Frost Forecast and Early Alarm System based on Deep Learning

Frost events represent a risk to food security around the world. In Bolivian highlands, this risk is aggravated because of the geographic condition and it may provoke negative economical repercussions. Considering that the population in those regions mainly depends on agricultural activities to survive, any hazard to crops also implies several risks to their subsistence. Traditionally, preventive tasks in these regions are manually performed and consist of dropping water nearby plants’ locations, however, they must be correctly synchronized with the time of the frost event. Unfortunately, due to the lack of local weather systems, farmers cannot forecast precisely when these events will take place which results in having higher chances of losing crops. In this work, we propose a system to forecast frost events and notify such through SMS alarms. As a proof of concept, a weather station was installed in the crop area of the Unidad Académica Campesina in the Batallas municipality to collect measurements of the ambient temperature and humidity. The data is used as input for the developed forecast model which is a deep learning model that combines both the discrete wavelets transform and a Wave Net block. The model’s target is to generate a forecast of ambient temperature for the next 24 hours with hourly resolution. The obtained results show a range between 99.64% and 91.26% of efficacy at the moment to forecast frost events. Moreover, the temperature prediction error (using the RMSE test) ranges between 0.15 and 1.37 Celsius degrees, also, the MAPE test shows a ratio between 0.04% to 0.34%, along the forecast horizon.