The weed plants of Urtica Dioica (UD) are considered as widely witnessed, injurious and vital weed witnessed in wide areas of grass. It needs rigid techniques of handling (removal or herbicide). These harmful plants have multiple effects on dairy and food production and quality degradation, its removal is a necessary procedure. Although manual techniques of removal are precise, they are extremely costly in the form of time and effort. On the other side, herbicides treatments cause negative impacts on environmental pollution and crops quality. Smart farming and weed processing depend on computer vision techniques to (automatically & efficiently) detect the weeds in order to treat them effectively. Traditional techniques of Machine learning (ML) have significant challenges in weed detection due to the lack of features in real-field imaging. The new fashion of neural networks (Deep learning (DL)) recorded higher accuracy regarding self-learning features, which enhances accurate weed detection. Some recent publications adopted a single DL method, achieving high performance in detecting well-separated weeds and controlled-illumination images, yet they recorded low performance for overlapping uncontrolled illumination or occluded leaves challenges. In this paper, an ensemble of DL networks is used, where each network is robust against specific challenges in UD detection. The proposed ensemble scheme adopts Convolutional Neural Networks (CNN) networks for feature extraction and image classification. The proposed framework includes three major stages such as data preparation, preprocessing, and classification phases. The 3-extractor ensemble scheme is adopted as the cornerstone of the classification stage. The adopted data in this paper is real-field captured images with (overlap, occlusion, and illumination) challenges. The three extractors are: robust Inception-v3 network against illumination issues, robust Visual Graphics Group-16 network (VGG-16) against overlapping issues and robust Residential Energy Services Network-50 (ResNet-50) against occlusion issues. Combining these networks to construct one ensemble scheme enhances the discrimination performance of real-field weed detection. The ensemble- experiments were tested on two benchmarking UD datasets, where both of them contain captured images with real-field challenges. The two datasets consist of (900 &677) samples, respectively, which are separately tested in the ensemble scheme. Set of standard evaluation measures are used, which showed considerable results like accuracy (96.89%), recall (98.08%), precision (94.45%), and F1-score (95.78%). This paper recorded superior performance in their accuracy and significance regarding the State of Art published literature. [ABSTRACT FROM AUTHOR]