# Assessment of three broadleaf weed species classification in rice field using UAV hyperspectral imaging and machine learning

## TL;DR

This study uses drone-based hyperspectral imaging and machine learning to accurately detect and classify three types of weeds in rice fields, helping improve crop management.

## Contribution

The study introduces an effective UAV hyperspectral imaging and SVM-based machine learning method for early-stage weed classification in rice fields.

## Key findings

- SVM achieved over 99% classification accuracy for three weed species at all growth stages.
- Weed vegetation cover increased over time, while rice cover fluctuated and soil cover decreased.
- The method supports scalable and efficient weed detection for precision agriculture.

## Abstract

Broadleaf weed (BLW) infestation is a major challenge in rice cultivation, particularly during the early vegetative stages when competition for resources is most critical. This study aims to enhance early-stage detection and classification of three prevalent BLW species—Monochoria vaginalis (MV), Limnocharis flava (LF), and Sphenoclea zeylanica (SZ)—in rice fields using unmanned aerial vehicle (UAV)-based hyperspectral imaging integrated with machine learning techniques. The research was conducted in a 1-hectare rice plot (Block L5A, Plot 121) near Pusat Benih Padi Felcra Sdn Bhd, Perak, Malaysia, a site characterized by high weed density. Hyperspectral data were acquired using a DJI Matrice 600 UAV equipped with a Resonon Pika L hyperspectral camera flown at 40 meters altitude. ENVI Classic 5.3 software was used to perform supervised classification based on selected regions of interest (ROIs) for training. Three classification algorithms—Support Vector Machine (SVM), Minimum Distance (MD), and Parallelepiped (PP)—were compared at 15, 25, and 30 days after sowing (DAS). Among them, SVM consistently achieved the highest classification accuracy, exceeding 99% for all weed species across all growth stages, with minimal omission and commission errors. Vegetation cover analysis showed an increasing trend in BLW expansion over time, while rice cover fluctuated and soil cover declined, indicating the competitive dominance of weeds. The findings underscore the effectiveness of UAV hyperspectral imaging combined with machine learning—especially SVM—as a scalable, accurate, and efficient approach for early weed detection. This methodology can support precision agriculture by enabling timely and targeted weed management strategies, ultimately improving rice yield and sustainability.