Using Semi-Supervised Learning to Predict Weed Density and Distribution for Precision Farming

Abstract

If weed growth is not controlled, it can have a devastating effect on the size and quality of a harvest. Unrestrained pesticide use for weed management can have severe consequences for ecosystem health and contribute to environmental degradation. However, if you can identify problem spots, you can more precisely treat those areas with insecticide. As a result of recent advances in the analysis of farm pictures, techniques have been developed for reliably identifying weed plants. . On the other hand, these methods mostly use supervised learning strategies, which require a huge set of pictures that have been labelled by hand. Therefore, these monitored systems are not practicable for the individual farmer because of the vast variety of plant species being cultivated. In this paper, we propose a semi-supervised deep learning method that uses a small number of colour photos taken by unmanned aerial vehicles to accurately predict the number and location of weeds in farmlands. Knowing the number and location of weeds is helpful for a site-specific weed management system in which only afflicted areas are treated by autonomous robots. In this research, the foreground vegetation pixels (including crops and weeds) are first identified using an unsupervised segmentation method based on a Convolutional Neural Network (CNN). There is then no need for manually constructed features since a trained CNN is used to pinpoint polluted locations. Carrot plants from the (1) Crop Weed Field Image Dataset (CWFID) and sugar beet plants from the (2) Sugar Beets dataset are used to test the approach. The proposed method has a maximum recall of 0.9 and an accuracy of 85%, making it ideal for locating weed hotspots. So, it is shown that the proposed strategy may be used for too many kinds of plants without having to collect a huge quantity of labelled data.