Weed density estimation in soya bean crop using deep convolutional neural networks in smart agriculture

Weeds are those unwanted plants that grow between cultivated crops, which reduce the purity of the crops. Crops are severely affected by weeds for their quality and yields. Farmers use the traditional method for weed removal that is time-consuming and also makes it difficult to identify the difference between weed and crop. This research proposes deep convolutional neural network based Inception V4 architecture approach for identifying weed density in soya bean crop fields using crop weed field image dataset (CFWID). This work uses RGB weed and crop images. It offers a data cleaning to eliminate background, and foreground vegetation using segmentation masked. Thereafter, the weed-density area is identified using vegetation segmentation, which is a major challenge in many of such research works. This approach is validated using the CFWID weed and crop dataset that consists of 1100 broadleaf, 2548 grass weed, and the remaining 736 weed images collected from soya bean crop fields and close-to-crop weeds. The proposed model achieves an accuracy of 98.2% using 4384 weed images. Therefore, the proposed approach has been generalized to different weed species in the soya bean crop without the need for extensive labelled data with the precision value of 97%, recall value as 99%, and F1 score as 98%.

Automated summary

This research proposes a deep convolutional neural network approach to identify weed density in soybean crop fields, achieving high accuracy in weed recognition.