# Estimation of Kenaf seedling canopy coverage in saline soil using semantic segmentation of UAV RGB images

## TL;DR

This study uses drone images and deep learning to monitor Kenaf seedlings in saline soil, enabling efficient growth analysis and variety selection.

## Contribution

A novel method combining UAV RGB imagery and enhanced U-Net semantic segmentation for precise Kenaf seedling canopy coverage estimation in saline-alkali soils.

## Key findings

- U-Net with a Self-Attention mechanism achieved an IoU of 85.99% and Dice score of 92.44% in plant segmentation.
- Three models (FCN, U-Net, DeepLabV3+) effectively segmented Kenaf plants, each with distinct advantages in accuracy, detail, and efficiency.
- Varieties Xiao 3, K5, and Xiao 2 showed strong saline-alkali adaptability based on canopy coverage and yield analysis.

## Abstract

The growth status of Kenaf (Hibiscus cannabinus L.) seedlings directly impacts its yield and quality. Addressing the challenges of inefficient monitoring and quantitative assessment of Kenaf seedlings under saline-alkali conditions, this study developed an automated method for plant identification and canopy coverage estimation during the seedling stage. This approach leverages high-resolution visible light imagery captured by unmanned aerial vehicles (UAVs) combined with deep learning semantic segmentation techniques. First, a UAV imagery dataset of Kenaf seedlings was constructed through geometric and radiometric calibration, image cropping, and sample annotation. Subsequently, three classical semantic segmentation models—FCN, U-Net, and DeepLabV3+—were trained and compared using image enhancement strategies. Model performance was quantitatively evaluated using metrics including Intersection over Union (IoU), accuracy, precision, and F1 score. Results indicate that all three models effectively segmented Kenaf plants from soil backgrounds. U-Net demonstrated optimal overall accuracy and detail retention, DeepLabV3+ exhibited advantages in small-scale object recognition, while FCN offered high computational efficiency, making it suitable for applications demanding real-time processing. Building upon this, the U-Net architecture was enhanced by incorporating a Self-Attention (SE) channel mechanism, further improving model performance to achieve an IoU of 85.99% and an Dice of 92.44%. Based on segmentation results from the enhanced UNet, plant canopy coverage during the Kenaf seedling stage was calculated. Combined with measured dry bark yield per mu, this enabled analysis of growth performance across varieties under saline-alkali conditions, identifying Xiao 3, K5, and Xiao 2 as materials exhibiting strong saline-alkali adaptability. The study demonstrates that this method enables high-precision identification and quantitative analysis of Kenaf seedlings, providing effective technical support for monitoring seedling growth and variety selection in saline-alkali soils.