Designing deep learning neural network for auto-detection of growing weeds within grass lands.

Accurate herbicide spraying offers a significant reduction in herbicide consumption, which reduces negative effects on other plants. An adopted system of detection is a crucial component of the smart procedure of spraying, which enhances the detection of target weeds and the decisions of spraying. This paper reports a set of Deep Convolutional Neural Network (DCNN) methods that recorded high accuracy in inspecting Bermuda-grass of type [Cynodon dactylon (L.) Pers.] for growing weeds. Visual Geometry Group (VGG) Net recorded a higher than (0.95) F1 score value, which outperforms ResNet-101 in detecting dollar weed of types (Verticillata plant.), Florida pusley (Richardia scabra L.), and old-world diamond-flower (Hedyotis corymbose L. Lam.) in growing grassland areas. A single model of RCNNet provided reliable detection of weeds within areas of grassland over different surface conditions and heights of mowing. Best recorded results of DCNN were yielded by DetectNet architecture in detecting Poa trivialis, which grow with different types of broadleaf weeds in the dormant stage of grassland. DetectNet provided considerable performance in weed detection during the growing stage of dormant grassland, where F1 scores exceeded (0.99). Due to the significant performance level, DCNN-based weed provided an efficient decision system, where machine-vision subsystems are adopted in precision herbicide applications to control weed growing within grassland areas. [ABSTRACT FROM AUTHOR]