Remotely piloted aircraft in spatial multispectral modeling of stressors in coffee plantantions

Abstract

The use of Remotely Piloted Aircraft for remote sensing in coffee plantations can assist the producer to identify management strategies to be adopted, making the activity more competitive, increasing productivity, and reducing the environmental impact. Therefore, this study aimed to develop a methodology to evaluate and monitor the spatial variability of biotic and abiotic stress in coffee (Coffea arabica L.), using multispectral Remote Sensing data obtained from Remotely Piloted Aircraft. In the first study, 09 vegetation indices were applied to evaluate the damage caused by frost in coffee plants. The results show that the vegetation indices have a strong relation and great precision to the frost damage identified in the coffee plants. In between the indices assessed, the Normalized Difference Vegetation Index (NDVI) has shown the best performances (r = -0.89, R2 = 0.79, MAE = 10.87, and RMSE = 14.35). Additionally, the spatial distribution of the vegetation indices allowed the verification of topography’s influence on the frost occurrence in coffee plantations. The second study evaluated the potential of the Random Forest machine learning method applied to vegetation indices to measure the nitrogen content in coffee leaves. The proposed model presented global accuracy and kappa coefficient of up to 0.91 and 0.86, respectively. The best results were achieved with the indices Green Normalized Difference Vegetation Index (GNDVI) and Green Optimal Soil Adjusted Vegetation Index (GOSAVI). Furthermore, the indices made it possible to verify that only 22% of the entire crop area had symptoms of nitrogen (N) deficiency in the plants, which would result in a 78% reduction in the amount of N applied by the producer. Finally, in the third study, 63 vegetation indices were combined to different machine learning methods, and decision trees to detect the severity of rust disease in coffee plants. The study concluded that the method Logistic Model Tree (LMT) was the one that most contributed to the accurate prediction of the disease. This method achieved overall precision, recall, and f-measure of 0.672, 0.747, and 0.695, respectively. Still, for classes 1 and 4, it returned F values of 0.915 and 0.875, being a good indicator of early Coffee Leaf Rust (CLR) between 2 and 5%, and at later stages of CLR, between 20 and 40 %. It also concluded that this model could help in precision farming practices, as it offers efficient, non-invasive, and spatially continuous monitoring of the disease.