AN EXPERT SYSTEM FOR VINEYARD MANAGEMENT BASED UPON PERVASIVE COMPUTER TECHNOLOGIES Giuseppe AIELLO, Mario ENEA, Cinzia MURIANA Dipartimento di Tecnologia Meccanica, Produzione e Ingegneria Gestionale Università degli Studi di Palermo Viale delle Scienze, 90128, Palermo Italia aiello@dtpm.unipa.it ; enea@unipa.it; muriana@unipa.it Abstract: Determining the optimal maturity level for performing viticulture operations and harvesting activities is a difficult task, because, depending on the variety, the climatic conditions and cultural practices, the phenologic maturation process occurs at different times. Recently, ubiquitous computing technologies allow an extremely precise and cost effective monitoring of environmental conditions by means of an RFID based sensor networks. The implementation of such technologies in vineyard management is nowadays under development, however, besides the possibility of gathering data, the need is perceived of developing decision support tools to fully exploit the potential opportunities of these new technologies. The present research aims at establishing a suitable method to support the decision process with the environmental data gathered automatically by a sensor network. The paper reports the results of an experimental study on a Sicilian vineyard showing that by means of the data collected by an RFID infrastructure it is possible to forecast the occurrence of phenologic maturity stage. Keywords: Decision support system, RFID, Production planning INTRODUCTION Sicily is the Italian region with the highest winemaking heritage and wine production constitutes a fundamental resource for the local economy accounting for 15% of the gross output of the entire agricultural turnover. In order to achieve high quality production, viticulture operations and harvesting activities must occur at the correct phenologic maturation level. In particular scheduling of the harvest operations has a fundamental role, as in fact grapes that are harvested before a desired maturity, result in the production of acidic wines while late harvesting generally results in unbalanced fruit composition. Determining the optimal maturity level for performing viticulture operations and harvesting activities however is a difficult task, because, depending on the variety, the climatic conditions and cultural practices, the phenologic maturation process occurs at different times. The environmental conditions that essentially affect grape ripeness are temperature, relative humidity, solar irradiation, and rainfall. In particular the temperature plays a fundamental role for grape maturation, including the aroma and the coloration [4]. In order to achieve the desired quality, hence, fruit maturity level must be closely monitored. Although several indices have been proposed, ripeness is currently an entirely subjective judgment, based upon the right mix of phenolic compounds, aroma, colour, sugars and acidity levels. In particular, the evaluation of the sugar and acid content, or ph of the berries, is of fundamental importance to manage the vineyard operations and schedule the harvesting process [2]. The sampling activities related to the evaluation of the maturity level must be manually performed, thus resulting in a costly and time consuming process. In such sense recent ubiquitous computing technologies may represent an effective solution which easily integrates with existing decision processes, providing extremely detailed and reliable information about environmental data. Pervasive computing technologies such as sensor network systems in fact give new capabilities for sensing and gathering data about an environment and new digital processing opportunities. Innovative infrastructures based on Wireless Sensor Networks (WSN) are a real-time, pervasive, non intrusive, low-cost, and
highly flexible data analysis technologies that can ensure high accuracy in detecting climatic conditions on the ground. Recently, WSNs have been employed in the specific area of farming monitoring and a few preliminary works describe applications for precision agriculture. [1]. According to such considerations, in this research the decision process regarding the timing of the vineyard operations is analyzed and a model is presented in order to forecast the occurrence of phenologic maturation stages on the basis of the information gathered by a wireless sensor network. The opportunities of practical application of the proposed methodology have also been verified against an experimental campaign. Sensor based vineyard management system In this section the decision processes concerning vineyard operations are investigated and a proactive vineyard management system is proposed to be employed in conjunction with ubiquitous computing technologies. The analysis of the decision processes involves the preliminary identification of the objectives they are supposed to fulfill and the subsequent formalization of the input data and the methodology required to produce the output decisions. Typical winemaking operations and harvesting decisions in particular, involve the interaction between the enologist and the vineyard manager, which have different priorities and objectives. The enologist is mostly concerned about wine quality, while the vineyard manager considers more specific agricultural variables, including operational costs. The assessment of grape maturity level is a primary information for winemakers and enologists. The date when optimal maturity is reached or when a new phenologic phase is entered, varies depending upon the quality of wine, the varietal typology of the vines, the site climatic conditions, the seasonal specific factors and the viticulture practices. Due to the complexity of the decision processes involved in vineyard management, viticulturists have developed an assessment method based upon the establishment of a set of indicators to be evaluated experimentally by periodically sampling the vineyard. The decision about the harvesting time grapes is such an important decision that most winemakers and grape growers start sampling grapes several weeks before the harvest time and continue sampling with increasing frequency as harvest time approaches. The maturation process is then monitored on the basis of maturation curves reporting sugar content, ph and total acidity of the samples. According to the previous considerations, the main drawback of the current winemaking practice is that it requires a direct sampling process of the berries, which must be manually carried out. The present research aims reducing the amount of manual sampling operations by supporting the decision processes with the environmental data gathered automatically by a sensor network. In particular the decision process here adopted is referred to the well known growth models proposed in the literature. Branas (1946), Winkler (1975) and Huglin (1986) in different years have studied parameters that affect grape ripeness and proposed some indices that allow to monitor the development of grapes. According to such models the five phenological phases (sprouting, flowering, fruit set, veraison, ripening) which characterize the grape maturation process can be related to environmental temperature. Based on studies conducted from the aforementioned authors, grapes ripeness can be related to the heat that the plant has stored during its growing period, that allows it to move from one phenological phase to the next. Heat needed to reach one phenological phase is commonly expressed through the heat quantity that a plant can store depending on daily temperatures, that define the Sum of Active Temperatures (STA), expressed in degree day (DD). STA allows to express phenological cycle length or single phase length in terms of thermal units according to (1): STA (1) where T Max and T min are the maximum and the minimum of daily air temperatures; if T Max is more than 35 C T Max is posed equal to this value; cardinalemin is the zero vegetative, i.e.
sprouting temperature. On the basis of this growth model some researchers have proposed numerical indexes capables to predict when one phenological phase happens. In our study we will refer to Winkler Index that is calculated as in (2): / / 10 (2) where negative values of (T med -10) should be set to zero. The establishment of referenced growth models based upon the aforementioned index allows to determine the phenologic maturation process by means of suitable thresholds. Winkler index varies depending on the quality of wine you want to produce. The typical values of the thresholds range from 1200 to 1400 for quality of wine as Cabernet Sauvignon (Red) or from 1800 2000 for Nebbiolo (Red) and Malvasia (White) [3]. According to the above considerations, Heat summation based growth models in conjunction with thresholds allow to roughly assess the maturation process and approximately establish the optimal dates for viticulture operations. In this research the use of heat summation based growth models is supported by extremely detailed micro-climatic information measured continuously by means of a sensor network. This methodology will not allow to predict exact harvest dates at the beginning of the season due to inherent approximations and to the many variables that influence the rate of fruit ripening. However, with experience and record keeping, it is possible to make reasonably accurate projections as the season progresses. It is also expected that the performance achievable with the initial setup of the system, based on referenced parameters and thresholds will increase as actual records of different seasons will be available. At the initial deployment of the system, the dates of major phenological stages, along with harvested yield and fruit quality parameters obtained by standard fruit sampling procedures should be recorded for each area of the vineyard together with local weather records and degree-day accumulations, in order to tune up growth models and the threshold values thus improving the quality of the assessments. This is a common feature in the deployment of expert systems. The relationship between temperature (heat summation) and grapes growth, in fact, defined by the degree-day, combined with developmental thresholds can be exploited to estimate the optimal timing of viticulture operations. The ripening condition in fact can be related to a Cumulative Degree Days threshold, representing the total number of degree days necessary for the grape berries to reach the ripening stage. In such conditions it is possible to predict the effects of the changing seasons and climatic conditions during the growth period of the berries that determine an initial slow heat accumulation process, which dynamically increases as the weather gets hot in summer, thus resulting in a non-stationary process. In such context, the most common forecasting methods are adaptive and nonadaptive regression models, generalized exponential smoothing methods. The classical Bayesian linear regression models are unable to reproduce some of the features frequently observed in non-stationary processes, while, on the contrary, in such cases time series methods are extremely effective. In this research the growth of the berries is predicted by means of the previously discussed growth model on the basis of the heat summation. The heat summation value is updated daily with field measurements, and results in a nondecreasing series of values, originating trended time series. Future values of heat summation can hence predicted by means of the well known Holt s model, which is an exponential adaptive forecasting method for trended data, based on (3), (4), (5): (3) 1 (4) 1 (5)
The three updating equations result in the evaluation of the updated component at a future time t as a weighted average of the (adjusted) previous estimate and the most recent information acquired at time t, while, the trend is updated by averaging the previous trend component (Tt 1) with the difference between the two most recent level estimates (the trend is defined as the change in level). On the basis of the updated component and trend at time t, the forecasted value can be evaluated for any future time t+k. This method requires the establishment of two smoothing constants, α, and β. A common approach is to determine the values of α, and β that minimize the mean or median absolute error, or a similar measure. Finally, as stated before, the decision about the harvesting time is based upon the comparison of the predicted value with a pre-established threshold. The violation of this threshold means that a new phenologic phase has been entered. In order to support the decision process, hence the estimation of the probability of the forecast violating a threshold must be evaluated. This is generally accomplished by estimating the mean and variance of the future observations and relaying on the assumption of gaussian error terms to estimate the probability of violating the threshold. The standard deviation of the tracked data can be calculated as usual referring to the Jensen s inequality for the Gaussian distribution. According to such relation, the standard deviation is about 1,25 the mean absolute deviation. The Probability of the signal exceeding the ripening threshold T (see Figure 1) can hence be evaluated as in (6): 1 (6) Figure 1: Probability of exceeding a fixed threshold. For the decision model hence, the establishment of an acceptance probability (A) is required, resulting in a risk of accepting the hypothesis of the achievement of the optimal ripening level at time t+k when it is not (1-A), and rejecting this hypothesis when optimal ripening level is achieved with probability (A). In conclusion, the methodology proposed for predicting the harvest date involves getting available information on the lower developmental threshold temperature, cumulative degree days, and the observed temperatures for the area considered. Additionally the proposed system takes into account the possibility that the same variety may not behave in the same ripening state from one area to another. Ubiquitous Computing for climatic data acquisition The growth forecast model above described is based on the measurement of local temperatures for the evaluation of the heat summation index. The practical application of the proposed methodology therefore required a reliable measurement system allowing fast and precise evaluation of local micro-climatic parameters such as air temperature, relative humidity, and solar irradiation.[4]. Sensor based ubiquitous computing techniques are spreading nowadays as a cost effective technology for such purposes. In this research a sensor network has been designed and deployed in the vineyard in order to have extremely
precise local estimates of the air temperature. Deploying such technology in a vineyard requires the definition of the density and number of sensors according to the vineyard topology. Vineyards are typically organized into a hedgerow system, which is characterized by a supporting structure made of zinc-plated iron, wooden, or concrete poles, and some lines of steel wires to hold the vine canopy. The microclimate of the grapevine is affected by the environmental conditions of a limited area close to the rootstock. In order to obtain more accurate data than weather station or satellite monitoring, temperature wireless sensors have been placed in each pole in the vineyard. Battery powered. Sensor nodes placed along the rows of grapevines form a connected multi-hop network which, once configured, runs unsupervised. For the purpose of this research Sensor nodes placed in the vineyard are equipped with temperature sensors only, however different technologies allow the evaluation of light exposure, and humidity. Such elements, therefore, can be further integrated in the decision support system as soon as suitable decision models are developed. Temperature Data are collected via the wireless network, gathered at a central storage unit. Commercially available boards have been used equipped with the Sensirion SHT11 combined temperature/relative humidity sensor. Temperature range of the sensor is (-40 C -123,8 C), while Humidity range is (0 100% RH). The distance between the hedgerows is 2.20 m, while iron poles are positioned at 80 cm from each other. Three nodes were positioned on each pole at 90 cm, in order to obtain measurements about the micro-climate at the productive area of the grapevine, measurements about the micro-climate of the leaf-covered area, measurements from the top of the green canopy to be used as reference for the lower areas; all nodes were TelosB motes equipped with temperature and relative humidity sensors. An immediate advantage arising from the adoption of a WSN-based approach is that corrective actions on the cultivations may be timely and selectively chosen; furthermore, the system allows to build a history of past events, and stored data may be analyzed in order to extract potential hidden correlations among the sensed environmental variables and the obtained result. The availability of a considerable amount of precise data, superior to what is commonly attainable through traditional random sampling, allows for the construction of accurate models, and thus favors the proposals of improvements in the cultivation process. In addition to this parameters surveyed around the plants allow us to determine more accurately the microclimate in the vineyard and, consequently, assess the ripening process. Experimental analysis In order to evaluate the effectiveness of the proposed methodology an experimental analysis has been carried out involving the forecasting of the harvest date on the basis of the experimental values gathered by means of the previously described sensors network. The experiments have been carried out in the area of Monreale (Sicily) where many varieties of DOC wines (Ansonica or Insolia, Cabernet Sauvignon, Chardonnay, Muller Thurgau etc.) are grown. Monreale has the typical Mediterranean climate, with mild and rainy winters and hot and dry summers; with an average annual rainfall of approx. 700 mm and average annual temperature of 18 C. The experiments were carried on Chardonnay variety, which typically sprouts between the third decade of March and the first decade of April, flowers between the third decade of April and the second decade of May and ripens between the second and the third decade of August. The vineyard has been divided in 15 zones with homogenous slope of the soil and solar exposure. Sensors have been set to record temperature values every hour from April 2008 to May 2008. Data collected were analyzed to determine daily average temperature for each zone in order to calculate Winkler index. As a result of our study we evaluated the date of flowering related to each of 15 areas. The Winkler index (expressed in DD) for the period from April to May for each of 15 areas under study is shown in Figure 2.
MITIP 2009,, 15-16 Octobe er, Bergamo 700 600 500 400 300 200 100 0 W Winkler Indexx 8/4 1/4 1 15/4 22/4 29/4 6/5 13 3/5 20/5 27/5 F Figure 2: Winklerr Index of 15 zo ones of vineyard d. To determine e start and end flowering da ate we referre ed to the refere enced thresho olds given in Table 1 that show medium m Winkler ind dex values for the major stages of grap pes ripeness calculated ovver the three years y for chard donnay varietyy [3]. Cu ultivar Start Flo owering E End Flow wering Sttart Vera aison En nd Vera aison ation Matura Charrdonnay 270 0 ± 30,2 386,9 ± 3 32,2 1010,3 ± 82 2,7 125 56 ± 96 6,3 1319,6 ± 130,,5 Table 1. Trien nnal average (20 005-2007) of Winkler Wi index durring the phenolo ogical phase in some s present cultivars in DOC D Monreale appellation a (ave erage ± standard d deviation). Results are shown in Ta able 2, where the average e value of the e days to Sta art and End d the range be etween them are a calculated for all zones. The average value of the flowering and days at Startt flowering and days at End d flowering arre 30 and 41,33 respectively, while the average valu ue between sta art and end flo owering is 11. Averag ge Days from m 1 April to Start flowering f Days from 1 April ering to End flowe Disttance between Start and End floweriing 30± 3,288 41,33± 2,7 748 11,00 Table 2. Number of zones Figure 3 repo orts distribution of Start date es for every off 15 zones of vineyard. v 4 3; 4; 5; 14 8; 9;13 3 11; 12 2 1 10 7 6 2 1 4/5 6/5 15 0 24/4 26/4 28/4 30/4 2//5 Da ay Start flowe ering Figure 3. 8/5
Winkler index has highlighted that the 15 zones under study have different date of flowering, due to different sunlight of every zone or different soil slope. This is an important information to realize a precision farming policy. Predicting when every phase happens allows to decide when is the moment to fertilize or when it needs to intervene to avoid the peronospora. In the same manner Winkler index calculated from April to October allows us to predict optimal harvest date and to plan harvest for every zone of the vineyard. As we have already said one of the most important goals of employ of eliothermal indexes is the possibility to predict when a thermal phase happens. To do this we analyzed the thermal sum of Winkler index for all zones of vineyard. Value of Winkler index expected with linear regression model and Holt s model are compared with those calculated with data collected and some parameters were calculated. As we have said in the previous paragraph to apply Holt s model it needs to determine for every zone of vineyard α and β values. For every zone we have determined the value of these parameters that minimize the Mean Absolute Error. The values of α and β determined for the area examined are respectively equal to 0,99 e 0,5446. Through the estimates made by the Regression model and Holt s model Winkler index value were determined for each zone of vineyard. Figure 4 reports expected value of Winkler index for one zone of vineyard. 550 Data Observed 350 150 Holt Linear Regression 26/4 29/4 2/5 5/5 8/5 11/5 14/5 17/5 20/5 23/5 26/5 29/5 Figure 4: Winkler index predicted with Holt s model and Linear Regression for one zone of vineyard. Table 6 reports MSE, MAD, ET and SD value for one sample of data (26 April- 31 May) of the same Zone of vineyard. They indicate how the expected value of Winkler Index differs from that observed. As you can see the Holt s model seems to predict Winkler index value better than the Regression model. This result is verified in almost all zones of vineyard. Average values of MSE, MAD, ET and SD, calculated for all zones resulted respectively equals to 657,98, 20,53, 194,47, 25,66 for Holt s model and 1610,57, 29,91, -302,95, 37,39 for Linear Regression. Holt s model Linear Regression MSE (DD) 981,78 6713,11 MAD (DD) 27,37 67,37 ET (DD) 153,07-2425,28 SD (DD) 34,22 84,21 Table 3. Based on the thresholds chosen for start and end flowering dates of start and end flowering predicted with the two models for every of 15 zones were calculated. Because the expected
values are affected by uncertainty because of their variance, it is useful to determine the probability of a single value exceeding a threshold chosen, for example, the value of the Winkler start or end of flowering. The normal distribution can be used to determine R(σ,µ t ),where σ is the standard deviation of data provided and µ t is the value of Winkler index provided at the time t. Results are shown in Figure 5, where the probability (1-R), to exceed the threshold of beginning of flowering was calculated for each of the two linear models used for forecasting. (1-R) 1,0 0,5 0,0 Figure 5: Probability of exceeding the threshold of beginning flowering. Conclusions In the present research a method that allow to predict the evolution of maturation process for the grape wine is proposed. The research aims to demonstrate that traditional systems can be replaced by innovative and non expensive technologies such as sensors networks. However, these capabilities pose several questions in the application space regarding for example what data should be gathered and how often, how information must be processed and how should the result be presented to the user, how can the knowledge based be employed to support the decision processes. Such issues must be considered in the general framework of the design of farming decision support systems, involving the analysis of the decision processes, the establishment of decision parameters, the identification of required information and datasets and the selection of the most suitable technologies for data acquisition. As a consequence, the potential of such technologies is currently exploited to a very limited extent due to the lack of suitable decision models and post-processing procedures to be included in farming decision support systems. The proposed methodology can be effectively employed to plan and schedule the harvesting operations taking into consideration the optimal maturity of the berries and coping with constraints related to the limited stocking capacity of the winery and the availability of the harvesting machineries. References Holt Linear Regression 26-Apr 29-Apr 2-May 5-May 8-May 11-May 14-May [1] Anastasi G., Farruggia O., Lo Re G., Ortolani M., (2009). Monitoring High-Quality Wine Production using Wireless Sensor Networks, in 42st Hawaii International Conference on Systems Science : Waikoloa, Big Island, HI, USA. [2] Hellman E., (2004). How to judge Grape Ripeness Before Harvest, in 23rd Annual Meeting of New Mexico Grape Growers and Wine Makers. Albuquerque, New Mexico.February 27-28. [3] Policarpo M., Pernice V., Dimino G., Cartabellotta D., (2008). Agroclimatic characterization of Monreale DOC appellation for vine growing, in VII th International terroir Congress, (Nyon, Switzerland). [4] Tonietto J., Carbonneau A.,(2004). A multicriteria climatic classification system for grape-growing regions worldwide, in Agricultural and Forest Meteorology, Volume 124, Issues 1-2, 20 July 2004, Pages 81-97. [5] Watson B., (2003). Evaluation of Winegrape Maturity, in Oregon Viticulture, Editor, E.W. Hellman, Oregon State University, p. 241.