SMARTerra, a High-Resolution Decision Support System for Monitoring Plant Pests and Diseases

Featured Application

The SMARTerra decision support system provides high-resolution risk maps related to plant diseases and pests, enabling prompt and targeted intervention that minimizes environmental impact while minimizing economic losses.

Abstract

The prediction and monitoring of plant diseases and pests are key activities in agriculture. These activities enable growers to take preventive measures to reduce the spread of diseases and harmful insects. Consequently, they reduce crop loss, make pesticide and resource use more efficient, and preserve plant health, contributing to environmental sustainability. We illustrate the SMARTerra decision support system, which processes daily measured and predicted weather data, spatially interpolating them at high resolution across the entire Sardinia region. From these data, SMARTerra generates risk predictions for plant pests and diseases. Currently, models for predicting the risk of rice blast disease and the hatching of locust eggs are implemented in the infrastructure. The web interface of the SMARTerra platform allows users to visualize detailed risk maps and promptly take preventive measures. A simple notification system is also implemented to directly alert emergency responders. Model outputs by the SMARTerra infrastructure are comparable with results from in-field observations produced by the LAORE Regional Agency. The infrastructure provides a database for storing the time series and risk maps generated, which can be used by agencies and researchers to conduct further analysis.

1. Introduction

In this paper, we use the term “pest” to refer to unwanted and potentially harmful organisms that inhabit and proliferate within crops causing damage to plants, such as weeds, insects, and other animals. We reserve the term “disease” for pathologies caused by microorganisms, including fungi, bacteria, and viruses. The effective management of pests and diseases is essential for ensuring the sustainability and prosperity of agriculture. In field emergencies, time is a critical factor. Traditional approaches to this problem, such as field observation, disease diagnosis, pest identification and counting, are hardly feasible given the current demands of precision agriculture for medium- to large-scale crops. The accurate prediction and timely monitoring of harmful diseases and insects enable growers to implement targeted preventive measures, optimize resource use, and preserve crop health. These considerations are of paramount importance globally, acknowledged by governmental entities, research institutions, and non-governmental organizations [1,2,3,4,5]. While numerous methods exist for the early detection of crop emergencies, based on in-field or remote sensors installed on drones or ground vehicles and governed by artificial intelligence (AI) or computer vision algorithms (see [6] and references therein), the focus here is on the prediction before problems arise. Models for such risk forecasting come in various forms, all sharing a common feature: they heavily rely on the strong interplay between crop behavior as a whole and weather conditions. The latter are well known to affect plant growth and development, trigger pathological outbreaks, and regulate pest spread. Agrometeorological parameters such as temperature, humidity, and rainfall are invariably involved in most agricultural risk models, like those cited above, which have proven to be effective. SMARTerra, the infrastructure we propose, has been designed with a dual purpose: (i) to provide a fully automatic system capable of processing measured and predicted weather data to generate near real-time forecasts of the risk associated with plant pests for the entire Sardinia region (Italy) and (ii) to perform this task at the single-farm scale (10 m). The latter objective is only partially guaranteed by standard interpolation techniques provided in common software libraries and therefore deserves a specific, innovative mathematical approach. This is the subject of a methodological paper that the authors are currently submitting for publication [7]. SMARTerra is designed to provide growers and agricultural professionals with a reliable, accessible tool based on advanced ICT technology. Its web interface allows users to view detailed risk prediction maps, facilitating timely, informed decisions and the implementation of targeted prevention strategies. Currently, the system integrates models for predicting the risk of two specific agricultural threats in Sardinia, detailed below, which can be regarded as working examples. However, SMARTerra is easily adaptable to any geographic area and a range of agricultural emergencies, given an available network of agrometeorological stations and suitable risk models for the relevant crops.

Few studies exist in the literature on this topic, mainly referred to as Integrated Pest Management (IPM). Reference [8] reviews several IPMs, either leveraging crop rotations or typically starting from the diagnosis of existing pests, rather than attempting to anticipate their appearance and spread. Notably, [9] presents a Decision Support System (DSS) for tree fruit growers based on meteorological data (measured by a single station—the nearest one) and model predictions, accessible via an annual fee. Such systems highlight a common limitation: they rely on non-optimal approximations of measured agrometeorological parameters. On a different ground, data-driven methods hold significant promise for pest prediction. Unlike deterministic models, they do not require the careful calibration of numerous parameters that are crop-specific and dependent on soil features, plant cultivar, general environmental conditions of the local area (including the dynamics of previous pest occurrences), and other factors. This represents a significant advantage. In our view, the main limitation of deep learning analysis is the need for large, annotated training datasets. While long-term time series of agrometeorological data are often available (as they are in our case of interest), a corresponding systematic, quantitative characterization of a particular pest or disease for the considered time range is seldom available. Therefore, whenever a new model for a particular emergency is desired, an on-field data recording of the dynamics of such emergency (the annotation for training) is necessary. This step is usually more demanding and time-consuming than the calibration of deterministic algorithms. Systematic and valuable reviews of different modeling methods, both deterministic and data-driven, can be found in [10,11,12], where the authors propose a holistic approach for linking crop, insects, and environmental conditions, and in [13] for plant disease modeling as part of a comprehensive approach to develop decision-making tools for IPM.

To illustrate the workflow and outcomes of the web-based method we propose, we consider two case studies related to plant pathologies and crop-feeding insects particularly invasive in Sardinia: the rice blast (Pyricularia oryzae) disease and the infestation of Moroccan locust (Dociostaurus maroccanus). SMARTerra is currently accessible to a selected group of end-users for testing. It will soon be available to farmers, agricultural agencies, and industry professionals once the terms of service, which are currently being finalized, are released. An online demo is available upon request to showcase the system’s capabilities. A precursor to SMARTerra exists, where weather data—both measured from a network of stations and predicted by a local meteorological model (BOLAM)—are manually processed to evaluate risk indices for a limited set of locations and at low resolution. This process employs spreadsheet-based technology, which is time-consuming and requires the daily involvement of several human experts susceptible to errors, such as the incorrect import or processing of data (e.g., location or date swapping). The detection of outliers in weather measurements is not systematically performed at the current development stage of SMARTerra: the implementation of this functionality will be described in [7]. The outcomes of SMARTerra allow for the timely notification of affected users not only through an online publication but also through automated and/or personalized bulletin delivery via email, smartphone applications, and other instant messaging applications. The data, methodologies, infrastructure of our decision support system, and the geographical areas of the study are illustrated in Section 2. Examples of weather and risk maps for rice blast disease and locust egg hatching, which are outputs of SMARTerra, are shown in Section 3. A discussion on the improvements, innovations, and benefits delivered by SMARTerra is provided in Section 4. Finally, Section 5 presents the conclusions along with future development scenarios.

2. Methods

The geographical areas, i.e., the island of Sardinia and the rice-growing districts, for which the SMARTerra decision support system is currently employed, and the meteorological data are described in Section 2.1. The back-end and front-end infrastructures of SMARTerra, summarized in Figure 1, are presented in Section 2.2 and Section 2.3, respectively.

The back-end of SMARTerra is responsible for the activities and functionalities that are not directly visible to end-users but are crucial to the operation of DSS as a whole. In particular, the back-end undertakes daily the following tasks, all performed in a full automatic way:

• Downloading and pre-processing the weather measurements and forecasts (Section 2.2.1);
• Efficiently storing the large volume of meteorological data (Section 2.2.2);
• Spatially interpolating the temperature, humidity and rain accumulation measurements with the Kriging with External Drift technique (Section 2.2.3);
• Spatially interpolating the temperature, humidity and rain accumulation forecasts from the BOLAM model (Section 2.2.4);
• Forecasting the risk indices and alert levels associated to the rice blast disease (Section 2.2.5);
• Forecasting the dates of the hatching of locust eggs (Section 2.2.6);
• Producing the raster files for the measured weather data and for the agricultural pests (Section 2.2.7) and generating a PDF report summarizing the whole output in a few pages.

On the other hand, the WebGIS portal (Section 2.3) serves as an interface with the user, who can access maps and data of the following:

• The hourly meteorological measurements over the entire territory of Sardinia, spatialized at $100 \mathrm{m}$ resolution;
• The risk indices and level alerts for Pyricularia oryzae at $10 \mathrm{m}$ resolution over the area of the rice-growing districts of Oristano and San Gavino;
• The cumulative rainfall and cumulative growing degree days, and the dates predicted for the hatching of locust eggs over the entire regional territory at $100 \mathrm{m}$ resolution.

2.1. Geographical Areas and Weather Data

2.1.1. The Island of Sardinia and the Rice Districts

The SMARTerra infrastructure is specifically developed for Sardinia, which, with an area of over 24,000 ${\mathrm{km}}^2$, is the second largest island in the Mediterranean Sea. Nevertheless, it can be readily adapted to other geographical regions. The Sardinian landscape ranges from fertile plains to rugged coastlines and mountain areas, resulting in distinct microclimates that may variably influence the development of pests and plant diseases (Figure 2, left).

This study will show, at $10 \mathrm{m}$ resolution, the alert level for rice blast disease on paddy fields in the rice-growing districts of Oristano and San Gavino and, at $100 \mathrm{m}$ resolution, the prediction for locust egg hatching for the entire region of Sardinia.

The SMARTerra backbone can be adapted, with minor modifications, to other geographical areas equipped with agrometeorological station networks, provided that local weather models are available.

2.1.2. Regional Network of Meteorological Stations

ARPA Sardegna collects the measurements sent by the Regional Meteorological Network (Rete Unica Regionale—RUR). The network comprises 109 thermo-pluviometric stations and 77 meteorological stations, totaling 186 stations (Figure 2, right) distributed throughout the island,  typically installed near population centers. In addition to the sensors, the network includes the instrumentation required for data transmission via radio or GPRS/UMTS, a series of radio repeaters, and the two data acquisition stations located in Cagliari and Sassari. The thermo-pluviometric stations provide measurements of air temperature ($°\mathrm{C}$) and precipitation ($\mathrm{m}\mathrm{m}/\mathrm{h}$), while the meteorological stations, in addition to the aforementioned measurements, also supply relative humidity ($\%$), global solar radiation ($\mathrm{W}/{\mathrm{m}}^2$), and wind speed ($\mathrm{m}/\mathrm{s}$) at $10 \mathrm{m}$ above the soil. While the calibration and maintenance of the station instruments is outsourced to specialized third parties, the Meteoclimatic Department of ARPA Sardegna is responsible for the quality control of the above measurements and for their processing into hourly and daily time series. The quality control of weather measurements adopts standard and automated data validation criteria such as physical consistency (i.e., the recorded values must fall within the expected range) and climatological consistency (as above, but relative to the typical climate of the region under consideration), temporal consistency (i.e., variations between consecutive readings), and persistence of measurements. On the other hand, spatial consistency check is performed manually and almost daily by comparing the measurements of neighbouring stations. All of the above quality control procedures generate validation flags that are attached to the data, thus denoting the level of reliability of the measurements. The data along with the corresponding validation codes are then stored in a relational database. Dedicated software therefore extracts only the validated data and prepares the files to be delivered daily to CRS4 via an FTP server. Consequently, barring sensor and transmitter malfunctions, thousands of measurements are collected daily by CRS4.

2.1.3. BOLAM Weather Forecasts

Weather forecasts for temperature, relative humidity, precipitation, wind speed, and global solar radiation are produced by the ARPAS Meteoclimatic Department using the BOlogna Limited Area Model (BOLAM) hydrostatic model simulations [15,16,17,18]. The forecasts are made on a grid of 1922 nodes, arranged with a horizontal resolution of $0.04°$, corresponding to a spacing of approximately $4.5 \mathrm{k}\mathrm{m}$ (Figure 2, right). For each meteorological parameter, hourly predictions are provided, covering a 2-day period. Overall, the BOLAM time series received daily at CRS4 contain nearly half a million forecast values.

2.2. Back-End

This section briefly illustrates the entire back-end workflow, from the preprocessing of measured and forecast weather data to the generation of raster files with risk indices and their storage on the CRS4 filesystem. Notably, all analysis components are performed fully automatically and in near real time.

2.2.1. Data Preprocessing

Every day D, ARPA Sardegna programmatically sends CRS4 a compressed file containing the following meteorological data: the measurements from all stations of the regional network for the previous day $D-1$ and the forecasts for the same meteorological variables calculated by the BOLAM model for days D and $D+1$. Automatic data reception at CRS4 is ensured each morning by an event oriented software system that (i) waits for the archive to be uploaded to the FTP server, (ii) downloads it locally to the storage area of the computing cluster, also creating backup files, and (iii) extracts and processes the data which are then ready to be stored in the InfluxDB database. An email is sent to SMARTerra maintainers, summarizing the contents of the received datasets and issuing an alert in case of anomalies. The entire procedure is performed automatically, without the need for human intervention, all days of the year.

2.2.2. Database

InfluxDB is a time-series database (TSDB) management system, i.e., a database designed to store, query, and analyze time-varying data, selected for the SMARTerra project. Being optimized for time-series data, such as those from sensors, IoT devices, infrastructure monitoring, and other applications where understanding the evolution of variables over time is critical, InfluxDB is therefore the ideal solution for SMARTerra: it allows the immediate retrieval of time series measurements for one or more weather stations, for one or several more time points, and for a delimited area or for the whole region, thus facilitating the execution, in addition to the interpolation techniques described in Section 2.2.3 and Section 2.2.4, of further interpolation methods and data analyses, such as the outlier detection technique described in [7]. In the next few years, when hopefully the SMARTerra application is more mature, no longer limited to the region of Sardinia alone, and the database contains several multiple-year time series of agrometeorological variables, the capabilities of InfluxDB will be fully exploited for, e.g., calibrating the models to predict the rice blast disease infection (Section 2.2.5), the hatching date of locusts (Section 2.2.6) or any other pest or disease forecast model yet to be implemented in the DSS and enabling the implementation of other agricultural models, such as those for the growth, the yield and the planning of crops, for the automated irrigation of fields, for studying the phenology of fruits and flowers. Another advantage of InfluxDB is the ease of integrating database operations into the SMARTerra procedure via API calls. Data storage is performed in double precision, but the use of compression techniques greatly reduces the disk space required. The dedicated bucket is organized into four measurement categories (Figure 3): hourly and daily measurements from the RUR network stations and hourly time-series from the BOLAM model for the same day (hours 1 to 24) and for the next day (hours 25 to 48) forecasts.

InfluxDB tags are associated with the measurements to facilitate querying the time series, e.g., station name, geographic coordinates, distance from the sea, elevation above sea level (EASL), and elevation above the valley floor (EAVF). The last feature is defined, for each station s, as ${\mathrm{EAVF}}_s={\mathrm{EASL}}_s-min{\left({\mathrm{EASL}}_s\right)}_{3 \mathrm{k}\mathrm{m}}$, where $min{\left({\mathrm{EASL}}_s\right)}_{3 \mathrm{k}\mathrm{m}}$ is the minimum value of EASL evaluated within a $3 \mathrm{k}\mathrm{m}$ circle around the station s. The EAVF is introduced to account for the different meteorological dynamics between points with the same EASL that correspond, for instance, to a mountain peak and those lying on an upland. These additional variables are then used as the external variables in the spatial interpolation with the Kriging with External Drift technique (see Section 2.2.3).

2.2.3. Geostatistical Interpolation of Measured Data

To interpolate agrometeorological values measured at stations, which are typically distributed in an irregular pattern, Kriging with External Drift (KED) is used. KED belongs to a well-known family of hybrid geostatistical interpolation methods that estimate the values $z\left(s_0\right)$ of a variable z at unsampled locations $s_0$ based on its known measurements $z\left(s_i\right)$ at n locations $s_1,\dots ,s_n$, i.e., the stations. The underlying idea of such hybrid approaches is to merge the deterministic contribution of attribute values, known both at the stations and at the unsampled points, with a spatially autocorrelated stochastic component. As the formulation of KED is not straightforward, we resort to a closely related method, Regression Kriging (RK).

RK originates from Ordinary Kriging (OK), which fundamentally involves a linear combination of the observed values.

$${\widehat{z}}_{\mathrm{OK}}\left(s_0\right)=\sum _{i=1}^n{\lambda }_{i}z\left(s_i\right)$$

(1)

with the constraint ${\sum }_{i=1}^n{\lambda }_i=1$. The weights ${\lambda }_i$ minimize the estimation variance and depend on the spatial correlation pattern of z as described by the variogram (see [19] and the references therein).

RK combines the Ordinary Kriging method described above with standard regression analysis to achieve improved spatial predictions. In our case, five features $q_1,\dots ,q_5$ (longitude, latitude, distance from the sea, elevation above sea level EASL, and elevation above the valley floor EAVF) are used to estimate

$$\widehat{z}\left(s\right)={\beta }_0+{\beta }_1{q}_1+{\beta }_2{q}_2+{\beta }_3{q}_3+{\beta }_4{q}_4+{\beta }_5{q}_5$$

(2)

The RK approximation is obtained by merging Equation (2) with the Kriging contribution:

$${\widehat{z}}_{\mathrm{RK}}\left(s_0\right)=\widehat{z}\left(s_0\right)+\sum _{i=1}^n{\lambda }_i\left(z\left(s_i\right)-\widehat{z}\left(s_i\right)\right)=\widehat{z}\left(s_0\right)+\sum _{i=1}^n{\lambda }_iϵ\left(s_i\right)=\widehat{z}\left(s_0\right)+\widehat{ϵ}\left(s_0\right)$$

(3)

Kriging with External Drift is similar to RK, but instead of explicitly separating the regression and Kriging contributions as in RK, it incorporates the “external drift” (i.e., the information from the features $q_k$) directly into the Kriging process. We do not provide mathematical details here and refer interested readers to [20]. The reason we prefer KED over RK is its slightly better performance in our numerical tests and more robust expected convergence properties.

For an extensive application of KED to temperature interpolation, see [21]. The use of KED and RK for the mapping of soil properties is illustrated in [22,23], respectively.

A comprehensive analysis of different Kriging methods, along with considerations about their accuracy for the problem at hand, is a demanding task and cannot be performed here. Furthermore, Kriging methods are not the only possible approach; machine learning (ML) models can also be exploited for the optimal interpolation of agrometeorological parameters. A dedicated study on the mathematical aspects of different methods, including those from the Kriging family and various ML algorithms, with exhaustive assessments and benchmarks, is currently in preparation [7].

2.2.4. Spatial Interpolation of Weather Forecasts

The nodes in the BOLAM model are arranged according to a Cartesian grid in terms of geographic coordinates, where unnecessary far offshore points have been removed to save computational effort (see Figure 2, right). Interpolating from such a pseudo-structured grid is mathematically easier than interpolating from the irregular pattern of the stations’ distribution and can be carried out, for instance, by standard bicubic spline approximation. This enables the interpolation of meteorological quantities on a finer grid, increasing the resolution from approximately $4 \mathrm{k}\mathrm{m}$ (the BOLAM grid step size) to $100 \mathrm{m}$ over the entire island or to $10 \mathrm{m}$ in regions critical for SMARTerra where farm-scale resolution is required.

2.2.5. Forecast of Rice Blast Disease

Rice is one of the most widely grown foods in the world, particularly in Asia and European regions bordering the Mediterranean Sea, including Sardinia. Rice blast disease, caused by the fungus Pyricularia oryzae, is a major threat to this crop. Countering the infection caused by the fungus is of primary importance, as the disease can significantly reduce production. The development of rice blast is favored by high temperatures, relative humidity, and the number of consecutive hours of leaf wetness. A review of various models capable of determining the day-to-day risk associated with infection, mostly based on air temperature, relative humidity, and rainfall, is presented in [24]. Depending on the level of risk predicted by the model, appropriate countermeasures such as fungicide application are taken, thus optimizing and reducing chemical usage. Our high-resolution model aims to provide such a decision support system at the rice field scale.

The LAORE agency delivers a daily bulletin to rice stakeholders showing the Rice Blast Disease Alert Level (LAB—Livello di Allerta del Brusone), calculated for the individual irrigation districts in the areas of Oristano and San Gavino. The LAB can take three values: low alert level (1), moderate (2), and high (3). It is obtained by processing meteorological data collected by the RUR network stations closest to the irrigation districts and weather forecasts from the BOLAM model. The LAB values calculated for individual stations are then interpolated onto a low-resolution grid, with each district assigned the highest LAB value from the pixels within its territory. The procedure for calculating the LAB values for the bulletin issued on day D, i.e., ${\mathrm{LAB}}_D$, makes use of meteorological data from three consecutive days: the measured data from the previous day $D-1$, and the BOLAM weather forecasts for the same day D and the following day $D+1$. The meteorological data input to the model are the hourly mean values of temperature T and humidity H and the hourly rainfall accumulation R. Thus, each station is represented by three time series of 72 elements each: $T_i$, $H_i$ and $R_i$ for $i\in \{1,\dots ,72\}$. The first step is to identify hour i for which the condition of leaf wetness ${\mathrm{LW}}_i$ occurs. For this model, with the leaf wetness sensors unavailable, the condition arises when the relative humidity is sufficiently large or when some rain is recorded:

$${\mathrm{LW}}_i=\left\{\begin{array}{cc}1,\hfill & \mathrm{if}{H}_i>85\%\mathrm{or}{R}_i>0 \mathrm{m}\mathrm{m}\hfill \\ 0,\hfill & \mathrm{otherwise}\hfill \end{array}\right.$$

(4)

Accordingly, the sporulation efficiency function $f\left(T_i\right)$ indicates when the temperature, in the presence of leaf wetness, is favorable for the onset of rice blast infection:

$$f\left(T_i\right)=\left\{\begin{array}{cc}{\displaystyle \frac{T_{\max }-{\mathrm{T}}_i}{T_{\max }-T_{\mathrm{opt}}}}·{\left({\displaystyle \frac{{\mathrm{T}}_i-T_{\min }}{T_{\mathrm{opt}}-T_{\min }}}\right)}^{{\displaystyle \frac{T_{\mathrm{opt}}-T_{\min }}{T_{\max }-T_{\mathrm{opt}}}}},\hfill & \mathrm{if}{\mathrm{LW}}_i=1\hfill \\ 0,\hfill & \mathrm{otherwise}\hfill \end{array}\right.$$

(5)

Here, $T_{\min }=10 °\mathrm{C}$, $T_{\mathrm{opt}}=27.5 °\mathrm{C}$ and $T_{\max }=35 °\mathrm{C}$, which represent the optimal temperature range and value for the fungus development, are chosen based on a literature survey.

The number of hours $W\left(T_i\right)$ required for the infection to develop is given by

$$W\left(T_i\right)=\left\{\begin{array}{cc}{S}_h/f\left(T_i\right),\hfill & \mathrm{if}f\left(T_i\right)>0\hfill \\ 0,\hfill & \mathrm{otherwise}\hfill \end{array}\right.$$

(6)

In Equation (6), $S_h$ represents the minimum number of consecutive hours of leaf wetness required for the rice disease to initiate. The literature typically reports values ranging between 2 and 4 hours; we opted for an intermediate value of $S_h=3$.

The following equation counts the number of hours $W_i^{\mathrm{eff}}$ in which leaf wetness occurs, resetting the count if the condition does not occur anymore in the latest four hours (in fact, if the leaf dries, the infective process stops):

$$W_i^{\mathrm{eff}}=\left\{\begin{array}{cc}{W}_{i-1}^{\mathrm{eff}}+{\mathrm{LW}}_i,\hfill & \mathrm{if}\sum _{j=i-3}^i{\mathrm{LW}}_j>0\hfill \\ 0,\hfill & \mathrm{otherwise}\hfill \end{array}\right.$$

(7)

By combining the above variables, the infection condition ${\mathrm{INF}}_i$ for hour i is achieved when

$${\mathrm{INF}}_i=\left\{\begin{array}{cc}1,\hfill & \mathrm{if}{W}_i^{\mathrm{eff}}>W\left(T_i\right)>0\hfill \\ 0,\hfill & \mathrm{otherwise}\hfill \end{array}\right.$$

(8)

Then, for each day $d\in \{D-1,D,D+1\}$, the risk index (IR) for the rice blast disease is computed as follows:

$${\mathrm{IR}}_d=\left\{\begin{array}{cc}0,\hfill & \mathrm{if}\sum _{h=1}^{24}{\mathrm{INF}}_{d,h}=0\hfill \\ 1,\hfill & \mathrm{if}\sum _{h=1}^{24}{\mathrm{INF}}_{d,h}=1\hfill \\ 2,\hfill & \mathrm{if}2\le \sum _{h=1}^{24}{\mathrm{INF}}_{d,h}\le 3\hfill \\ 3,\hfill & \mathrm{if}\sum _{h=1}^{24}{\mathrm{INF}}_{d,h}>3\hfill \end{array}\right.$$

(9)

where the subscripts d and h mean the hour h of day d. Finally, the alert level ${\mathrm{LAB}}_D$ is

$${\mathrm{LAB}}_D=\left\{\begin{array}{cc}1,\hfill & \mathrm{if}\sum _{d=D-1}^{D+1}{\mathrm{IR}}_d<4\hfill \\ 2,\hfill & \mathrm{if}4\le \sum _{d=D-1}^{D+1}{\mathrm{IR}}_d\le 6\hfill \\ 3,\hfill & \mathrm{if}\sum _{d=D-1}^{D+1}{\mathrm{IR}}_d>6\hfill \end{array}\right.$$

(10)

SMARTerra computes the risk indices and the alert levels for the whole grid of Sardinia with resolution 100 m and for the higher resolution (10 m) grids of the rice districts. This allows us to assign $\mathrm{IR}$ and $\mathrm{LAB}$ values to each single pixel of the grids as opposed to the procedure employed by LAORE which performs the computation for a very limited number of locations.

2.2.6. Forecast of Locust Egg Hatching

The model for forecasting the egg hatching date for the locusts is developed by LAORE based on the FAO eLocust model [25]. The procedure, illustrated in Algorithm 1, utilizes the following time series obtained from meteorological measurements: daily maximum temperature $T_{\max }$, daily average temperature $T_{\mathrm{mean}}$, and total daily precipitation R. Forecast data predicted by the BOLAM model are not used.

Currently, LAORE runs the model for a limited number of locations where the pest is most prevalent, e.g., the Ottana plain. SMARTerra can run the model simultaneously for the whole raster grid of Sardinia by using the interpolated measured data. The procedure identifies, for each pixel, the dates ${\mathrm{day}}_1$, ${\mathrm{day}}_2$, and ${\mathrm{day}}_3$ when the respective conditions are achieved, i.e., (i) the maximum daily temperature is lower than $T_{\mathrm{th}}=30.0 °\mathrm{C}$ for $n_{\mathrm{days}}=5$ consecutive days, (ii) the rain accumulation since ${\mathrm{day}}_1$ is $R_{\mathrm{sum}}\ge R_{\mathrm{th}}=100.0 \mathrm{m}\mathrm{m}$, and (iii) the sum of growing degree days since ${\mathrm{day}}_2$ is $D_{\mathrm{sum}}\ge D_{\mathrm{th}}=200.0 °\mathrm{C}$.

Algorithm 1 Procedure for estimating the egg hatching date of locusts

$\mathrm{today}\leftarrow \mathrm{day}, \mathrm{month}, \mathrm{year}$ ${\mathrm{day}}_0\leftarrow \mathrm{August} 1\mathrm{st}, \mathrm{year}-1$ for i in ${\mathrm{day}}_0,\dots ,\mathrm{today}$ do     $C_i=0$     if $T_{\max ,i}<T_{\mathrm{th}}$ then         $C_i=1$     end if     if ${\sum }_{j=i-4}^i{C}_j=n_{\mathrm{days}}$ then         ${\mathrm{day}}_1\leftarrow i$ return     end if end for $R_{\mathrm{sum}}=0$ for i in ${\mathrm{day}}_1,\dots ,\mathrm{today}$ do     $R_{\mathrm{sum}}=R_{\mathrm{sum}}+R_i$     if $R_{\mathrm{sum}}>R_{\mathrm{th}}$ then         ${\mathrm{day}}_2\leftarrow i$ return     end if end for $D_{\mathrm{sum}}=0$ for i in ${\mathrm{day}}_2,\dots ,\mathrm{today}$ do     $D_i=max\left\{T_{\mathrm{mean},i}-T_{\mathrm{base}},0\right\}$     $D_{\mathrm{sum}}=D_{\mathrm{sum}}+D_i$     if $D_{\mathrm{sum}}>D_{\mathrm{th}}$ then         ${\mathrm{day}}_3\leftarrow i$ return     end if end for

2.2.7. Weather and Risk Rasters

The measured agrometeorological values are interpolated at a resolution of $100 \mathrm{m}$ for the entire territory of Sardinia and $10 \mathrm{m}$ for smaller, specific areas like rice field districts. Hourly raster images are generated for the mean temperature, mean relative humidity, and precipitation, along with daily maps for rainfall accumulation, daily maximum temperature, and daily mean temperature. Additionally, maps depicting the risk index for rice blast disease are generated for the previous, current, and following days, as well as a raster map indicating the alert level. Finally, maps are produced to illustrate the variables necessary for predicting the date of locust egg hatching.

2.2.8. Data Storage

The back-end pipeline generates a substantial daily dataset, approximately 5–6 GB in size. This dataset comprises interpolated measurements and predictions of three agrometeorological parameters (temperature, humidity, and rainfall) across various scales, from small rice-growing districts to the entire region of Sardinia. Additional weather variables may be included in the future, depending on the integration of new risk models in SMARTerra. Over time, the interpolated BOLAM predictions are phased out. This is because the locust egg hatching prediction model requires only data from meteorological stations, while the rice blast disease risk prediction model relies solely on BOLAM forecasts for the current and next day. Moreover, the BOLAM interpolated data are swiftly recalculated using cubic interpolation instead of the computationally intensive KED technique.

2.3. Front-End

The front-end includes a browser-based GIS application that represents the geographical area of interest and visualizes data from various sources, such as numerical series, shapes, and rasters. Users can locate meteorological stations and nearby BOLAM grid points. For each station or BOLAM node, the time series of selected meteorological variables—temperature, relative humidity, and rainfall—can be retrieved from the InfluxDB database and displayed in a dedicated popup window. Users can also compare values between different points. The front-end is divided into two areas: “Weather”, which displays maps with meteorological indicators, and “Pests and Diseases”, which visualizes risk maps. Examples and screenshots from both areas will be shown in Section 3. Here, we mention that users can view both raster maps and time series of either measured weather variables (sampled up to $10 \mathrm{m}$) or forecasts for the next days, both at hourly intervals (see Figure 4). In the “Pests and Diseases” section, users can view rasters associated with the 3-day rice blast (“brusone”) risk maps for the paddy fields at different resolution levels, along with supplementary information on the rice crops and the selected farm. An example is shown in Figure 5 for 12 July 2023. Similarly, detailed and interactive graphical outcomes are provided for locust egg hatching, as illustrated in Section 3.

On the technological side, the geographical data are exposed through GeoServer (version 2.5), and the entire application is written with the JavaScript framework MapWall (version 0.6), developed in-house at CRS4.

3. Results

In Section 3.1 an overview of the maps produced daily by SMARTerra is presented, while its capabilities as a tool for predicting and assisting the countering of pests and diseases in agriculture are illustrated in Section 3.2 and Section 3.3.

3.1. Weather and Risk Maps

The back-end section of SMARTerra generates several maps available online on the dedicated geoportal and published daily in a PDF report accessible only to internal users, primarily rice farmers and technicians involved in the campaign against locust spread. These maps include hourly visualizations of mean temperature, relative humidity, and rainfall accumulation obtained from the KED interpolation of measurements from the RUR network stations. Additionally, visualizations of daily maximum and mean temperature values, as well as rainfall accumulation, are available. Based on this information, the two risk models currently implemented provide maps of rice blast disease risk for all of Sardinia and specifically for the Oristano and San Gavino rice districts, as well as maps of the expected dates of locust egg hatching. SMARTerra is designed to facilitate the inclusion of new risk models. For example purposes, maps from the report showing daily average and rainfall accumulation are displayed in Figure 6. The thumbnails from the PDF report, available only internally, are shown in Figure 7; their purpose is to provide a quick overview of the produced raster maps along with a synthetic recommendation about field treatments from the LAORE agency (not included here).

For end-users, whether farmers or agronomists, the alert levels are of significant interest and can serve as a decision support system for field operations. Figure 8 illustrates the risk indices computed from interpolated weather measurements and forecasts and the alert level obtained from the risk indices (see Section 2.2.5).

As for locusts, Figure 9 shows an example of the prediction model outcome: the cumulative sum of degree days (left) and the dates corresponding to egg hatching (right).

3.2. Prediction Assessment: The Rice Blast Disease

The most significant test for SMARTerra is arguably the comparison with in-field observations, both for rice blast and locust cases. Comparing the predicted results for rice blast is challenging because there are no quantitative parameters to be measured in the field. Counting the number of affected plants is difficult to implement due to the variability in pathology signs among plants, the high density of rice crops (up to 300 plants per ${\mathrm{m}}^2$), challenging access conditions (flooded paddy fields), and the fragile ecosystems involved. Consequently, observations are typically made from several meters away, mostly at the edges of paddy fields, with risk classification based largely on individual experience. In this situation, it is reasonable to seek qualitative agreement rather than a strict, measurable correspondence. Few in-field observations are available as these are time-consuming operations not performed on a daily basis. Generally, SMARTerra fairly predicts the pattern of rice blast spread, as illustrated in Figure 10.

In both comparisons (top and bottom figures), a common pattern clearly emerges between the observations and predictions, although a bias is present in the first case, manifested as a shift (yellow is associated with green and red with yellow). It is important to understand the fundamentally different nature of observation and prediction results. Observations are inherently piecewise due to the limited number of inspections, as previously illustrated, and the labor organization of regional agencies, which does not allow for an extensive and continuous monitoring campaign across all rice districts. Consequently, a constant risk index is attributed to the entire district based on the assumption that variation is not that significant. In reality, agrometeorological data can change rapidly over short distances due to shoreline proximity and other factors. High-resolution predictions made by SMARTerra account for these effects. A fair comparison requires more systematic observation activity. Fortunately, while human experts, typically agronomists from the agencies, handle this activity rather than the farmers, LAORE has recently set up a data collection project based on the smartphone application ArcGIS QuickCapture (version 1.20.21) [26], which is now widely used in research for wildlife and vegetation monitoring [27,28]. This tool allows rice producers to report the appearance of blast in their paddy fields by sending a georeferenced notification, optionally accompanied by a photograph of the affected plants. Given the app’s ease of use and speed, LAORE has exerted significant moral suasion on the farmers, who are very aware of and motivated by the importance of such feedback and data collection. We expect to have more quality data in the coming seasons, enabling better model calibration and optimization and more precise and significant comparisons between observations and predictions.

3.3. Prediction Assessment: The Locust Egg Hatching

Locust presence is monitored by LAORE through field surveys to observe and georeference insect growth. We were provided with survey data from the Ottana Plain area, collected from early spring until the end of May 2023. Each survey record includes the georeferenced location and the locust development stage (“instar”), which ranges from 1 (newborns) to 6 (adults). For our purposes, we compare the dates corresponding to instar 1 with the egg hatching dates predicted by SMARTerra at the same locations. An initial model run, performed with parameters taken from the literature ($T_{\mathrm{th}}=30 °\mathrm{C}$, $R_{\mathrm{th}}=100 \mathrm{m}\mathrm{m}$, $D_{\mathrm{th}}=200 °\mathrm{C}$), shows (Figure 11, top left) values of ${\Delta }_d$, the difference (in days) between the observed instar 1 date and the prediction of SMARTerra (Algorithm 1), spatially distributed in two homogeneous clusters. ${\Delta }_d>0$ means that SMARTerra predicts the locust egg hatching earlier than observed, and vice versa. The light-grey observations have $\left|{\Delta }_d\right|<30$ days and include predictions in good agreement with the observations, while dark-grey hexagons correspond to a distribution of records centered around ${\Delta }_d=150$ days.

SMARTerra helps us understand the cause of this major misprediction. Analyzing Figure 11, middle left, it is evident that the red area corresponding to the dark-grey observations reaches the rainfall threshold $R_{\mathrm{th}}$ at dates (${\mathrm{day}}_2$) approximately 3–4 weeks earlier than the orange area corresponding to the light-grey observations. This large difference in ${\mathrm{day}}_2$ naturally affects the predicted egg hatching dates for ${\mathrm{day}}_3$ (Figure 11, bottom left). In fact, the dark-green and light-green areas exhibit a significant difference of about 4–5 months. This can be explained as follows: since the accumulation of degree days is the summation of the difference, if positive, between the average daily temperature $T_{\mathrm{mean}}$ and the base temperature $T_{\mathrm{base}}=10 °\mathrm{C}$, the pixels in the dark-green area could continue accumulating during the still-warm days of October and November. In contrast, the summation for the pixels in the light-green area essentially halted during winter due to the low temperatures, requiring them to wait for the return of spring.

However, predicted dates can change dramatically with even minor modifications in the model parameters. For example, running the model implemented in SMARTerra with the slightly different choice $T_{\mathrm{th}}=29 °\mathrm{C}$, $R_{\mathrm{th}}=105 \mathrm{m}\mathrm{m}$, and $D_{\mathrm{th}}=200 °\mathrm{C}$ (unchanged) results in a single-mode distribution of ${\Delta }_d$, except for a few outliers, as shown in Figure 11, top right. The map in Figure 11, middle right, indicates that achieving the rainfall accumulation threshold $R_{\mathrm{th}}=105 \mathrm{m}\mathrm{m}$ is slightly delayed in the area under examination compared to the middle left plot of the same figure. Consequently, only three observations fall in the dark-red area, while the others lie within the light-red and orange areas. Finally, Figure 11, bottom right, shows that nearly all instar 1 observations fall within a predicted egg hatching period between March and April. As a matter of fact, by adjusting the model with slightly different parameters, the prediction assessment relative to field observations can be notably improved. Mathematically, this indicates that the model we use is ill-conditioned, meaning small changes in the parameters can cause large changes in the results. This suggests that, on the one hand, a careful calibration of the model parameters is necessary, and on the other hand, the model would benefit from introducing some form of regularization, which would make the model more stable and robust.

4. Discussion

SMARTerra is a near real-time tool built on a WebGIS platform that may be easily extended to any agricultural area with access to hourly data from a meteorological station network. It automatically computes risk indices for various pests and diseases, generates daily bulletins, and sends them to registered users to support crop management planning.

Any risk model based on agrometeorological inputs can be easily accommodated within SMARTerra; we provided two examples concerning rice blast and locust egg hatching. Beyond the service and implementation features, our system is innovative in a key aspect: data visualization and risk index predictions are made at the farm scale, with resolutions up to 10 m.

This approach shifts DSSs for agriculture from the district to the farm level, potentially enabling measured data and risk predictions tailored to specific crops.

Testing the accuracy of emergencies forecast at such high resolution is challenging. For rice blast, the main difficulty lies in the lack of precise experimental data. Monitoring is typically conducted at only a few points per crop and across a limited number of crops per district, leading to a subjective assessment of the presence of Pyricularia oryzae. In this frame, only qualitative estimates are feasible, showing significant pattern similarity.

The case study of locust egg hatching highlights a significant issue that is likely to become more pronounced with the shift towards high-resolution DSSs: the need for a well-defined approach that incorporates regularization techniques to reduce sensitivity to minor parameter changes.

Old, standard approaches relying on district-scale data analysis often employ averaging and pooling strategies, which can act as simple yet effective regularization methods that may obscure very localized dynamics of diseases and pests, as well as their dependence on model parameters. Fortunately, the egg hatching case considered in this study benefits from a substantial number of observations, thanks to the efforts of regional agencies. This allows for the implementation of systematic analysis methods, with a portion of the annotated data reserved for calibration and testing. Such an approach enables fine-tuning the model to accurately reflect the dynamics of locust spread in Sardinia and, more broadly, the spread of crop pests where effective risk models are available. It also supports the training of data-driven algorithms using machine learning or deep learning.

5. Conclusions

In Agriculture 4.0 and beyond, data analysis as a tool for decision support systems will become standard. Advances in sensors and hardware technologies will make collecting data from agrometeorological stations and other sources increasingly affordable. For example, the LAORE agency is currently promoting new initiatives for in-field data acquisition using smartphone technology, which is widely accessible to producers. This will enable the anticipation of crop dynamics, including the risk of diseases and pest occurrence, and pave the way for more sustainable, economical, and environmentally friendly agriculture. Platforms like SMARTerra, which integrate sensor networks, software systems, and advanced mathematical tools for data analysis, will likely become more popular among end-users, including farmers, agronomists, and technicians from local agricultural agencies. This convergence of leading-edge ICT technologies facilitates a shift from the traditional district-based approach to a new paradigm where analysis at the single crop scale becomes possible but also poses some problems, including the following:

• The risk to exaggerate data from individual stations (outliers problem);
• The need for improved tools for the interpolation of irregularly distributed data;
• The potential exposure to ill-conditioned mathematical models;
• Data confidentiality concerns as the high-resolution approach can reveal information at the individual farm scale, posing a risk of market disturbance.

Some of these issues are addressed in a forthcoming paper [7]. The last point requires careful consideration on a case-by-case basis and close interaction and feedback from end-users, including experts in farm business planning and economics.

Several promising future developments can be anticipated, starting with a more systematic strategy that integrates simple deterministic algorithms with advanced, data-driven models, typically based on machine learning. This technology is widely recognized for its ability to combine different types of data and includes effective regularization tools and other valuable numerical resources. This would also allow for the integration of remote imagery from public or commercial satellites to study the interactions between soil properties and the complex ecosystem, including plants, insects, animals, and disease agents [29].

Beyond improvements in data analysis and mathematical methods, substantial advancements are also expected from the agronomic side. For example, our model for predicting the spread of rice blast does not consider the effects of the disease on different rice cultivars. This is a significant point [24] that also pertains to the observation phase, which currently treats susceptible and resistant cultivars the same. This distinction can be incorporated into the risk model through careful parameter calibration, an approach that will become feasible with the expected increase in the dataset.

As previously mentioned, both the back-end and front-end of SMARTerra are designed to easily accommodate additional risk models for agriculture based on agrometeorological data, including those available from the ARPAS network that are not currently used (e.g., solar radiation or wind data), thus potentially addressing a wide range of pests and diseases. It also includes crop emergencies that were not considered important in the past but have become increasingly relevant due to changing climate conditions and external factors.