Microclimate Hourly Downscaling Using Wireless Sensor and Online Sequential Extreme Learning Machines

In agriculture weather forecast can be considered a valuable asset for planning and also when giving guidance for decision support systems. However, if the forecast is based on a meteorological station which is not located close to the field it could fail to reproduce important microclimate factors. In this paper, we used a non-linear algorithm called online sequential extreme learning machine (OSELM) to perform a point statistical downscaling from 1 to 6 hour ahead of the Global Forecast System (GFS) data to a wireless sensor in the field which collects weather data. Online sequential learning algorithms do not require expensive retraining whenever new data are received, thus whenever new data are received the algorithms adjust the model parameters, maximizing the accuracy of the next forecast. We compared the OSELM against the online sequential multiple linear regression (OSMLR) and the GFS forecasts. Using local observations recorded from 10 stations on different locations over US, we concluded that OSELM is an attractive approach as it consistently outperformed OSMLR and GFS when comparing the RMSE.