APRENDIZADO DE MÁQUINA APLICADO EM IMAGEM NDVI PARA PREVISÃO DA PRODUTIVIDADE DA CANA-DE-AÇÚCAR

Abstract

This article presents an approach through models based on ML (Machine Learning) applied to NDVI (Normalized Difference Vegetation Index) images to estimate productivity in the sugarcane crop. The use of human techniques based on cognitive experiences is predominant to anticipate productivity. The images used were provided by the NDVI Sentinel-2 satellite, since the datasets were obtained from two georeferenced points, two plots and applied to the images for extraction and processing. Two predictive algorithms are used for the models: (i) CNN (Convolution Neural Network), (ii) KNN (K-Nearest Neighbors), (iii) RF (Random Forest), (iv) SVM (Support Vector Machie) , (v) AdaBoost (Adaptive Boost). The RF algorithm was presented or more efficient, so that the results for the DP (Standard Deviation) and the formula for the MSE (Mean Square Error) obtained 30.71 tons (t) and the MAE (Mean Absolute Error) obtained 3.73(t). Regarding the estimates, the DP formula for the MSE obtains 34.71 (t) and the MAE of 3.97 (t). The EM (Mean Error) for the estimates was -8.80% and the RF algorithm was 0.012%. The results will show consistency for the productivity estimates in the sugarcane crop.