Weeds detection efficiency through different convolutional neural networks technology

The preservation of the environment has become a priority and a subject that is receiving more and more attention. This is particularly important in the field of precision agriculture, where pesticide and herbicide use has become more controlled. In this study, we propose to evaluate the ability of the deep learning (DL) and convolutional neural network (CNNs) technology to detect weeds in several types of crops using a perspective and proximity images to enable localized and ultra-localized herbicide spraying in the region of Beni Mellal in Morocco. We studied the detection of weeds through six recent CNN known for their speed and precision, namely, VGGNet (16 and 19), GoogLeNet (Inception V3 and V4) and MobileNet (V1 and V2). The first experiment was performed with the CNNs architectures from scratch and the second experiment with their pre-trained versions. The results showed that Inception V4 achieved the highest precision with a rate of 99.41% and 99.51% on the mixed image sets and for its version from scratch and its pre-trained version respectively, and that MobileNet V2 was the fastest and lightest with its size of 14 MB.