Seite - 131 - in Intelligent Environments 2019 - Workshop Proceedings of the 15th International Conference on Intelligent Environments

thepredictionofweather conditions [2,6], yieldprediction [1,13],water saving through irrigationmonitoring [3,7] amongothers. In this studywe focus on the paradigmof precision agriculture facing the problem of predictingweather conditions in particular to avoid frost in crops both in extensive cultivation and in greenhouses. The problem of frost prevention is a problem that de- pends on a large number of factors and that it is not possible to tackle using only the global weather forecast, but it is necessary to have local information on the area/plot forwhich the prediction ofweather conditions is to be carried out [5]. The problemof predicting temperatures to palliate and act on low temperatureswithin agriculture is a latentproblemtoday. If thepossibilityofasignificant temperaturedropisknownbefore- hand, anti-freeze techniques can be activated, such as connectingwindmills and stoves or connecting the heating in a greenhouse. If the anti-frost techniques are not activated sufficiently in advance, great economic losses can be predicted for the farmerwhen all orpart of thecrop is lost. This studyproposesa temperaturepredictionmodelbasedon timeseries, topredict the temperature of a local plot using previously collected temperatures.With this pre- diction, the farmer will be able to know the possibility of a drop in temperatures and thuswill be able to activate and/or prepare all thenecessary resources to apply the anti- frost technique.The idea is to integrate thepredictionmodel intoan IoTsystemtoauto- mate decisions. Thus. this paper focuses on the IoTparadigmand complements one of themost important parts of the architectureof the IoTparadigm, specifically the intelli- gentcomponent.Given thecomplexityof theproblemaddressed, thepredictivemodel is basedondeep learning.DeepLearning represents a set ofmachine learning algorithms basedonasetof artificial neuralnetworkscomposedofcomplexhierarchical levels [4]. Deep learningmodels arebeginning tobeused in theworldofagriculture to solvecom- plexproblemssuchas theclassificationofdiseasesand/orplants throughimagesoryield predictions incrops [10]. In thisstudyatypeofrecurrentneuralnetworkisproposedtocreate the temperature predictionmodel, specificallyLongshort-termmemory(LSTM)neuralnetwork isused. This type of neural networks obtain very satisfactory resultswhen the data have a tem- poral tendency, as is the caseof the temperaturedataof aplot [17].Therefore, themain objectiveof thisworkis toperformapreliminaryanalysisanddesignofanLSTMneural network to create a temperature predictionmodel to be integrated into an IoT system deployed in several agriculturalplots. This study isorganizedas follow. In section2abriefbackground reviewonaspects related to deep learning and precision agriculture is presented. Section 3 describes the data and techniques used for this study de temperature prediction. Finally, Section 4 presents the results and an analysis of themandSection 5presents the conclusions and futurework. 2. Background Deeplearningtechniqueshavebeguntobeintroducedinthefieldofprecisionagriculture tohelpcompleteandrealize thechallenges thatagricultureposes [10].Among thefields of application deep learning has been applied to find the classification of plant species, identification of plant diseases, identification of soil cover, classification of crop type, M.Á.Guillén-Navarroetal. /AnLSTMDeepLearningScheme forPredictionofLowTemperatures 131