Identifying and quantifying weeds is a crucial aspect of agriculture for efficiently controlling them. Weeds compete with the crop for nutrients, minerals, physical space, sunlight, and water, causing problems in crops ranging from low production to economic losses and environmental deterioration of the land. Weed quantification is generally a manual process requiring significant time and precision. Convolutional Neural Networks (CNN) are very common in weed quantification. Thus, the purpose of this research is the adaptation of the ResNeXt50 CNN architecture for semantic segmentation tasks, focused on the automatic quantification of weeds (Broadleaf dock, Dandelion, Kikuyo grass, and other unidentified classes) in potato fields using RGB images acquired by the DJI Mavic 2 Pro drone. The analytical model was trained following the Knowledge Discovery in Databases (KDD) methodology using Python and the TensorFlow-Keras frameworks. The results indicate that the modified ResNeXt50 model presented a mean IoU of 0.7350, a performance comparable to the values reported by other authors considering fewer weed classes. The Student´s t-test and Pearson correlation coefficient were applied to contrast the weed coverage from the model predictions and the ground truth, indicating no statistically significant differences between both measurements in most weed classes.