precipitation estimation, which is only one of the elements that
Unfortunately, hydrology involves atmospheric, surface, and determine the level of the water in basins. Currently, one of
underground water systems, which are difficult to model on the most effective techniques for weather prediction is based
their own, and even more so when considered as a whole. on weather radars, which provide models with high density
As a result, modern hydrology often relies on a number of spatial information on precipitation. However, they have the
mathematical and empirical models that focus on isolated disadvantage of requiring high investment and operation costs
portions of the whole water cycle, thus providing only partial, and would only provide a partial answer to the estimation of
defective and oftentimes inconsistent information. Moreover, the variability of hydrological processes. Indeed, since there
such models are either based on complex physical theories is no equivalent measurement technique for the remaining
that involve a large number of variables and parameters, processes causing spatio-temporal variability, as for example
which are often difficult to observe in practice and do not infiltration/runoff or flood routing, their benefits remain un-
have a clear physical meaning, or are empirically obtained complete.

© 2023 International Federation for Information Processing (IFIP). ISBN: 978-3-903176-56-0

 Fig. 1: Hydrological processes within a catchment.

This paper provides the following contributions:

(i) a high-level introduction to the fundamentals of hydrol-
ogy to equip ICT practitioners with the basic knowledge
to understand (and appreciate) the various challenges
offered by the hydrological domain;
(ii) a review of the main solutions proposed in the literature
regarding the problems of hydrological data detection,
processing, and visualization; Fig. 2: Example of hyetographs for two different days in the
(iii) a selection of results obtained in real-world scenarios to Madrid basin. Each measurement refers to the amount of water
ground the discussion and exemplify possible solutions; falling in one hour.
(iv) a reasoned discussion of the open challenges to be
addressed and the possible approaches that can be applied
to tackle such problems.
The rest of the paper is organized as follows. Sec. II provides
a very compact introduction on basic hydrological processes.
Then, Sec. III, IV and V discuss the data collection, process-
ing, and visualization solutions, respectively. Finally, we con-
clude the paper with a discussion about possible approaches
to tackle the open challenges in Sec. VI.
II. BASICS ON HYDROLOGY
Fig. 3: Example of spatial interpolation techniques. Markers
Predicting the water level along a river during and after a
indicate the positions of the gauging stations in the Madrid
rainfall event requires to know a number of variables: how
catchment. The map on the left is obtained by using Voronoi
much water falls on different parts of the catchment; what
tessellation, while that on the right is derived from a Barycen-
fraction of such water will infiltrate the ground and what will
tric interpolation.
flow to the surface (runoff) to the drainage channel; when
and where the runoff water reaches the discharge channel.
Fig. 1 exemplifies the main processes at catchment scale, with
interest for events-based hydrology. A more detailed hydro- image on the right, instead, reaches its maximum in the first
logical analysis, considering the specificity of the environment hour of rain (11:00), then the rain almost stops and starts
(topology, soil structure and absorbtion capabilities, ...) is out again in the late afternoon. Also from these examples, it is
of the scope of this investigation, which aims to obtain more clear that the weather stations can only provide a discrete,
general considerations and methodologies. sampled version of the two-dimensional time-varying contin-
The amount of rained water can be determined from the uous process of rainfall intensity. It is possible to obtain a
so-called hyetograph, i.e., the graphs of rainfall intensity over measure of the total amount of water rained over the area
time in the catchment. Empirically, hyetographs can be built during a given period by spatially interpolating the samples
from sensors that measure the rain intensity during a storm. collected by different weather stations in the same basin,
Fig. 2 shows an example of two hyetographs obtained in and then integrating over that period of time. However, the
different days from the weather stations monitoring network of accuracy of the result depends on ow the interpolation is
the Madrid City Council.1 These data will be used also for the performed. In Fig. 3, for example, we report two interpolation
following examples, unless otherwise mentioned. In the figure, maps obtained from the same dataset with different methods:
the hyetographs show storms with different behaviors: the since the computation of the interpolated values is different,
graph on the left represents a storm that evolves with gradual the estimation of the total rained water obtained by integrating
intensity through several hours (9:00 – 24:00), reaching its the values over the area of interest will yield to quite different
peak in the late afternoon. The rain event represented by the results. Hence, choosing the proper interpolation techniques is
critical.
1 [Link] Only part of the rained water flows into the drainage river,
while the remaining fraction is absorbed along the way by the to accurately measure processes and state variables is a
ground. Knowing the absorption rate of the soil, it is virtually priority for improving the understanding of processes. . .”, it
possible to find the rate of surface runoff at each point in the seems evident that the accurate process monitoring has to be
basin and, in turn, the net contribution to the discharge river. imperatively based on a rational strategy for optimal moni-
Unfortunately, the infiltration process is difficult to monitor, toring of hydrological processes as a whole. A comprehensive
and can be extremely heterogeneous over time and space. In approach for determining the optimal monitoring strategy, i.e.,
fact, not only the capability of the ground to absorb water what variables have to be measured, where, how often and
depends on the use of the terrain and on its composition, but what is the optimal spatial and temporal combination of remote
it also changes during a storm with the humidity of the terrain sensing and ground-based sensors, is still lacking.
itself. Finally, the time of concentration, i.e., the time taken Moreover, it might be necessary to increase the density of
by water to flow from a certain point in a watershed to the weather stations and sensors on the ground to collect more
watershed outlet, depends on topography, geology, and land accurate data and build precise maps of the hydrological
use and, hence, is very difficult to compute or predict. processes. This requires the deployment of a large number of
From this quick overview it should be clear that processes nodes, thus increasing the capital and operational expenditures.
involved in atmospheric water, surface water and subsurface Such deployments can benefit from energy-neutral gauging
water subsystems jointly evolve and interact with high spatial stations to reduce infrastructure and maintenance costs. In this
and temporal variability, determining the whole hydrological way, sensor nodes exploit energy harvesting capabilities to get
system behaviour. Therefore, despite the numerous theoretical the energy required for sensing, processing and transmitting
and empirical models that have been proposed to describe operations. This makes it possible to realize more extended
each of these phenomena, modeling their interactions in a real networks and place sensors that can not be powered by a power
catchment remains a formidable challenge. grid. Furthermore, when compared to the use of batteries,
solutions with energy harvesting are more sustainable and
III. M EASURING AND COLLECTION OF HYDROLOGICAL eco-friendly, since they do not require battery replacement.
DATA
Although, the performance of these solutions can strongly de-
Finding strategies for the optimal sensor location and tech- pend on the energy availability, and require an accurate design
niques to combine the measurements of different types of of sampling and communication protocols that increase the
sensors is fundamental to provide the hydrological models energy efficiency of data collection [15]–[17]. Some examples
with informative data. The problem of optimal sensor posi- are applications where it is possible to harvest energy from
tioning and sampling rate has been addressed in other contexts the environment (e.g., solar light, wind) or from the monitored
(e.g., [1], [2] including water distribution systems or storm process itself (stream water flows, geothermal phenomena).
water infrastructures [3]), but it remains widely unexplored in From this discussion, it is apparent that a trade-off is present
hydrology, where the problem is exacerbated by the spatio- for sensor density, quality of information and costs. Indeed,
temporal dynamics of the observed variables. Some results the deployed network and its configuration (i.e., sampling
based on mathematical theories for optimal process monitoring time, sensors position) should be tuned to provide reliable
have been recently published in [4]–[6]. A number of studies information, enough to provide accurate and precise results
on monitoring hydrological processes have been previously for the desired purpose but, at the same time, avoiding the
conducted (a detailed list can be found in [7]), but the collection of redundant data, which would lead to the useless
monitoring network is either oversized (e.g., [8] deployed 300 waste of memory, computational, and economic resources.
sensors for a 1 km2 catchment), or severely undersized, with
just a few gauging stations for large areas [9]–[11]. IV. DATA PROCESSING AND FORECASTING
Also, it is possible to exploit the temporal correlation of Hydrological theories provide tools to deal with complex
the monitored variables to optimize their sensing and storage. random phenomena with high spatial and temporal variability,
Some works have explored the possibility of using real-time in which physical approaches require relevant simplifications
weather forecasts and hydrologic conditions to autonomously and/or high computational efforts. On the other side, empiri-
adjust the measurement frequency, in order to intensify obser- cal models, much more frequently used, are more scenario-
vations during interesting events [12], [13]. specific and lack generality. Instead, machine-learning al-
How to optimally merge remote sensing and ground-based gorithms are instrumental in automatically estimating the
measurements, paying attention both to spatial and tempo- models’ parameters based on the actual data collected from
ral variability of the observed processes, remains an open the environment, in order to make them more flexible and
problem. The optimal combination will have to cope with adaptable to the underlying system characteristics. Such an
risks from potential statistical correlation between satellite approach has been successfully applied in other disciplines,
and in situ information [14], and allow for a significant e.g., telecommunications [18] and energetic engineering [19],
improvement in the speed and accuracy of the identification of but has only recently been considered in hydrology. For ex-
relevant events, for a given measurement rate and density of ample, Random Forest, Support Vector Machine, and Extreme
sensing locations. Following the International Association of Learning Machine have been used in [20] for the estima-
Hydrological Sciences (IAHS): “. . .the ability of hydrologists tion of daily reference evapotranspiration in the presence of
 TABLE I: Main statistical moments considering the entire storms dataset.
Metric Duration [h] Volume [mm] Average rainfall intensity [mm/h] Maximum rainfall intensity [mm/h]
Maximum 21 47.1 8.07 23.7
Minimum 1 0.1 0.1 0.1
Average ± std 2.56 ± 3.04 1.69 ± 4.3 0.3 ± 0.65 0.81 ± 2.03
Mode 1 - - -

TABLE II: Main statistical moments considering storms generating runoff according to the Green-Ampt model.
Metric Duration Volume Average Maximum Runoff volume Maximum runoff
[h] [mm] rainfall intensity [mm/h] rainfall intensity [mm/h] [mm] runoff intensity[mm/h]
Maximum 21 47.1 7.35 21.4 24.3 24.63
Minimum 2 5.7 0 1.9 0.1 0.17
Average ± std 8.44±0.9 14.54±1.62 1.82±2.93 5.81±2.12 5.55±1.74 5.2±1.71
Mode 5.44 8.75 1.34 4.25 5.97 5.59

meteorological data, while in [21], the authors developed a A hybrid approach can be combining hydrological models
study for the prediction of water flow using artificial neural with data-driven techniques. For example, Tab. I shows the
networks. In [22] Xu et al. integrated deep learning and the statistics of all rain phenomena in the Madrid dataset. By
fire-fly algorithm for optimizing the parameters of a Support considering [28], it is possible to obtain the soil parameters
Vector Regression method to predict the hydrological process for the catchment and to apply the Green-Ampt model [29]
of Huangfuchuan in Fugu County, China. Multiple linear to estimate the occurrence of runoff. This makes it possible
regression, K-nearest neighbors, Support Vector Regression, to select the events generating the runoff and, then, applying
Cubist, Random Forest and Artificial Neural Networks have ML techniques to investigate the behavior and the impact of
been widely used for mapping soil physical properties [23]. other variables (e.g., pressure, relative humidity, temperature).
Furthermore, ML and deep learning techniques have been Tab. II shows the statistical values when considering only
recently applied for water flow and flood prediction. In [24] the subset of precipitation events generating the runoff. As
the authors compare the performance of different types of expected, these include heavier storm events, corresponding
neural networks for monthly flow prediction. However, the best to longer storm duration and more intense rainfall.
solution turns out to be dependent on the specific case study, It is worth noting that ML algorithms themselves may suffer
amount of available features and training observations. In [25] from the a priori choice of the measured variables and sam-
the authors couple convolutional neural networks with transfer pling frequency. As an example, we clusterized metereological
learning to reduce the training time of the neural network in stations in the Veneto region, in the north of Italy,2 by con-
flood prediction problems. sidering different combinations of the available variables and
Other ML techniques have been applied to specific problems different sampling frequencies, employing the data collected
(signals decomposition, model parameters estimation, model in year 2017, with the purpose of identifying areas having
matching, etc), overlooking the general problem of modelling homogeneous weather characteristics. Fig. 4 shows the results
the whole water cycle. One of the reasons is that the com- for different combinations when fixing the number of clusters
plexity of such general problem overwhelms the capability of to 3 and using K-medoids clustering algorithm. Timeseries
classic, centralized ML approaches, which are not able to scale of the values of the variables of interest were given as input
up with the number, high-dimensionality and heterogeneity of features characterizing each weather station. It is apparent
the input variables involved in the general model. Recently, that, although the groups of stations are similar, roughly
promising results have been obtained by pre-processing the corresponding to the three geographical regions (mountains,
input signals to make them more informative, before feeding plain, Venetian lagoon/coast), the actual clusters are different.
them into ML algorithms [26], [27]. However, no attempt has For example, while the clusters obtained considering only
been made to model the whole water cycle, given its huge the precipitation (Figs. 4a, 4d) are geographically compact,
complexity, and the variety of input processes it hinges upon. those obtained with the relative humidity (Figs. 4b, 4e) are
Existing hydrological models can be improved in different more spread. Then, considering Figs. 4c, 4f, where all the
ways. First, simplifying the existing models can provide a variables are considered, we can observe that also the sampling
better understanding of the most important parameters to be frequency plays a role in the cluster determination. This
accurately estimated, helping to fit the model to the actual example, although simple, shows how the choice of the input
geographical scenario. Second, the exploitation of suitable ML data, previously discussed in Sec. III, can impact the result of
techniques, specifically selected to the purpose, can contribute ML approaches, thus proving the importance of proper data
to automatically configure such parameters according to the selection.
characteristics of the region, using satellite maps and mea- 2 ARPAV dataset: [Link]
surements provided by on-ground sensors. 2022 [Link].
 (a) Precipitation, 1 sample/day. (b) Relative humidity, 1 sample/day. (c) All variables, 1 sample/day.

(d) Precipitation, 4 samples/hour. (e) Relative humidity, 4 samples/hour. (f) All variables, 4 samples/hour.

Fig. 4: Clustering of meteorological stations in the Veneto region (IT) using as features different combinations of environmental
variables (precipitation, relative humidity, air temperature, solar radiation) and sampling periods (24 h and 15’). Clusters are
identified by different colors.

We stress out that the aim of data-driven approaches should It is hence fundamental to develop new techniques to
not be too specific to the addressed scenario, but rather provide visualize the output of the hydrological models in a graphical
a framework that can be applied to different cases. For ex- format, making it possible to “observe” the processes in the
ample, unsupervised deep learning and sensitivity/uncertainty different subsystems and the water fluxes and storage within
analysis techniques for identifying the most relevant param- and among hydrological subsystems. Furthermore, the tool
eters for the accuracy of the model should be considered, should make it possible to observe how the water level in
as proposed in [26]. These methodologies, therefore, acquire the catchments is predicted to change when varying some
general purpose and applicability, and can become an effective of the system input variables (e.g., rainfall/sun/wind inten-
tool to study an area of interest. Once developed, the general sity, geomorphology, land uses, water allocation within the
ML architecture can thus be fed offline with historical data catchment). The uncertainty of the estimation should also
collected from the area to be studied, and be employed as a be graphically represented, to provide an indication of the
black-box tool to estimate the soil parameters and predicting reliability of the prediction. Digital visualization tools in
the water distribution and movement according to the mea- Geographic Information System (GIS) can be used to represent
sured variables. Because of its use and possible computational the results (e.g., surface morphology of terrain, soil moisture,
cost, it is expected that such a framework will run offline and surface water flow) [30], [31].
in a centralized manner, employing the available data collected
previously. VI. WAY FORWARD
In this section we give an overview of the tools and research
V. V ISUALIZATION TOOLS directions that can be explored to tackle the problem of water
management.
Currently, hydrological models are accessible only to ex- In complex systems, including hydrologic ones, macro-
perts, which know the meaning and role of the different scopic effects arise from the interactions among lower-layer
models parameters, and have the skills to correctly read and processes. To deal with such systems, it is hence natural to fo-
interpret the models’ outcome, which may be given in the form cus on the individual underlying elements, before attacking the
of probability distributions or time-series with uncertainty system as a whole. In the case of hydrology, the macroscopic
intervals. However, decision making in the water management effects at basin level emerge from the interactions among at-
domain involves a number of stakeholders, not all equipped mospheric (rainfall), surface and subsurface water subsystems.
with the technical background to correctly use the model, or In order to tame this complexity, it is worth considering each
interpret its outcomes, especially thematic maps. of the subsystems individually, before making an attempt to
 - Hydrological models from - Collected (heterogeneous) data
the literature
- Measurements in the fields
validation ML TECHNIQUES

HYDROLOGICAL
MODEL
SELECTION - Model abstraction of primary/dependent
variables and their relation

- Prediction (runoff,
infiltration, surface - Optimize sensor location and measurement
water levels, …) frequency
- Identification and estimation of the
parameters of hydrological models

For every subsystem/subproblem MERGE AND UPSCALE

Fig. 5: Diagram representing the relations and dependencies among the various methodology steps.

derive a more comprehensive model for the whole water cycle movement of groundwater can be expressed by a variation of
in a catchment. Darcy’s law dependent on permeability aquifer’s coefficient
For the rainfall process, models should focus on both short- (or complementarity by transmissivity aquifer’s rate) which in
term forecasting, seeking to model individual storm events, general terms can be easily determined by field measurements
and medium/long term for detecting trends relevant for water paying particular attention to soil spatial heterogeneity.
management and allocation. Conversely to current approaches In summary, to address the challenges described in this
in the literature, an attempt should be made to model individual paper, we recommend the following general methodology,
storm events focusing on physical laws instead of defining whose main steps are represented in Fig. 5 along with their
purely autoregressive or model-free solutions. This “white dependencies.
box” approach, if successful, will not only provide efficient
• Select reliable and accurate hydrologic models of the
models to predict storm events, but also shed light on the
individual water subsystems whose parameters can be
parameters that have a stronger impact on the physical models,
more easily estimated from environmental data.
thus contributing to a better understating of the theoretical
• Exploit ML techniques to extract abstract models of
aspects of hydrology. Algorithms for detecting and predicting
heterogeneous sensing data, including remote sensing
medium/long term trends are also needed to guide water man-
(satellite and aerial 2D/3D images), and ground measure-
agement decisions, plans and policies. Satellite information
ments of primary variables (solar radiation, temperature,
can help with medium/long term forecasts, while short term
humidity, soil water content, water table level, etc.)
characterisation and forecast can be aided by both ground-
and dependent variables (evaporation, evapotranspiration,
based sensors or radar measurements. In addition, portable
precipitation, infiltration, surface water levels, velocity,
radars can be used for retrieving highly detailed spatial infor-
depletion levels, etc.)
mation (for example portable X-band weather radars) to ben-
• Use such models to optimise sensor location and mea-
efit from accurate and high spatial resolution measurements.
surement frequency, in order to improve the quality
On atmospheric water subsystems, theoretical aerodynamic,
(accuracy, reliability, information provided) of the data
energy balance or combined approaches, should be explored
and reduce monitoring costs.
for modelling both evaporation and evapotranspiration.
• Develop techniques to automatically configure the hy-
Surface water subsystem encompasses infiltration and
drologic models parameters based on the information
runoff-related processes. Remote sensing has been proved to
obtained from the data, in order to improve their accuracy
be valid for directly measuring water content at upper soil
while reducing complexity.
layers, but the required temporal resolution for estimating infil-
• Merge and upscale the previous models to come up with a
tration evolution during storm events is not ensured by satellite
comprehensive model for the whole hydrological system,
images. For this reason, infiltration should be measured using
following the Reynold Transport Theorem, and solve the
ground-based sensors. On infiltration models, physically based
system by applying distributed learning techniques or
approaches as those proposed by Richards, Green-Ampt and
PDE approximation methods based on neural networks
Soil Conservation System Curve Number method can be of
with prior knowledge of the physical world.
use.
Infiltration phenomena are also linked to the groundwater The fusion of different types of measurements is hence a
dynamics, which affect the subsurface water subsystem. The key characteristic of the proposed approach: the underlying as-
sumption is that the ensemble of remote, aerial, and on-ground of Excellence” initiative (Law 232/2016). In Spain, this
measurements contains all the information needed to correctly paper belongs to PCI2020-120694-2 project funded by
set the parameters of the selected hydrological models, and MCIN/AEI/10.13039/501100011033 and the European Union
that such information can be automatically extracted from the “NextGenerationEU”/PRTR.
data by means of proper ML approaches.
For validation and calibration of the developed ML models,
R EFERENCES
physically based theories can be used namely: Reynolds’
transport theorem for defining mass and energy fluxes at [1] D. J. Chmielewski, T. Palmer, and V. Manousiouthakis, “On the theory
the whole hydrological system; theoretical aerodynamic, en- of optimal sensor placement,” AIChE journal, vol. 48, no. 5, pp. 1001–
ergy balance or combined approaches, for modelling both 1012, 2002.
evaporation and evapotranspiration; precipitable water column, [2] A. A. Alonso, I. G. Kevrekidis, J. R. Banga, and C. E. Frouzakis, “Op-
timal sensor location and reduced order observer design for distributed
thunderstorm cell models, frontal storm, mesoscale convective process systems,” Computers & chemical engineering, vol. 28, no. 1-2,
systems for rainfall estimate; Richards [32], Green-Ampt [29], pp. 27–35, 2004.
and SCS Curve number methods for estimating infiltration and [3] M. Bartos and B. Kerkez, “Hydrograph peak-shaving using a graph-
theoretic algorithm for placement of hydraulic control structures,” Ad-
soil water dynamics; unit hydrographs, lumped flow routing vances in Water Resources, vol. 127, pp. 167–179, 2019.
models and distributed flow methods based on Saint Venant’s [4] S. Zubelzu, L. Rodrı́guez-Sinobas, D. Segovia-Cardozo, and A. Dı́ez-
equation for surface water dynamics; and modified Darcy´s Herrero, “Optimal locations for flow and velocity sensors along a river
channel,” Hydrological Sciences Journal, vol. 65, no. 5, pp. 800–812,
law for groundwater dynamics. 2020.
[5] W. H. Asquith, “The use of support vectors from support vector
VII. C ONCLUSIONS machines for hydrometeorologic monitoring network analyses,” Journal
of Hydrology, vol. 583, p. 124522, 2020.
An effective management of water resources is precious
[6] L. Zhuo, Q. Dai, B. Zhao, and D. Han, “Soil moisture sensor network
for social, economical and safety reasons. Addressing this design for hydrological applications,” Hydrology and Earth System
problem is even more urgent when considering the presence Sciences, vol. 24, no. 5, pp. 2577–2591, 2020.
of critical events produced by climate change. However, as [7] F. Tauro, J. Selker, N. Van De Giesen, T. Abrate, R. Uijlenhoet,
M. Porfiri, S. Manfreda, K. Caylor, T. Moramarco, J. Benveniste
explained and exemplified in the previous sections, the meth- et al., “Measurements and Observations in the XXI century (MOXXI):
ods that are currently used (i.e., hydrological models) rely on innovation and multi-disciplinarity to sense the hydrological cycle,”
many parameters and variables, whose meaning is not always Hydrological sciences journal, vol. 63, no. 2, pp. 169–196, 2018.
[8] B. Kerkez, S. D. Glaser, R. C. Bales, and M. W. Meadows, “Design and
related to physical quantities. Furthermore, they often need to performance of a wireless sensor network for catchment-scale snow and
be calibrated and fed with values coming from estimated or soil moisture measurements,” Water Resources Research, vol. 48, no. 9,
measured values that describe the soil characteristics, which 2012.
[9] E. Varouchakis, D. Hristopulos, and G. Karatzas, “Improving kriging of
prevents them from being generally applicable and requires groundwater level data using nonlinear normalizing transformations—a
extensive measurement campaigns on the field. field application,” Hydrological Sciences Journal, vol. 57, no. 7, pp.
Current ICT solutions, such as sensor networks and artificial 1404–1419, 2012.
intelligence can be useful in the model definition and param- [10] A. Nebenzal, B. Fishbain, and S. Kendler, “Model-based dense air pollu-
tion maps from sparse sensing in multi-source scenarios,” Environmental
eter tuning, based on measured data. Though, as explained in Modelling & Software, vol. 128, p. 104701, 2020.
this work, modeling and predicting water flows and distribu- [11] S. Grimaldi, A. Petroselli, and F. Serinaldi, “A continuous simulation
tion means considering a complex system, which is affected model for design-hydrograph estimation in small and ungauged water-
sheds,” Hydrological Sciences Journal, vol. 57, no. 6, pp. 1035–1051,
by many variables and elements (e.g., soil characteristics and 2012.
use, weather conditions). Hence, to attack the problem, there [12] B. P. Wong and B. Kerkez, “Real-time environmental sensor data:
is a need to gain a better understanding of several aspects, An application to water quality using web services,” Environmental
Modelling & Software, vol. 84, pp. 505–517, 2016.
related to the observation, processing, and interpretation of [13] ——, “Adaptive measurements of urban runoff quality,” Water Resources
the hydrological processes. How many gauging stations are Research, vol. 52, no. 11, pp. 8986–9000, 2016.
needed to correctly estimate the amount of rained water in a [14] J. Dong, L. Wei, X. Chen, Z. Duan, and Y. Lu, “An instrument variable
catchments? Which type of sensors should be provided to such based algorithm for estimating cross-correlated hydrological remote
sensing errors,” Journal of Hydrology, vol. 581, p. 124413, 2020.
stations? What should be the sampling period for the different [15] A. Biason, C. Pielli, A. Zanella, and M. Zorzi, “Access control for
environmental variables? How to avoid bias in the collected IoT nodes with energy and fidelity constraints,” IEEE Transactions on
data? To what extent is it possible to predict events? Which Wireless Communications, vol. 17, no. 5, pp. 3242–3257, 2018.
[16] C. Pielli, D. Zucchetto, A. Zanella, and M. Zorzi, “An interference-aware
hydrological theoretical/empirical models are more suitable channel access strategy for WSNs exploiting temporal correlation,” IEEE
to be combined with ML-aided parameters estimate? These Transactions on Communications, vol. 67, no. 12, pp. 8585–8597, 2019.
are only a few of the open questions that call for a stronger [17] M. Capuzzo, C. Delgado, J. Famaey, and A. Zanella, “An ns-3 im-
plementation of a battery-less node for energy-harvesting internet of
cooperation between ICT and hydrology experts. things,” in Proceedings of the Workshop on ns-3, 2021, pp. 57–64.
[18] C. Jiang, H. Zhang, Y. Ren, Z. Han, K.-C. Chen, and L. Hanzo,
ACKNOWLEDGMENT “Machine learning paradigms for next-generation wireless networks,”
This research was funded by the CHIST-ERA grant IEEE Wireless Communications, vol. 24, no. 2, pp. 98–105, 2016.
[19] X. Chen, K. Yu, W. Du, W. Zhao, and G. Liu, “Parameters identification
CHIST-ERA-19-CES-002 and by MIUR (Italian Ministry of solar cell models using generalized oppositional teaching learning
for Education and Research) under the ”Departments based optimization,” Energy, vol. 99, pp. 170–180, 2016.
[20] J. Fan, W. Yue, L. Wu, F. Zhang, H. Cai, X. Wang, X. Lu, and
Y. Xiang, “Evaluation of SVM, ELM and four tree-based ensemble
models for predicting daily reference evapotranspiration using limited
meteorological data in different climates of China,” Agricultural and
forest meteorology, vol. 263, pp. 225–241, 2018.
[21] M. Kovačević, N. Ivanišević, T. Dašić, and L. Marković, “Application
of artificial neural networks for hydrological modelling in karst,” Graev-
inar, vol. 70, no. 01., pp. 1–10, 2018.
[22] L. Xu, J. Zhao, C. Li, C. Li, X. Wang, and Z. Xie, “Simulation and
prediction of hydrological processes based on firefly algorithm with deep
learning and support vector for regression,” International Journal of
Parallel, Emergent and Distributed Systems, vol. 35, no. 3, pp. 288–
296, 2020.
[23] Y. Khaledian and B. A. Miller, “Selecting appropriate machine learning
methods for digital soil mapping,” Applied Mathematical Modelling,
vol. 81, pp. 401–418, 2020.
[24] Z. Alizadeh, J. Yazdi, J. H. Kim, and A. K. Al-Shamiri, “Assessment
of machine learning techniques for monthly flow prediction,” Water,
vol. 10, no. 11, p. 1676, 2018.
[25] N. Kimura, I. Yoshinaga, K. Sekijima, I. Azechi, and D. Baba, “Con-
volutional neural network coupled with a transfer-learning approach for
time-series flood predictions,” Water, vol. 12, no. 1, p. 96, 2020.
[26] M. Zorzi, A. Zanella, A. Testolin, M. D. F. De Grazia, and M. Zorzi,
“Cognition-based networks: A new perspective on network optimization
using learning and distributed intelligence,” IEEE Access, vol. 3, pp.
1512–1530, 2015.
[27] M. Rezaie-Balf, O. Kisi, and L. H. Chua, “Application of ensemble em-
pirical mode decomposition based on machine learning methodologies
in forecasting monthly pan evaporation,” Hydrology Research, vol. 50,
no. 2, pp. 498–516, 2019.
[28] R. F. Carsel and R. S. Parrish, “Developing joint probability
distributions of soil water retention characteristics,” Water Resources
Research, vol. 24, no. 5, pp. 755–769, 1988. [Online]. Available: https:
//[Link]/doi/abs/10.1029/WR024i005p00755
[29] W. Green, “Ampt. GA: Studies on soil physics. 1. Flow of air and water
through soils,” J. Agr. Sci, vol. 4, pp. 1–24, 1911.
[30] S. Cucchiaro, D. J. Fallu, H. Zhang, K. Walsh, K. Van Oost, A. G.
Brown, and P. Tarolli, “Multiplatform-SfM and TLS data fusion for
monitoring agricultural terraces in complex topographic and landcover
conditions,” Remote Sensing, vol. 12, no. 12, p. 1946, 2020.
[31] A. Pijl, L. E. Reuter, E. Quarella, T. A. Vogel, and P. Tarolli, “GIS-based
soil erosion modelling under various steep-slope vineyard practices,”
Catena, vol. 193, p. 104604, 2020.
[32] L. A. Richards, “Capillary Conduction of Liquids through Porous
Mediums,” Physics, vol. 1, no. 5, pp. 318–333, 1931.