Scalable coffee crop classification using NDVI time series from SATVeg and machine learning models

Coffee production is one of Brazil's most important agricultural activities, making monitoring essential for sustainability and production management. This study introduces a method for identifying coffee plantations using publicly available NDVI (Normalized Difference Vegetation Index) time series data from the SATVeg platform. We analyzed sixty georeferenced samples from Patrocínio, Minas Gerais, over six years with six classification models: Logistic Regression, K-Nearest Neighbors Dynamic Time Warping (KNN-DTW), Support Vector Machine (SVM), Random Forest (RF), XGBoost, and TiDE. This is the first published study to report 100% classification accuracy using solely NDVI time series from the SATVeg platform, without incorporating high-resolution or multispectral imagery. TiDE outperformed the others, achieving 100% accuracy in real-world validation, a mean probability of 0.996, and a standard deviation of ±0.001, indicating high prediction stability. SVM, RF, and Logistic Regression also reached 100% accuracy, while XGBoost showed strong recall but less robustness, and KNN-DTW performed poorly. This study demonstrates that NDVI can accurately classify crops using moderate-resolution data, reducing the cost of model processing and simplifying scalable solutions for agriculture. • Sixty georeferenced NDVI time-series samples were extracted from SATVeg for coffee and non-coffee crops in Patrocínio, Brazil. • NDVI temporal patterns revealed seasonal and biennial cycles supporting crop classification without high-res imagery. • Six models were evaluated: Logistic Regression, KNN-DTW, SVM, Random Forest, XGBoost, and TiDE. • TiDE and Random Forest reached 100% accuracy on real-world validation, with TiDE providing the highest stability. • NDVI alone enabled accurate classification, confirming the viability of low-cost, scalable monitoring of crops.