Weed Density and Distribution Estimation for Precision Agriculture Using Semi-Supervised Learning

Uncontrolled growth of weeds can severely affect the crop yield and quality. Unrestricted use herbicide for weed removal alters biodiversity cause environmental pollution. Instead, identifying weed-infested regions aid selective chemical treatment these regions. Advances in analyzing farm images have resulted solutions to identify plants. However, a majority approaches are based on supervised learning methods which requires huge amount manually annotated images. As result, economically infeasible individual farmer because wide variety plant species being cultivated. In this paper, we propose deep learning-based semi-supervised approach robust estimation density distribution across farmlands using only limited color acquired from autonomous robots. This be useful site-specific management system infected areas work, foreground vegetation pixels containing crops first identified Convolutional Neural Network (CNN) unsupervised segmentation. Subsequently, fine-tuned CNN, eliminating need designing hand-crafted features. The is validated two datasets different crop/weed (1) Crop Weed Field Image Dataset (CWFID), consists carrot (2) Sugar Beets dataset. proposed method able localize maximum recall 0.99 estimate with accuracy 82.13%. Hence, shown generalize without extensive labeled data.