Detecting Coffee Leaf Nitrogen with UAV-Based Vegetation Indexes and Machine Learning

This study evaluated the potential of vegetation indexes (VI) based on Unmanned Aerial Vehicle (UAV) and machine learning to estimate nitrogen content (N) in coffee leaves. The study was carried out in a coffee plantation located in Santo Antônio do Amparo, MG, Brazil. In this crop, 10 sampling points were demarcated, consisting of 5 plants. The sampling points were classified as N deficient (<2.5%), N critical (2.5 to 3.0%) and N adequate (3.0 to 3.5%). The images were captured using a commercial UAV 3DR Solo, equipped with a Parrot Sequoia multispectral camera. Chemical analyses of N in coffee leaves and VI values at sampling points were used as input parameters for image training and its classification by Random Forest (RF). The RF performance was evaluated using the metrics global accuracy and kappa coefficient. In general, the classification showed a good performance in evaluating N in the coffee leaves, with global accuracy and kappa coefficient values ranging from 0.64 to 0.78 and 0.52 to 0.73, respectively. Among the evaluated indexes, the Green Normalized Difference Vegetation Index (GNDVI) presented the best results, while the Normalized Difference Red Edge Index (NDRE) was the worst. In addition, it was possible to estimate the spatial distribution of N classes throughout the crop. The GNDVI also presented the best class definitions, demonstrating that 24% of the crop had N deficiency, 49% critical N and 27% sufficient N. The model proposed in this study can offer a promising approach to mapping and quantifying N in a coffee crop.