Using Geospatial and Computer Vision Techniques for Weed Detection and Mapping in Agricultural Systems

Abstract

Site-specific treatment of weeds in agricultural landscapes has been gaining importance in recent years due to economic savings and minimal impact on the environment. Weed detection and recognition is a major component of any site-specific treatment method. Different crop-weed complexities and weed control objectives may require the strategic implementation of various weed recognition approaches. This research aims at evaluating various weed detection and mapping approaches in different crops including cotton, corn, soybean, and wheat using remotely sensed digital RGB imageries. The first experiment was conducted in mid-season cotton infested with early-mid growth stage weeds to evaluate crop row detection methods for weed mapping and density estimation. The second experiment was conducted in wheat to evaluate the pixel-based detection approach for detecting Italian ryegrass and developing grid maps for competitive interactions. The third experiment was conducted to test the cross-crop species applicability of a convolutional neural networks (CNN)-based weed detection model trained for cotton over other row crops such as corn and soybean. The fourth experiment was conducted to explore various image synthesis techniques for training deep learning models to detect weeds in cotton. The first experiment revealed that the crop row detection approach can provide high accuracy levels for weed mapping and weed-density estimation. The second experiment showed that grass weeds such as Italian ryegrass can be effectively classified from wheat and competitive effects of ryegrass on wheat can be predicted early with reasonably high accuracy using the pixel-based machine learning approach. The third experiment indicated that a deep learning-based weed detection model trained for cotton can be used for soybean with more confidence compared to corn. The final experiment revealed that synthetic images can provide comparable accuracy to real images for training weed detection models. In addition, the experiment showed that above-ground biomass of broadleaved weeds may be better predicted than grass using canopy mask results. Overall, these findings improve sensor-based weed detection, which is expected to advance precision weed management.