AGROMETEOROLOGICAL-SPECTRAL MODELS WITH MACHINE LEARNING TO FORECAST SWEET ORANGE YIELDS FOR DIFFERENT MATURITY GROUPS

• Grant number: 24/19044-2
• Support Opportunities: Scholarships in Brazil - Master
• Start date: April 01, 2025
• End date: March 31, 2027
• Field of knowledge: Agronomical Sciences - Agronomy - Agricultural Meteorology
• Principal Investigator: Glauco de Souza Rolim
• Grantee: João Antonio Lorençone
• Host Institution: Faculdade de Ciências Agrárias e Veterinárias (FCAV). Universidade Estadual Paulista (UNESP). Campus de Jaboticabal. Jaboticabal , SP, Brazil

Abstract

Sweet orange (Citrus sinensis (L.) Osbeck) is one of the most important agricultural crops globally and one of the main economic and social pillars of the agricultural sector in Brazil, contributing significantly to job creation and exports. The project aims to develop yield prediction models, as early as possible before harvest, for sweet orange for early, mid-season and late ripening groups in irrigated and rainfed systems. The productivity data (tons ha-1) will be collected in partnership with the multinational company Citrosuco for 27 properties distributed in 21 municipalities in the state of São Paulo and one in Minas Gerais, covering a total of 2,121 plots monitored over 12 years (2012-2024). Citrosuco will also make climate data available through automatic weather stations installed on all the properties, recording daily information on precipitation, temperature (maximum, minimum and average), relative humidity, solar radiation, wind speed, incident solar radiation, solar radiation at the top of the atmosphere, dew point temperature, soil moisture for the root system and soil moisture on the surface. In addition, edaphic data such as soil texture and cation exchange capacity will be obtained from the SoilGrids platform, while spectral data will be collected from the SATVeg platform (Embrapa), which provides vegetation indices such as normalized difference vegetation index and enhanced vegetation index. Topographical data, such as altitude and slope, will be collected from the Shuttle Radar Topography Mission (SRTM). Four models will be used for yield predictions: Multiple Linear Regression, Random Forest Regression, Support Vector Regression and XGBoost. The understanding of the influence of climate, soil and vegetation indices on productivity will be evaluated based on Principal Component Analysis and Spearman's Correlation. The data set will be divided between training (70%) and testing (30%), using cross-validation techniques to ensure the robustness and generalizability of the models. The variables will be organized into decennia (10-day periods). General and specific models will be developed for different regions: North, Central, South and Far South, in order to obtain accurate forecasts. The best models will be specialized for the entire citrus belt using the kriging method, using ArcGis Pro. At the end of the project, at least 164 forecasting models will have been trained and tested, with Support Vector Regression and Extreme Gradient Boosting Regression expected to show the best results. The hope is to develop models that can be applied by the sector to predict productivity in different climatic contexts, helping to adapt to climate variability and contributing to the resilience of the citrus sector in the face of uncertainty.