Do classificador ao estimador: adaptando transformers para predição contínua de produtividade agrícola para soja a partir de séries temporais de imagens de satélites

Abstract

Agriculture is an important sector for the global economy and food security; however, it faces challenges related to environmental variability and the scarcity of accurate data for analysis and decision-making. Forecasting agricultural productivity, especially for crops such as soybeans, is crucial for resource optimization and agricultural planning. In recent years, remote sensing through satellite imagery has provided frequent and detailed data, helping to understand the dynamics of vegetation cover, plants and land use patterns. When combined with deep learning techniques, it has become a powerful tool for monitoring and analyzing agricultural areas. This work proposes the adaptation of a model originally designed for crop classification, the SITS-FORMER, to a regression task, validated through soybean productivity estimation in small agricultural plots. To this end, a specific dataset was developed, combining time series from the Sentinel-2 satellite with historical soybean productivity data from selected Brazilian municipalities. This was done in light of the limited availability of reliable agricultural data and the need to validate crop presence across multiple years. The construction of this dataset was relevant to ensure the quality and representativeness of the information used for training and validating the model. By transforming the SITSFORMER from a classification model into a regression model, it became possible to leverage its ability to learn temporal, spatial and spectral representations from image series, enabling continuous productivity prediction instead of simple categorization. This approach allows for yield estimation with practical applicability, meeting the demand for quick and accessible information in precision agriculture. The experiments conducted demonstrated that the adapted model is promising and performed better when using a complete temporal window aligned with the crop calendar, reaching a coefficient of determination (R²) of 0.067. The use of the MSE loss function further enhanced the results under this temporal configuration, yielding a value of 0.10. However, the model showed limitations in capturing extreme productivity values, both high and low, indicating low sensitivity to outliers.