The experiment was conducted in Yancheng City, Jiangsu Province, China (32°34′N–34°28′N, 119°27′E–120°54′E). Located in the core region of the Yangtze River Delta, Yancheng features flat terrain and a temperate climate, transitioning from the northern subtropical to warm temperate zones. Bordering the East China Sea, the region encompasses approximately 7 million mu (473, 333 hectares) of tidal flats with widespread saline-alkali soils. The experimental area exhibits elevated soil salinity levels, imposing stress on crop germination and early growth (Figure 1).

Soil pH ranged from 8.88 to 9.07, with total water-soluble salts (TSW) content between 0.12% and 2.86%. Kenaf was sown using broadcast seeding, with seeds evenly distributed across each experimental plot. Each plot measured 40 m in length and 1.4 m in width. Eleven test varieties were included: H368, Xiao 1, Xiao 2, Xiao 3, K1, K2, K3, K4, K5, K6, and K7.

Visible light imagery was captured using a quadcopter UAV (DJI Inspire 2, DJI Innovation, Guangdong, China) equipped with an RGB camera featuring a resolution of 5472 × 3648 pixels. Multispectral imagery acquisition employed a quadcopter (DJI Phantom 4 Multispectral, DJI Innovation, Guangdong, China) equipped with one visible light sensor and five multispectral sensors. Flight operations occurred from 10:00–11:00 on July 9, 2023, with aerial photography parameters set as follows: flight altitude 25 m, lens tilt angle −90°, flight speed 6 m/s, and both fore-and-aft and sideways overlap rates set to 75%. To minimize image distortion and target movement, data collection was conducted under clear, windless weather conditions. Ultimately, 335 high-resolution visible images and 1, 734 multispectral images were acquired during the Kenaf seedling stage. Image calibration and stitching were completed using DJI Terra software.

Prior to harvesting, ground growth data were collected at the Kenaf maturity stage. Fresh plant weight was determined by randomly selecting 20 Kenaf plants and summing their fresh weights; plant height and stem diameter were each taken as the average value from the same batch of 20 plants; bark thickness measurements were obtained by stacking the bark layers from five selected plants to ensure more stable and representative data.

All image processing and deep learning experiments were conducted on the same high-performance workstation. The hardware configuration includes a 13th-generation Intel Core i5-13490F processor (base frequency 2.5 GHz), 64 GB DDR4 memory, and an NVIDIA GeForce RTX 4070 graphics card. The software environment utilized Anaconda-Miniconda3 to build isolated virtual environments running Python 3.8, CUDA 11.3, and PyTorch 1.10. Development and debugging were performed using PyCharm Community Edition 2024. Remote sensing feature extraction was conducted with phenoAI air (AgriBrain, Nanjing, China). This hardware and software configuration provided stable and efficient computational support for model training and testing.

Orthoimages were cropped based on neighborhood boundaries to uniform widths and lengths, yielding sub-images of 340 × 512 pixels (Image A) and 340 × 8880 pixels (Image B: single-image representation of a single Kenaf variety). From image A, 600 images were randomly selected and manually annotated using LabelMe software to generate binary semantic segmentation labels for Kenaf seedlings/background. To enhance model robustness and generalization, multiple data augmentation strategies were employed during training, including:

Scale change: Images were scaled at six levels (0.5×, 0.75×, 1.25×, 1.5×, 1.75×, 2×) and uniformly cropped to 512×512 pixels;

Geometric flip: Including horizontal flipping and vertical flipping by 180°;

Photometric distortion: Random adjustments to brightness, contrast, and saturation.

To ensure image quality and consistency, all images underwent mean and standard deviation normalization during dataset construction to minimize color bias caused by lighting variations. Background pixels were filled with 0, while segmented label regions were filled with 255 to ensure label consistency. Finally, the dataset was divided into training, validation, and test sets at an 8:1:1 ratio, providing ample data support for deep learning model construction and evaluation.

The objective of this study is to establish a standardised, reproducible workflow for monitoring and segmenting Kenaf seedlings using UAV. A systematic comparison will be made of the performance of representative classical semantic segmentation models on this task. In selecting architectures for this study, three were selected based on the following principles.

Based on the optimal segmentation model trained, semantic segmentation was performed on the Image B dataset. The model automatically identified and distinguished the Kenaf canopy coverage from the background areas in the images, generating a binary segmentation result map where plant area pixels are valued as 1 and background pixels as 0.

Subsequently, using NumPy—a Python module for image analysis—the number of pixels within the plant canopy coverage was counted, and its proportion relative to the entire image was calculated. This proportion reflects the coverage of Kenaf in the field image, i.e., the percentage of plant area. The calculation formula is as follows:

To further investigate the relationship between Kenaf seedling growth information and remote sensing characteristics, this study employed PhenoAI Air software to systematically analyze and extract features from remote sensing data across all experimental plots. This software can remove soil background from images. For each Kenaf variety area, through information annotation, it automatically calculates and outputs 138 remote sensing feature values across three major categories: color features, texture features, and spectral indices (Fu et al., 2023).

Color features primarily reflect the color composition and brightness variations of plant canopies, indirectly indicating crop health status. Texture features describe the spatial distribution patterns of grayscale values in images, aiding in distinguishing plants from soil backgrounds or weeds. Spectral indices quantitatively reflect vegetation coverage, photosynthetic activity, and biomass changes, serving as crucial indicators for vegetation growth monitoring.

Following feature extraction, multiple feature selection methods, including Pearson correlation analysis, Lasso, Competitive Adaptive Reweighted Sampling (CARS), and Least Angle Regression (LARS), were conducted between various remote sensing feature values and the plant canopy coverage of Kenaf seedlings obtained through optimal model segmentation. By comparing the key features identified through different methods, determine the optimal set of remote sensing feature values. The complete data collection and processing, model construction, and analysis workflow diagram is shown in Figure 3.

To visually demonstrate the performance of different models in segmenting Kenaf seedlings, the segmentation results were visualized. Plant regions are represented in green, while soil background appears in blue. The overlay layer’s opacity was set to 50% to enable simultaneous observation of the original image and segmentation results.

The segmentation visualizations reveal that all three models accurately distinguish Kenaf plants from the soil background, though differences exist in detail handling and boundary recognition. The FCN model achieves overall coherent segmentation with fewer missed detections, yet some soil background misclassification persists within plant regions—particularly in areas with heavy plant shadows. Additionally, when plants densely cluster in small patches, FCN struggles to effectively separate adjacent individuals, leading to minor plant adhesion in small areas.

The DeepLabv3+ model excels at identifying small-scale individual plants, effectively preserving the morphological features of independent plants. However, it underperforms in recognizing some plants with lighter leaf colors, resulting in a few missed detections (Figure 4).

In contrast, the UNet model delivers the best overall segmentation accuracy and regional coherence. Within small areas, UNet preserves plant details more effectively. Although minor missed detections still occur in isolated cases, its overall visual consistency and recognition accuracy surpass those of the other two models.

The segmentation results of three models—DeepLabv3+, FCN, and U-Net—were compared (Table 3) to evaluate their performance in identifying Kenaf seedlings during the seedling stage. Overall, all three models effectively distinguished Kenaf from soil background areas, though differences existed in accuracy and computational efficiency.

Among segmentation accuracy metrics, the U-Net model demonstrated the best performance, achieving an Intersection over Union (IoU) of 84.29%, an average precision of 91.91%, and an F1 score of 91.45%. This indicates that UNet effectively fuses shallow spatial information with deep semantic features through its skip-connection architecture, better preserving Kenaf leaf edge details and texture information, thereby enhancing overall segmentation quality.

The DeepLabv3+ model achieved an average IoU of 82.52%, slightly lower than U-Net, but its precision reached 91.17%, indicating fewer false positives in plant pixel classification. This high precision relates to the ASPP module’s multiscale feature extraction capability, which effectively captures the complex morphology of Kenaf plants. However, DeepLabv3+’s model execution time is approximately 2.5 times longer than the other two models, resulting in relatively low inference efficiency that is less suitable for large-scale rapid processing.

The FCN model achieved an average Intersection over Union (IoU) of 80.54%, performing slightly below the other two models overall. Nevertheless, it maintained a high accuracy rate and stable recall rate. Its simpler model architecture enables the fastest inference speed, making it suitable for applications requiring real-time processing or with limited computational resources.

A comprehensive comparison of the three models reveals that U-Net strikes a favorable balance between segmentation accuracy and computational efficiency. Notably, it achieves the highest recall rate of 92.33% in identifying Kenaf areas, demonstrating its most stable and reliable performance in the task of extracting Kenaf plants.

To enhance and optimise the model architecture, we compared UNet, SE-UNet, and SE-UNet with AdamW (Table 4), the SE module consistently achieved improvements across all metrics. mIoU increased by 1.03%, while mF1-score rose by 0.61%. Replacing SGD with AdamW further enhanced SE-UNet’s performance, achieving an overall mIoU improvement of 1.7%, accuracy gain of 1.01%, and mF1-score increase of 0.99%.

To quantitatively assess the individual contribution of different data augmentation strategies, a series of ablation experiments were conducted using the SE-UNet model optimized with AdamW. In each experiment, only one augmentation strategy was applied while all other augmentation operations were disabled. The segmentation performance under different augmentation configurations is summarized in Table 5.

When scale change was applied, SE-UNet achieved an IoU of 85.08% and an F1-score of 91.91%, indicating that scale-based augmentation can moderately improve the model’s robustness to variations in seedling size and spatial resolution. However, its contribution to overall segmentation accuracy was relatively limited compared with other augmentation strategies.The use of geometric inversion with IoU and F1-score increasing to 85.34% and 92.07%, respectively. Among the tested strategies, photometric distortion resulted in the highest segmentation performance. Under this configuration, SE-UNet achieved an IoU of 85.52%, an accuracy of 92.93%, and an F1-score of 92.17%. These results demonstrate that photometric augmentation is particularly effective in improving model robustness to illumination variability.

In the context of saline-alkali conditions, a notable degree of variability in growth patterns was observed among 11 Kenaf cultivars. The dry bark yield per mu fluctuated between 241.76–538.65 kg/mu, with Xiao 3 and K5 yielding the highest, while K3 yielded the lowest.

During the seedling stage, the plant area exhibited marked inter-varietal variation, ranging from 0.2202 to 0.4743. Xiao 1 and Xiao 2 exhibited larger plant areas and vigorous early growth, whereas K2 and K3 displayed smaller plant areas, indicating greater susceptibility to salt stress inhibition during the seedling stage (Figure 5).

In comparison with conventional morphological indicators, the correlation between seedling-stage plant area and yield demonstrated enhanced stability. The sprout area exhibited a moderately strong positive correlation with dry bark yield (r ≈ 0.52), indicating that early canopy expansion plays a crucial role in subsequent biomass formation (Figure 6).

To further investigate the relationship between remote sensing characteristics and the growth status of Kenaf seedlings during the seedling stage, a comprehensive feature analysis framework was employed, integrating correlation analysis and multi-method feature selection. The three types of remote sensing characteristics used in this analysis were detailed in our previous study (Fu et al., 2024). The plant canopy coverage derived from optimal semantic segmentation was used as the target variable for subsequent analysis.

Initially, Pearson correlation Analysis results indicate that color characteristics and vegetation index features generally exhibit high correlations with plant area, demonstrating that spectral and color features extracted from imagery can effectively reflect differences in plant canopy coverage and biomass. Among these, 18 feature indicators showed extremely high correlations with plant area (correlation coefficient |r| > 0.8) (Figure 7).

These indicators primarily originate from combinations of reflectance values in the visible and near-infrared bands. Vegetation indices such as NDRE, GNDVI, and GRNDVI effectively reflect chlorophyll content and canopy coverage in Kenaf leaves, while color features like ExG, ExGR, and RGRI demonstrate enhanced green characteristics in the visible spectrum, showing significant correlation with plant area (Table 6).

To more effectively reduce feature redundancy and enhance the robustness of feature selection, this study further integrated three complementary feature selection methods: LASSO, LARS, and CARS. By comprehensively comparing the features selected by different methods, we extracted common key variables, thereby revealing important feature factors capable of stably and consistently predicting plant canopy coverage.

Comparative results demonstrate that features selected by LARS highly overlap with correlation analysis outcomes, including: RGRI, H, WI_mean, B, a, ExR, WI_std, and G, indicating these variables exhibit stable strong linear correlations. Similarly, features chosen by CARS—COM, ExR, ExGR, B, a, and NDRE—also share substantial overlap with correlation analysis. Moreover, both LARS and CARS jointly selected the four features ExR, B, S, and a, demonstrating their significant robustness across two distinct feature selection methods. Comprehensive analysis indicates that correlation analysis, LARS, and CARS consistently identified three core features: B, a, and ExR. In contrast, the feature subset selected by LASSO was more streamlined, comprising only ExR, B, and H.

This study systematically compares three representative semantic segmentation models—FCN, DeepLabv3+, and U-Net—regarding their efficacy in extracting ramie seedlings from high-resolution aerial imagery. Whilst all models successfully separate seedlings from soil backgrounds, performance variations stem from architectural characteristics and feature representation mechanisms.

As a benchmark end-to-end segmentation framework, FCN demonstrates stable performance and high computational efficiency. However, its limited feature fusion capability constrains its handling of fine boundaries and densely distributed seedlings. DeepLabv3+, leveraging dilated convolutions and multi-scale contextual modelling, excels in small object recognition accuracy. Yet, its high computational complexity prolongs inference time, limiting its application in large-scale or real-time processing.

U-Net achieves the optimal balance between segmentation accuracy and efficiency. Its encoder-decoder architecture with skip connections effectively preserves spatial details while integrating high-level semantic information, making it particularly well-suited for ramie seedling segmentation tasks—a scenario characterised by fragmented plant morphology, significant size variation, and strong background interference. These results confirm U-Net as a reliable benchmark solution for UAV-based crop seedling segmentation tasks.

To further enhance segmentation performance, we introduced a squeeze-excited (SE) channel attention module within the UNet architecture. This mechanism enables the network to amplify information-rich spectral and textural channels while suppressing background noise through adaptive calibration of channel-level feature responses. Improvements in intersection-over-union (IoU) and F1 scores demonstrate that channel attention effectively enhances the ability to distinguish small-scale features.

Furthermore, replacing the SGD optimiser with AdamW further elevates model performance. AdamW decouples weight decay from gradient updates, thereby achieving more stable convergence and stronger generalisation capabilities.

Data augmentation ablation experiments demonstrate that different augmentation strategies contribute disparately to segmentation performance. Among the tested strategies, photometric distortion yields the most pronounced improvement, followed by geometric inversion and scale variation. This phenomenon indicates that variability introduced by changing illumination conditions—such as solar angle and shading—constitutes a key source of uncertainty affecting segmentation outcomes in drone-captured images of red hemp seedlings.

Under saline-alkali stress, significant variations were observed among different Kenaf varieties in terms of seedling canopy development, stem structure, and yield factors. Varieties exhibiting greater canopy coverage typically demonstrated stronger early growth vigour and higher final dry bark yields, indicating superior early salt tolerance. The contribution of structural traits to yield varied by cultivar: Xiao3 achieved the highest dry bark yield despite possessing medium stem thickness and relatively thin bark; K5 attained high dry bark yield despite lower fresh bark weight, reflecting its outstanding dehydration and biomass conversion efficiency under salt stress. In summary, under the experimental conditions, Xiao3, K5 and Xiao2 demonstrated strong salt tolerance, whereas K3 and K6 exhibited greater sensitivity to salt stress.

Building upon semantic segmentation, this study systematically identified the four key radiometric features most strongly correlated with ramie seedling canopy coverage through correlation analysis and multiple feature selection (LASSO, LARS, CARS): B, a, H, and ExR. The robustness demonstrated by these features across different methodologies supports their potential as reliable phenotypic indicators. Future work may integrate precise segmentation results with the selected features to achieve high-precision, non-destructive monitoring of crop growth in saline-alkali environments.

Although this study systematically evaluated three mainstream semantic segmentation models for Kenaf seedling identification, several limitations should be acknowledged. First, the dataset primarily consisted of UAV imagery from a single growth stage, lacking multi-temporal data needed to comprehensively capture dynamic growth patterns. Second, the limited number of training samples and influences such as varying illumination and shadow conditions may have constrained model generalizability. Future research could advance this work in the following directions: