Um modelo de machine-learning para predição do tempo de colheita de árvores macieiras com base em dados fenológicos e parâmetros climáticos

Abstract

Machine learning approaches have been used in several areas. In the field of agricultural research, machine learning has been used to increase agricultural productivity and minimize its environmental impact, proving to be an important tool to support decision making.Different strategies are found in the literature to predict phenological stages of different cultures. From the current state of the art, we observed few works that address the prediction of the harvest date. We did not find works with an approach similar to the one proposed. Forecasting the time of harvest is a challenge to develop fruit production sustainably and reduce food waste. Fruits are perishable, of high value and seasonal, and sales prices are generally time sensitive, which makes harvest forecasts extremely valuable to growers. This study proposes the Pred- Harv model, a machine learning model that uses recurrent neural networks to predict the start date of the apple harvest, given the temperature-related weather conditions expected for the period. Predictions are made from the phenological stage of full bloom, based on historical series of phenology and meteorological data. The computational model contributes with the ability to anticipate information about the harvest date, enabling the producer to better plan activities, avoiding costs and improving productivity. The use of ML methods aims to make the predictive capacity of models based on thermal summation aimed at fruit growing more effective, allowing for the simulation of climate changes in the period. The PredHarv model is based on thermal sum models, but uses a multivariate approach. We use the thermal sum relating it to period length and other variables related to period temperature. We use a machine learning method, exploring the potential of LSTM networks to deal with problems involving time series. The model output returns the period length in calendar days, given the expected temperaturerelated weather conditions for the period. Additionally, a methodology for using the model is proposed in order to expand the predictive capacity, as a way to reduce the uncertainty implicit in the information provided by the user, necessary for calculating the forecast. We developed a prototype of the PredHarv model and performed experiments with real data from agricultural institutions. The combination of variables used in the model demonstrated an effective prediction strategy. We evaluated the metrics and the results obtained in the evaluation scenarios demonstrate that the model is efficient, with good generalization and capable of obtaining results with better accuracy compared to the linear model based on thermal accumulation.