Clustering Weather Time Series used for Agricultural Disease Alert Systems in Florida

Abstract

Meteorological observations are widely used as input for disease alert systems in agriculture. In Florida, USA, the AgroClimate Advisory Systems provide disease alerts to growers of various crops, including strawberries, blueberries, and citrus. Data observed in weather stations belonging to FAWN (Florida Automated Weather Network) are used to simulate disease risk, and growers are notified when environmental conditions are favorable for infection, helping them decide when to spray for prevention. However, observation problems in weather stations, such as sensor or communication failures, can compromise the reliability of these applications, which unfortunately are common in this context. Thus, this work explores the clustering of temperature and relative humidity data, in time series format, as a way to monitor the quality of the information provided by two plant disease alert systems. An approach based on clustering was used to group Florida weather stations according to their microclimate characteristics. The elbow and silhouette methods were used to help find the optimal number of clusters, found to be 3. The K-Means algorithm was used with multivariate time series to group the weather stations. Then, an improvement was proposed to flag suspicious observations and early identify inconsistent measurements, increasing the reliability of the system.