DSS fertirrigation system: An Italian case study

To the plan it is ABSTRACT The proposed paper deals with a Decision Support System (DSS) conceived for agronomists and farm managers involved in the fertirrigation (FER) process. The DSS has been developed within the Smart District 4.0 (SD 4.0) project, funded with the contribution of the Italian Ministry of Economic Development (MISE), aiming to sustain the digitization process of the Italian Small and Medium Enterprises (SMEs). The DSS is part of an integrated framework, mainly including advanced services for blockchain traceability and agriculture planning, optimizing the cooperation among different actors, such as farmers, agronomists and fertilizer suppliers. The most relevant and challenging part of the platform consists of Machine Learning (ML) supervised and unsupervised algorithms with high performance, to optimize production and quality in viticulture. Specifically, the Probabilistic Neural Network (PNN) model provides the performance of time shortage prediction when compared with other ML algorithm performance. Finally, the clustering k-Means algorithm is applied to extract information about active substances, doses, time of shortage, adversities, and pathogen elements, useful to optimize a FER plan. All the analyzed algorithms are applied to an experimental dataset. The work starts with the design of the processes and of the information flow able to optimize grape supply chain and traceability, and then focuses on data analysis. The data analysis approach is suitable to formulate precision farming rules.


INTRODUCTION
Fertirrigation (FER) is a growing field that is based on the mix of agricultural and technological knowledge and practice (Maloku et al., 2020). The technological domain of the FER is based on the following elements: wireless sensors, soft computing and Internet of Things (IoT) (Njoroge et al., 2018), analytics, Decision Support Systems (DSSs), cloud storage data (Power et al., 2019), geographic information system-GIS, precision planting (Kumar et al., 2017), global navigation satellite system (Marucci et al., 2017). FER systems can be integrated with 5G technology (Razaak et al., 2019). FER reduces the waste of resources (Sanghera et al., 2020;Sureshkumar et al., 2017). FER systems remove the presence of weeds from the soil (Lottes et al., 2018;Olaniyi et al., 2020). Robotization can boost the application of FER (Pretto et al., 2019;Fawakherji et al., 2021). FER improves the managerial efficiency of agricultural firms (Paustian et al., 2017) increasing the productivity of the soil (Baek et al., 2019). The usage of FER systems has a positive effect on promoting the efficiency of the nutrition method and the reduction of pollution (Septar et al., 2019). The implementation of photovoltaic systems can also improve the efficiency of a FER system (Shamsuddin et al., 2020;Abdullah et al., 2020). The efficiency of FER systems can also be improved with the usage of IoT (Visconti et al., 2020) and data management (Subić et al., 2016). The application of information technology can have a positive impact on FER (Incrocci et al., 2017) also in developing countries (Shraddha et al., 2018;Jacobs et al., 2018). Applications can also have a role FER management (Pérez-Castro et al., 2017). The application of FER systems, based on robotization, data management and IoT (Zamora-Izquierdo et al., 2019) has a positive role in reducing water scarcity Jiménez-Benítez et al., 2020;Kachor et al., 2019;Gaikwad et al., 2020), diminution of water waste Salvi et al., 2017) and optimization of nutrients (Dong et al., 2018;Karunanithy et al., 2020). The application of FER techniques also has a positive impact on the ability to calculate the appropriate metrics for planning irrigation (Smith, 1992;Wu, 1992) and evapotranspiration (Allen et al., 1998). The application of methods of estimation and prediction, can be optimized through the usage of statistical and computational methods based on Machine Learning (ML) (Choudhary et al., 2019;Krupakar et al., 2016). The application of fertilization plans is based on the estimation of quantities of fertilizers and nutrients present in the soil (Regar & Singh, 2014;Chuan & Zhao, 2016;Xu et al., 2017;Brandoli et al., 2021). The quantitative estimation of the mixture of fertilizers, water and nutrients is mainly based on statistical and computational methods i.e., ML (Chlingaryan et al., 2018;Qin et al., 2018;Nikoloski et al., 2019;Pedersen et al., 2019;Pohl et al., 2017). To optimize the FER plan it is also important to collect information about other variables such as: precipitation, temperature, humidity, pressure, dew point, wind speed, and intensity of solar radiation (Schlenker & Roberts, 2009;O'toole & Hatfield, 1983;Fritschen, 1967;Fischer, 1985). To perform an improved FER activity, it is necessary to gain the knowledge about the nutrients of the soil (Moreno & Garcia, 2018;Bondre & Mahagaonkar, 2019). Concerning irrigation and FER advances, specific methods can be adopted using DSS models based on data field detection Massaro & Meuli et al., 2018) and data fusion approaches (D' Accolti et al., 2018). Another technology useful for the traceability of field production processes is the blockchain (Massaro & Maritati et al., 2020). The described state of the art, highlights that the FER technologies and approaches can be different. For each crop can be developed specific approaches and technologies: the complexity is mainly to find the technology matching with the specific pilot case study, starting to the process optimizing. In this direction, a first step is to implement a software platform able to collect field data and actor's information. The basic principle to construct a FER DSS is to design and implement an information system platform, allowing the communication between different system actors such as agronomists, farmers, fertilizer suppliers, and analyzing collected data by means of specific methods and ML algorithms. Specifically, the study proposed in this paper is part of the Smart District 4.0 (SD 4.0) project, supported by the Italian Ministry of Economic Development (MISE), with the aim of stimulating the widespread digitization processes of SMEs in some typical sectors. In particular, as regards the Agro-Industrial sector, SD 4.0 has undertaken an experimental test case aimed at enhancing a solution for precision farming developed by Asepa Energy srl, a small company with high expertise and skill in industrial automation. Thanks to the proposed platform, it is possible to automate remotely combined cycles of irrigation and fertilization. The goals of the pilot case study of the project is to: • Provide a single method of access to the various users of the system platform: farmers, agronomists, suppliers and possibly actors in the distribution chain, such as buyers, logisticians and consumers. • Guarantee the certified traceability of field operations via blockchain; • Guarantee the management of stocks in the warehouse, with the automatic supply of fertilizers; • Allow the agronomist to be able to set the FER plan also in terms of proposals automatically generated and based on the updated status of the production site (sensors, weather, etc.) using automatic learning logics; • Provide graphical dashboards useful for precision farming in the viticulture field.
Following the main goals of the project, the results of the proposed work are discussed in the following steps: • Design of the information system of the pilot case study describing in details dataflow architectures; • Production process design by means of the Business Process Modeling Notation (BPMN) approach; • The description of blockchain and platform FER planning dashboards; • Comparison of different ML algorithm performances suitable for the prediction of the important variable of the time of shortage indicating company productivity; • The adoption of the clustering k-Means model to provide graphical dashboards indicating the effectiveness of fertilizers. Figure 1 is illustrated the architecture of the FER system related to the case of study: the architecture indicates the data flow between the whole SD 4.0 platform involving field data of a central unit located in a field, and other data available in the cloud.

Architecture design
The DSS is integrated into the platform and processes different data such as weather data and predicted ones, products used for pathogens and adversities, product doses and times of shortage. As shown in the architecture of Figure 2, the field data transmission is enabled by MQ Telemetry Transport or Message Queue Telemetry Transport (MQTT) messages. Input sensor data and field information are processed by ML algorithms (core of the DSS system), providing fertilizer and irrigation recommendations and predictions.

BPMN processes
From the analysis of the AS-IS processes, substantially manual procedures are revealed which allow to trace field operations in a paper book. All the operating processes are not automatized, and the agronomist used a paper scheduling the FER procedures to perform. The actors involved in the optimized TO-BE processes are therefore substantially the same as those considered in the AS-IS processes, by adding the FER manager reading the DSS ML outputs. Furthermore, the new process has added to the registration (notarization) of information via blockchain.
All the actors involved in the new processes are therefore the FER manager, the agronomist, the farmer and the fertilizer manufacturing company providing commercial field treatment. In Figure 3 is illustrated the BPMN diagram of the first TO-BE use case, involving the various actors of the system. This diagram is based on the concept that, concerning the FER phases, there are three different methods: • Manual fertigation: in this case the farmer can fertirrigate the field by activating the SD 4.0 system remotely using a mobile phone, sending the commands about the irrigation, the type of fertilizer and the doses to be used; • Scheduled fertigation: the farmer can set up a calendar containing a FER plan by specifying information such as date, start time, end time, amount of water, type and amount of fertilizer to adopt; • Automatic fertigation: in this case the SD 4.0 system proceeds with the automatic FER of the field, detecting data from the field sensors and applying the FER model set.
The manual and the scheduled procedures are optimized by setting some threshold values able to control the FER process. The farmer and the agronomist pools complete the first case study: the farmer set doses, data, operating hour, and could enable the automatic FER plan, while, the agronomist creates and updates the FER model by suggesting to the farmer the actions to perform according to the results provided by the DSS.
The BPMN diagram of Figure 4 indicates the second use case describing the SD 4.0 platform's main functions of traceability of treatments via blockchain and inventory management. The use case is able to trace the life cycle of the product transcribing the related information useful also for agro-industrial companies which could manage and monitor the entire production chain, learn about the trend of crops and the cultivation techniques adopted by suppliers. In this use case, a notarization procedure is proposed by the farmer relating to any information on the products used (fertilizers) and any activity performed in the field (sowing date, seed type, position of the field, associated area, treatment date, etc.). Information notarization via blockchain, offers in the first instance the possibility to the final customer or supply chain operators, to retrieve all the information on a specific product by simply reading the QR code as packaging label of the purchased product. In the same formal model, the process of replenishing the fertilizer stocks is also managed based on the trend in consumption related to their use during the season. The SD 4.0 platform receives the message, acquires all the information associated with the particular event ID from the DB, and notarizes them through blockchain. Then the platform updates the stocks and verifies the availability for other programmed FER plans. In the case of product unavailability, the platform automatically sends a notification to the fertilizer producer enabling the purchase process.
In Figure 5 is illustrated the third use case involving the "core" actor of the agronomist, enabling processes of creating and updating the FER plan. The goal of this use case is to create an intelligent and automatic FER system through the use of ML techniques, maximizing crop yields. The factors contributing to the crop yield are the environmental factors such as meteorological factors (rainfall, temperature, humidity, air pressure, dew point, wind speed, intensity of solar radiation), culture factors (type of crop, morphological composition of the soil, evapotranspiration conditions), and agronomic factors (quantity of fertilizer or water used in the last irrigation, estimated shortage time). The FER model is set using a calendar, where for each FER event the agronomist will have to specify the related information.

PLATFORM FRAMEWORK
The Unified Modeling Language (UML) scheme of Figure 6 indicates the Use Case Diagram (UCD), including all system functions and actors, and shows SD 4.0 platform data flow. In the diagram is distinguished the data warehouse system (Google Cloud BigQuery), and the relationships between all actors. As main DSS outputs are indicated in the figure the stock management indicators, the intelligent FER plan, the weather forecasting and data clustering dashboards. Figure 6 is representative of the BPMN "TO BE" process in Figure 5. The DSS "core" is constituted by the Konstanz Information Miner (KNIME) artificial intelligence algorithms. The adopted KNIME workflows are objects oriented Graphical User Interfaces (GUIs) suitable for industrial applications Massaro, 2021;. More details will be provided in the Results section.

SD 4.0 frontend implemented interfaces
The SD 4.0 platform is developed by using different technologies embedded into a unique framework. In particular, the Ethereum technology is adopted to include notarized information into blockchain blocks and to trace product transactions. In Figure 7 is illustrated a screenshot of the frontend interface controlling blockchain data.
The frontend interface related to SD 4.0 initial FER plan creation is shown in the Figure 8.
When a FER plan is created a unique Identification number (ID) is associated with each FER event. In Figure 9 is illustrated the Graphical User Interface (GUI), composed of the different fields structuring a FER plan.
The output of the created event is summarized by the layout interface of Figure 10.

DSS data processing of the experimental dataset
Some functions of the DSS are forecasting and the data clustering supporting agronomists to define FER plans. The experimental dataset concerns two years of FER process, related to three fields used for three grape typologies (Red Globe, Italia, Selvatico), and contains the following attributes for a total number of 94 records:      Where some commercial fertilizers have the same name as the active substances. The dataset is sorted by date. A first approach to analyze experimental data, is to find the best supervised algorithm for forecasting. In Table 1     Trees, and Linear Regression algorithms. The performance is estimated by adopting Konstanz Information Miner (KNIME) workflow [59] using the standard hyperparameters indicated in Table 2 (parameters provided by the KNIME tool as default parameters).
In order to optimize the metric for the best performance estimation, it is defined the BestAlgorithm of Eq. (1), defins the final algorithm ranking. By observing the results of Table 3, the PNN algorithm represents the best algorithm for the fixed hyperparameters. Different results can be achieved by changing hyperparameters by adopting an approach based on error minimization , but the proposed method is useful to find a primary an algorithm suitable for the experimental dataset, by considering also the BestAlgorithm indicator able to address the choice by estimating the sum of the Mean Absolute Error (MAE), Means Squared Error (MSE), and Root Mean Squared Error (RMSE). (1) In Figure 11 is illustrated the KNIME workflow implementing PNN algorithm composed by the following blocks (workflow objects): • Excel Reader: reader of dataset exported from BigQuery datawarehouse in the local repository; •  Figure 12 is illustrated the trend of the Pesticide dose during two operation years, where more records can coincide with the same data explaining why the second season appears to be longer.
The Time Shortage trends related to the two analyzed seasons, and to the prediction of the third year, are plotted in Figure 13: the days are limited to a period range between 10 and 35 days with some peaks observed in the first year and in the predicted one. The result of Figure 13 proves that for the three analyzed grape fields the human work is and will be controlled (correct execution of the FER plan).
The second approach adopted for data analysis is to find data clusters supporting dataset reading, and providing information about precision farming for the specific case study.   Table 3: Ranking related to the tested algorithms using the BestAlgorithm indicator Figure 11: KNIME workflow applied to the experimental dataset, implementing PNN algorithm.    In order to obtain DSS graphical dashboards, the KNIME workflow of Figure 14 is adopted implementing the k-Means algorithm. The choice of the cluster number k is defined by estimating the Silhouette Coefficient (SC) (Aranganayagi et al., 2007). The estimated SCs are listed in Table 4: a good approach is to select the k number providing the best overall SC number for the case of k =6.

DISCUSSION
By fixing k = 6 are estimated the clustering dashboards plot two variables for each graph. Reading the dashboards of Figures 15-19, the information resumed in Table 5 has been achieved, useful to optimize a FER plan.

CONCLUSIONS
The proposed work discusses the results of a project based on the technological improvement of a company working in FER. By Adopting a software platform, new processes are implemented to formulate FER plans according to data analysis criteria. In this direction, DSS dashboards of supervised and unsupervised algorithms provide a further tool supporting agronomist decisions focused on the analysis of key parameters such as adopted active substances and related doses, time of shortage and adversity or pathogen elements. The paper is oriented to describe the data flow of the pilot case study showing methods and approaches useful for data processing and data analysis. The proposed methods can be extended to other typologies of cultures requiring precision farming, and to Industry 4.0 systems.

ACKNOWLEDGEMENT
All the applications have been deployed by a unique IT collaborative framework developed within the Smart District 4.0 Project: the italian Fondo per la Crescita Sostenibile, Bando "Agenda Digitale", D.M. Oct. 15 th , 2014, funded by "Ministero dello Sviluppo Economico". This is an initiative funded with the contribution of the Italian Ministry of Economic Development aiming to sustain the digitization process of the Italian SMEs. Authors thank to the partner Noovle for the collaboration provided during the work development. The proposed results are used to verify the usability of the data on the platform.  Figure 18] The time of shortage usually does not overcome the 35 days using each active substance.