EVALUATING PERFORMANCE OF DIFFERENT NEURAL NETWORK MODELS ON CROP AND WEED DETECTION AND CLASSIFICATION

Abstract

One of the most significant elements affecting agricultural yield is weeds. The waste and contamination of rural environments caused by full-coverage herbicide spraying is becoming more visible. With the continual improvement in agricultural production levels, it is critical to differentiate crops from weeds and to achieve accurate weed-only spraying. However, precise weed and crop identification and localization are required for spraying. In order to increase the crop yield and reduce the threats imposed by weeds in agriculture, a measure is taken to identify and classify the weeds and crops with the help of deep learning techniques. Convolutional neural networks render a good way to identify the weeds that harms the crop’s growth. Aiming at achieving a greater accuracy, models such as CNN and MASKR-CNN were built. Comparatively, CNN resulted with 94.29% as training accuracy and 100% accuracy of validation in VGG16 architecture. Therefore, by the suggested method, there is a lot of possibility to reduce the manual work to identify crops and weeds. According to results, with the fine tuning of hyper parameters, accuracy can be increased