Accurate Weed Mapping and Prescription Map Generation Based on Fully Convolutional Networks Using UAV Imagery

Area of Science:

• Agricultural Science
• Remote Sensing
• Computer Vision

Background:

• Excessive herbicide use in rice cultivation leads to agronomic and environmental issues.
• Site-specific weed management (SSWM) offers a sustainable alternative by tailoring herbicide application.
• Accurate weed cover and prescription maps are crucial for effective SSWM.

Purpose of the Study:

• To develop and evaluate a method for generating precise weed cover and prescription maps using high-resolution UAV imagery.
• To assess the performance of fully convolutional networks (FCN) for pixel-level weed classification in rice fields.
• To determine the potential herbicide savings achievable with the proposed SSWM approach.

Main Methods:

• High-resolution unmanned aerial vehicle (UAV) imagery was captured over a rice field.
• Fully convolutional networks (FCN) were employed for pixel-level weed classification.
• A chessboard segmentation process was utilized to construct the prescription map grid framework.

Main Results:

• The FCN-based weed mapping achieved an overall accuracy of 0.9196 and a mean intersection over union (mean IU) of 0.8473.
• The complete process, from data collection to map generation, took less than 30 minutes for a 50x60m field.
• Prescription map generation with varying weed thresholds resulted in high accuracies (above 0.94) and herbicide savings ranging from 58.3% to 70.8%.

Conclusions:

• The developed method effectively generates accurate weed cover and prescription maps for SSWM applications.
• UAV imagery combined with FCN provides a rapid and precise tool for site-specific weed control in rice.
• This approach demonstrates significant potential for reducing herbicide usage and environmental impact in agriculture.