Neural Networks and GPU-Based Weed Detection Under Varying Light Illumination

Destructive plants that spread and compete with agricultural crops for water and nutrients are sometimes referred to as weeds. Earlier, traditional techniques were used in agriculture to remove unwanted plants or weeds. Since weeds are a major cause of crop yield loss, they compete with productive crops for soil, nutrients, and sunshine. As a result, safer and more effective herbicide products are always being developed. Today, uniform herbicide application is a major contributor to crop poisoning, environmental contamination, and high herbicide consumption costs. The issues with uniform spraying in fields can possibly be resolved by site-specific spraying. Hence, reliable weeds detection or identification is done precisely under varying light illumination. Scale invariance feature transform (SIFT) is employed for feature extraction, and a neural network is used to distinguish weeds from plants utilizing geometric features in the existing model. The primary objective is to improve the accuracy rate and to reduce the computational time. The proposed system is closed to be more accurate if convolution neural network and region-based convolutional neural network are used. One of the strengths of convolutional networks is their inherent translation invariance, and by using CNN, depth of the volume is increasing at each level so that can keep the computation time at reasonable level. Labor and pesticide costs will be lowered in a greater proportion, and manual weed removal will need less labor. In the case of smart weed control system, i.e., automatic system, it is expected to improve the accuracy rate. The performance rate can be improved by parallel processing that can be done in GPU will boost performance. The fast RCNN and the CNN classifier are the two most important factors affecting the performance of the proposed system.