CLASSIFICATION OF THE OCCURRENCE OF BROADLEAF WEEDS IN NARROW-LEAF CROPS

ABSTRACT

Considering the spectral differences between broadleaf weeds and narrow-leaf crops and the influence of terrain and soil variables on weed infestations, integrating such information into a machine-learning algorithm can lead to accurate weed maps. Therefore, we aim to evaluate the effectiveness of these variables in classifying the occurrence of broadleaf weeds in narrow-leaf crops. Weed data was collected at georeferenced points across two areas covering 200 ha (pasture) and 106 ha (sorghum), creating classes 0 (absence) and 1 (presence). For each sample point, we obtained 11 variables: soil clay content, cation exchange capacity, soil organic matter, terrain elevation, slope, NDVI, EVI, CIgreen, BGND (derived from PlanetScope images), and spatial information (X and Y coordinates). These variables were used as predictors of broadleaf weed presence and absence in the Random Forest classification algorithm. The presence and absence of broadleaf weeds were correctly classified in 84% and 74% of all predictions in the test sample sets for pasture and sorghum areas, respectively. This strategy represents an efficient way to map and manage the occurrence of broadleaf weeds in narrow-leaf crops.

precision agriculture; machine learning; remote sensing; weed management