Exploring the Impact of Species Participation Levels on the Performance of Dominant Plant Identification Models in the Sericite–Artemisia Desert Grassland by Using Deep Learning

Abstract

Accurate plant species identification in desert grasslands using hyperspectral data is a critical prerequisite for large-scale, high-precision grassland monitoring and management. However, due to prolonged overgrazing and the inherent ecological vulnerability of the environment, sericite–Artemisia desert grassland has experienced significant ecological degradation. Therefore, in this study, we obtained spectral images of the grassland in April 2022 using a Soc710 VP imaging spectrometer (Surface Optics Corporation, San Diego, CA, USA), which were classified into three levels (low, medium, and high) based on the level of participation of Seriphidium transiliense (Poljakov) Poljakov and Ceratocarpus arenarius L. in the community. The optimal index factor (OIF) was employed to synthesize feature band images, which were subsequently used as input for the DeepLabv3p, PSPNet, and UNet deep learning models in order to assess the influence of species participation on classification accuracy. The results indicated that species participation significantly impacted spectral information extraction and model classification performance. Higher participation enhanced the scattering of reflectivity in the canopy structure of S. transiliense, while the light saturation effect of C. arenarius was induced by its short stature. Band combinations—such as Blue, Red Edge, and NIR (BREN) and Red, Red Edge, and NIR (RREN)—exhibited strong capabilities in capturing structural vegetation information. The identification model performances were optimal, with a high level of S. transiliense participation and with DeepLabv3p, PSPNet, and UNet achieving an overall accuracy (OA) of 97.86%, 96.51%, and 98.20%. Among the tested models, UNet exhibited the highest classification accuracy and robustness with small sample datasets, effectively differentiating between S. transiliense, C. arenarius, and bare ground. However, when C. arenarius was the primary target species, the model’s performance declined as its participation levels increased, exhibiting significant omission errors for S. transiliense, whose producer’s accuracy (PA) decreased by 45.91%. The findings of this study provide effective technical means and theoretical support for the identification of plant species and ecological monitoring in sericite–Artemisia desert grasslands.

1. Introduction

Hyperspectral remote sensing technology, characterized by a high spectral resolution, a high multi-band capacity, and rich informational content, demonstrates significant advantages in identifying, classifying, and extracting surface material data [1]. Recent advancements in pattern recognition and machine learning have further enabled substantial progress in plant identification via remote sensing, offering a novel technical approach for species recognition that achieves rapid and high-accuracy results. Traditional machine learning models require manual feature selection and extraction, thereby constraining the model’s learning capacity and increasing reliance on human intervention, particularly when processing large-scale plant species datasets [2]. In contrast, deep learning networks autonomously extract image features, exhibiting superior learning capabilities, training efficiency, and classification accuracy—often exceeding 90% [3]. For example, Olariu et al. [4] employed maximum likelihood, support vector machine (SVM), random forest (RF), and VGG-19 convolutional neural network (CNN) algorithms to identify invasive woody plants, with VGG-19 CNNs achieving the highest accuracy (96.9%). Similarly, Pan et al. [5] utilized GoogleNet, ResNet-50, ResNet-101, and VGG-16 networks to classify 23 wild Vitis vinifera L. leaf varieties, reporting accuracies of over 95% across all the models. Salehi et al. [6] further demonstrated the efficacy of deep learning by attaining a 100% classification accuracy for Triticum aestivum L. genotypes using feature-based artificial neural networks (ANNs), outperforming PCA-SIMCA methods. Additionally, Sawarkar et al. [7], with the help of the majority multiclass voting (MajMulVot) algorithm, used supervised machine learning (ML) and deep learning (DL) to classify commercial bamboo species; the result shows that the DL-MajMulVot method is better at classifying bamboo than the ML method and had the highest accuracy at 92.8%.

Species composition is a fundamental parameter for characterizing vegetation’s community structures and serves as a key indicator of regional biodiversity and ecosystem health [8]. Accurately classifying grassland plant species and mapping their spatial distribution are critical for advancing ecological understanding, monitoring biodiversity, and implementing targeted management strategies [9,10]. For example, Lyu et al. [11] employed hyperspectral remote sensing imagery integrated with a multi-layer perceptron neural network (MLPNN) to distinguish dominant species and degradation indicator plants through a spectral analysis. Their results demonstrated that the spectral curve-based classification of typical grassland species achieved a 20% higher accuracy compared to mixed vegetation spectral identification. Ainiwaer et al. [3] used VGG16 and VGG19 models to classify plant communities in the oasis desert transition zone; the highest accuracies of the models were 95.25% and 96.73%. Similarly, Liu et al. [12] developed classification models—including RF, a SVM, a backpropagation (BP) neural network, a CNN, and an enhanced CNN—using hyperspectral data to identify key and degraded species in temperate grasslands. Their findings revealed that the enhanced CNN model exhibited a superior performance in species discrimination, achieving stable accuracy rates exceeding 98% for Artemisia frigida Willd. and Caragana microphylla Lam. Wang et al. [13] constructed a streamlined 2D-CNN model through feature selection to classify the main plants A. frigida and Stipa brevifora Griseb. in desert grassland; its overall classification accuracy was 99.216%.

Seriphidium transiliense (Poljakov) Poljakov desert grassland, a typical sericite–Artemisia desert grassland in Xinjiang [14], has undergone extensive degradation due to prolonged overgrazing and inherent ecological vulnerability [15]. To improve the precision and efficiency of grassland monitoring, our research team developed a Fisher discriminant model, integrating feature parameters and vegetation indices for identifying dominant species in grassland communities and achieving a peak accuracy of 98.87% [16]. Prior studies further demonstrated that higher species coverage enhances model inversion accuracy [17]. These methods rely primarily on linear discriminant approaches from a statistical perspective, limiting their ability to model complex nonlinear relationships. Integrating hyperspectral remote sensing with advanced classification algorithms not only enhances grassland species identification accuracy but also improves monitoring precision, thereby establishing a robust framework for grassland conservation research.

Given that plant composition directly modulates grassland spectral properties—thereby impacting remote sensing-based species identification—and that species participation reflects ecological dominance within plant communities, two critical questions emerge: (1) How do shifts in species participation influence the spectral signatures of dominant plants in sericite–Artemisia desert grassland? (2) How do hyperspectral image recognition models respond to variations in species participation levels in terms of performance? To address these questions, this study employed ground-based hyperspectral imaging to analyze spectral response mechanisms of sericite–Artemisia plants to species participation. Key discriminative spectral features were identified, and characteristic band images with different species participation conditions were synthesized and applied to three deep learning networks by assessing the classification efficacy of models. These results establish a theoretical basis for advancing species identification, ecological monitoring, and targeted conservation strategies in degraded grassland ecosystems.

2. Materials and Methods

2.1. Study Area and Experimental Design

The study area encompasses the Bayang river township in Midong district, Urumqi city, Xinjiang, China (87°52′59″–87°55′13″ E, 44°00′16″–44°01′20″ N; Figure 1), with elevations ranging from 895 to 954 m. Annual precipitation averages approximately 250 mm (National Earth System Science Data Center).The dominant species is S. artemisia, and the subdominant species are the annual plants Ceratocarpus arenarius L.

In April 2022, three 550 m long transects were established in areas of sericite–Artemisia desert grassland, spaced 200 m apart, following the principle of vegetation uniform distribution. A 0.5 m × 0.6 m quadrat was systematically placed every 10 m along each transect for the hyperspectral imaging of plant communities. Owing to topographic variability, 53, 56, and 61 quadrats were deployed across the three transects, yielding a total of 170 sampling units.

The plant species within each quadrat were documented, with coverage assessed visually and aboveground biomass harvested via ground-level cutting post-spectral data acquisition. All aboveground biomass was clipped at ground level, cleared of impurities, and stored in labeled sample bags. Fresh weight was determined using a precision electronic balance, and species participation (calculated as relative coverage and relative aboveground biomass proportion) was categorized into three levels (low, medium, and high) based on predefined thresholds; the relevant calculation formula is as follows (see also

Table 1. Participation level and sample division.

Participation Level	Value	S. transiliense Image	C. arenarius Image
Low	0.0–0.3	51	16
Medium	0.3–0.6	29	33
High	0.6–1.0	20	121
Total		100	170

) [18]:

$$\mathrm{Relative} \mathrm{Coverage} \left(\mathrm{RC}\right)={\displaystyle \frac{\mathrm{Coverage} \mathrm{of} \mathrm{species} i}{{\displaystyle \sum \mathrm{Coverage} \mathrm{of} \mathrm{all} \mathrm{species}}}}\times 100\%$$

(1)

$$\mathrm{Relative} \mathrm{Aboveground} \mathrm{Biomass} \left(\mathrm{RAB}\right)={\displaystyle \frac{\mathrm{Aboveground} \mathrm{Biomass} \mathrm{of} \mathrm{species} i}{{\displaystyle \sum \mathrm{Aboveground} \mathrm{Biomass} \mathrm{of} \mathrm{all} \mathrm{species}}}}\times 100\%$$

(2)

$$\mathrm{Participation}={\displaystyle \frac{\mathrm{RC}+\mathrm{RAB}}{2}}$$

(3)

2.2. Hyperspectral Image Acquisition

Spectral data acquisition using the Soc710 VP imaging spectrometer (Surface Optics Corporation, San Diego, CA, USA) was conducted on clear days (wind speed ≤ level 3) between 11:00 and 17:00 Beijing time (UTC + 8). The instrument operates within a spectral range of 400–1000 nm, with 256 bands at 4.68 nm resolution and an imaging resolution of 1392 × 1040 pixels. During measurements, consistent lighting conditions were maintained for both the target and the reference gray board, ensuring identical horizontal alignment with the plant canopy. The lens was positioned vertically to the ground and oriented perpendicular to the solar azimuth to eliminate shadows and obstructions within the sampling area. Gray board calibration was performed immediately before and after each measurement session.

2.3. Data Processing

The spectral reflectance values of target species were extracted using the SRAnal710 system (Surface Optics Corporation, San Diego, CA, USA). Using SPSS 27 software, the means and standard deviations of the spectral reflectance at different levels of participation for S. transiliense and C. arenarius were calculated, along with a correlation analysis of reflectance values across five sensitive spectral bands. ENVI 5.6 software was employed to subset hyperspectral imagery for these species at varying abundance levels, while MATLAB R2024a generated annotated data labels. Deep learning models were implemented in Python, and visualization was conducted in Origin 2024 (Figure 2).

2.4. Feature Band Selection

The acquired spectral image samples exhibit numerous bands and substantial data volume, resulting in spectral redundancy and multicollinearity. Many wavelengths within the collected spectra are irrelevant to classification tasks, and removing these redundant bands enhances computational efficiency [19]. Feature band selection aims to extract a subset of wavelengths from the original spectra that optimally represent ground object characteristics, prioritizing high information content, low inter-band correlation, and high class separability. Consequently, this study applied the optimum index factor (OIF) for feature wavelength selection [20]. The workflow involved partitioning all bands into five spectral range subsets, selecting the band with the highest standard deviation within each subset, computing pairwise correlation coefficients, and iteratively calculating OIF values for all possible triple-band combinations to identify the optimal feature band set.

$$\mathrm{O}IF={\displaystyle \frac{{\displaystyle {\sum }_i^{\mathrm{n}}{S}_i}}{{\displaystyle {\sum }_{i=1}^n{\displaystyle {\sum }_{j=i+1}^n\left|R_{ij}\right|}}}}$$

(4)

The OIF is defined as follows, where S represents the standard deviation of the ith band, R denotes the absolute value of the correlation coefficient between the ith and jth bands, and n is the number of bands selected from all bands.

2.5. Establishment of Recognition Model

The training server was set up with a Python 3.10 virtual environment configured on Windows 11. The model framework employed was the PyTorch 2.0.0 deep learning framework. The hardware was equipped with an RTX 3090 graphics card, and the installed version was CUDA 11.3 to enable GPU-accelerated training. The hyperspectral images were proportionally resized to 480 × 400 pixels and input into DeepLabv3p, PSPNet, and UNet networks to develop recognition models, enabling the comparative analysis of their performance. All models were configured with a learning rate of 0.001, were trained over 200 epochs, and had a batch size of 4. During model training, the dataset was divided into 70% for training and 30% for validation to assess accuracy, following established protocols [21].

DeepLabv3p (DeepLab Version 3 Plus), an advanced iteration of the DeepLab series, is a deep learning architecture optimized for semantic segmentation tasks. Developed as an extension of DeepLabv3 with the aim of enhancing segmentation accuracy and computational efficiency, it employs atrous spatial pyramid pooling (ASPP) to capture multi-scale contextual information. The ASPP module utilizes convolutional layers with varying dilation rates to extract hierarchical features from hyperspectral imagery, enabling multi-receptive field fusion and the robust encoding of spatial–spectral details [22]. This architecture enhances the model’s capacity to discern objects of diverse scales and is primarily applied for automated pixel-level semantic feature extraction and classification in remote sensing applications.

PSPNet (pyramid scene parsing network) is a deep learning architecture designed for semantic segmentation, leveraging pyramid pooling and scene parsing modules to address challenges in complex scenes with multi-scale objects. By aggregating multi-scale contextual features, the model enhances global context representation, improving object segmentation and pixel-level classification accuracy [23].

UNet (universal network) is a popular CNN architecture and has demonstrated exceptional performance in biomedical and remote sensing image segmentation tasks. Characterized by a symmetrical encoder–decoder architecture with skip connections, UNet preserves spatial details by bridging feature maps between corresponding encoder and decoder layers during upsampling. This design facilitates multi-scale feature integration, enhancing segmentation precision across varying object scales—a critical capability for analyzing heterogeneous targets in hyperspectral imagery. The architecture is adaptable to domain-specific tasks through targeted optimization [24].

2.6. Accuracy Verification

Classifier performance is evaluated using metrics derived from the confusion matrix: overall accuracy (OA), user’s accuracy (UA), producer’s accuracy (PA), and the Kappa coefficient. OA quantifies the probability that a randomly sampled classification result matches its ground truth label. The Kappa coefficient assesses agreement between classification results and reference data while accounting for chance, offering a more robust consistency measure than OA alone. UA and PA evaluate classification reliability at the category level, with UA reflecting commission errors and PA addressing omission errors [25]. The formulas are defined as follows:

$$U{A}_{\mathrm{i}}={\displaystyle \sum _{i=1}^k\frac{N_{ii}}{N_{\mathrm{i}+}}}$$

(5)

$$P{A}_{\mathrm{i}}={\displaystyle \sum _{i=1}^k\frac{N_{ii}}{N_{+i}}}$$

(6)

$$OA={\displaystyle \sum _{i=1}^k\frac{N_{ii}}{N}}$$

(7)

$$K\mathrm{appa}={\displaystyle \frac{N{\displaystyle {\sum }_{i=1}^{\mathrm{k}}{N}_{ii}-{\displaystyle {\sum }_{i=1}^{\mathrm{k}}(N_{i+}{N}_{+i}}})}{N^2-{\displaystyle {\sum }_{i=1}^k(N_{i+}{N}_{+i})}}}$$

(8)

The expression in which N represents the total number of samples, k is the total number of categories, N_ii is the number of samples correctly classified into the correct category, and N_+i and N_i+ are the actual number of samples in the i class and the number of samples predicted as the i class, respectively.

3. Results

3.1. Spectral Characteristics Analysis

The absorption and reflectance characteristics of S. transiliense and C. arenarius under varying participation levels are illustrated in Figure 3. Both species exhibit consistent spectral curves with analogous reflection peaks and absorption valleys, following a parabolic trend of initial increase followed by a gradual decline. As shown in Figure 3a, medium participation demonstrates significantly higher reflectance across the entire spectral range compared to other levels. While high participation exhibits minimal reflectance in the 376–736 nm range, it displays a pronounced upward trend in the near-infrared (NIR) region. As shown in Figure 3b, plant reflectance increases with higher participation levels. Additionally, the spectral curve of low participation diverges markedly from medium and high participation profiles. Within the visible spectrum (376–750 nm), the medium and high participation curves exhibit substantial overlap; however, in the NIR region, medium participation reflectance increases sharply, differentiating it from high participation.

The spectral profiles of both plant species exhibit pronounced amplitude peaks at 500 nm, 560 nm, 700 nm, 770 nm, 960 nm, and 1000 nm compared to other spectral regions, suggesting enhanced discriminative capacity for distinguishing plant species across participation levels (Figure 3):

3.2. Feature Band Selection in Different Participation Levels

3.2.1. Calculation of Band Standard Deviation

Variations in species type and participation levels exhibited minimal influence on sensitive band selection within the blue region (499.69 nm). In the green region, sensitive bands were identified at 556.60, 567.07, 559.22, and 561.83 nm. The red region featured two sensitive bands (603.98 and 698.17 nm), while the red-edge region included three (772.71, 775.50, and 778.30 nm). The near-infrared region demonstrated sensitivity at 958.87, 964.75, and 997.24 nm (

Table 2. Sensitive bands and standard (SD) deviations for plant screening at different levels of participation.

Classification	Low				Medium				High
	S. transiliense		C. arenarius		S. transiliense		C. arenarius		S. transiliense		C. arenarius
	Band	SD	Band	SD	Band	SD	Band	SD	Band	SD	Band	SD
Blue (B)	499.69	0.038	499.69	0.049	499.69	0.039	499.69	0.038	499.69	0.038	499.69	0.036
Green (G)	556.60	0.050	567.07	0.061	556.60	0.050	556.60	0.054	559.22	0.048	561.83	0.047
Red (R)	698.17	0.046	698.17	0.060	698.17	0.050	603.98	0.049	698.17	0.045	698.17	0.047
Red edge (RE)	772.71	0.152	778.30	0.143	772.71	0.147	772.71	0.152	775.50	0.127	772.71	0.132
Near-infrared (N)	958.87	0.153	997.24	0.155	964.75	0.156	964.75	0.158	997.24	0.146	958.87	0.128

).

3.2.2. Band Correlation Analysis

The analysis of composite Figure 4A reveals minimum correlation coefficients of 0.53 (low, R and N bands), 0.60 (medium, R and RE band), and 0.83 (high, B and N bands). In composite Figure 4B, minimum correlation coefficients were 0.62 (low, B and N bands), 0.70 (medium, B and RE bands), and 0.43 (high, R and N bands).

Across all three participation levels for both species, B and G bands exhibited strong correlations with R, RE, and N bands, suggesting spectral redundancy. Conversely, lower correlations between B and RE, as well as R and RE, may better capture discriminative features for species identification (Figure 5).

3.2.3. Optimum Index Factor Value Calculation

Analysis of Figure 5a reveals optimal band combinations (RREN) for S. transiliense across participation levels, with optimum index factor (OIF) values of 0.17 (low), 0.16 (medium), and 0.12 (high). Figure 5b demonstrates that C. arenarius exhibits optimal RREN bands for low (OIF = 0.16) and high (OIF = 0.17) participation levels, whereas the medium participation level shows optimal BREN bands (OIF = 0.15).

3.3. Model Evaluation

3.3.1. Evaluation of Desert Grassland Identification Model

As illustrated in Figure 6A, when targeting S. transiliense as the primary species, all three recognition models exhibited progressive increases in OA and Kappa coefficients with rising participation levels. The UNet model outperformed others in desert grassland identification, achieving peak OA values of 96.39% (low), 96.81% (medium), and 98.20% (high), with corresponding Kappa coefficients of 0.92, 0.92, and 0.94. Conversely, Figure 6B demonstrates that models targeting C. arenarius showed declining OA and Kappa values as its participation increased. Despite this inverse trend, UNet maintained superior performance, yielding OA values of 98.16% (low), 95.98% (medium), and 93.56% (high), alongside Kappa coefficients of 0.92, 0.88, and 0.85 across participation levels.

Under both experimental conditions, DeepLabv3p, PSPNet, and UNet achieved OA values exceeding 91% and Kappa coefficients above 0.79 for desert grassland identification, confirming robust classification capabilities across all models. However, model performance diverged with increasing dataset complexity: The performance of the model with S. transiliense as the main recognition object is gradually improving, while the performance of the model with C. arenarius as the main recognition object is gradually decreasing (Figure 6).

3.3.2. Evaluation of Desert Grassland Major Identification Objects Models

Model accuracy for all classes improved proportionally with increasing S. transiliense participation, with UNet achieving the highest precision. Feature visualization analysis demonstrates UNet’s superior ability to delineate edge features of both species, followed by DeepLabv3p. Conversely, PSPNet exhibited significant misclassification errors, frequently assigning S. transiliense and C. arenarius pixels to bare land (Figure 7).

Precision metrics indicate that the UNet model achieved superior overall performance, with PA and UA for S. transiliense across low, medium, and high participation levels reaching levels of 73.72%, 86.32%, and 92.09% and 82.92%, 92.33%, and 97.94%. For bare land, these values were 98.19%, 98.46%, and 99.14% and 97.61%, 98.04%, and 99.05%. The PSPNet model exhibited declining PA and UA for C. arenarius recognition as S. transiliense participation increased. In contrast, DeepLabv3p and UNet showed no significant performance trends for C. arenarius, with minimum PA and UA values of 94.68% and 85.63%, and although the model exhibits relatively weak recognition capabilities for C. arenarius, it can still categorize it with a fair degree of accuracy (

Table 3. The classification results of three types of recognition objects based on the characteristic bands of S. transiliense.

Model	Participation Level	PA (%)			UA (%)
		S. transiliense	C. arenarius	Bare Land	S. transiliense	C. arenarius	Bare Land
DeepLabv3p	Low	55.86	92.25	95.98	89.50	88.73	95.25
	Medium	80.83	93.46	97.16	93.18	87.65	97.30
	High	91.61	89.36	99.22	96.83	89.98	98.44
PSPNet	Low	50.46	88.81	94.76	92.17	85.72	93.44
	Medium	79.42	86.92	95.74	88.07	84.04	95.50
	High	90.91	84.08	98.00	91.16	82.90	98.04
UNet	Low	73.72	95.64	98.19	82.92	95.35	97.61
	Medium	86.32	94.68	98.46	92.33	93.61	98.04
	High	92.09	96.87	99.14	97.94	85.63	99.05

).

As the proportion of C. arenarius in the community increases, the classification accuracy for S. transiliense and bare land decreases. Visualizations reveal reduced edge detection accuracy for S. transiliense across all models (Figure 8). Precision statistics indicate that, at high C. arenarius participation, the PA values for S. transiliense drop to 39.91%, 41.34%, and 45.43%. While UA values also decline, they remain relatively high (91.76%, 89.27%, and 84.00%). This suggests that the models may overfit to C. arenarius features in high-participation datasets, causing the severe under-detection of S. transiliense but fewer false positives. Bare land exhibits PA and UA above 94% in all models, likely due to its distinct spectral characteristics. C. arenarius PA and UA values increase with participation, but low-participation cases show misclassification (e.g., as bare land or S. transiliense), resulting in lower precision. A comprehensive comparison reveals that the UNet model performs the best, with the high participation level C.arenarius achieving PA and UA values of 88.36%, 92.27%, and 91.16% and 87.75%, 86.19%, and 89.23% (

Table 4. The classification results of three types of recognition objects based on the characteristic bands of C. arenarius.

Model	Participation Level	PA (%)			UA (%)
		S. transiliense	C. arenarius	Bare Land	S. transiliense	C. arenarius	Bare Land
DeepLabv3p	Low	93.14	91.67	99.01	94.97	83.66	99.13
	Medium	76.37	83.74	96.94	90.92	81.65	95.62
	High	39.91	88.68	96.27	91.76	87.58	94.18
PSPNet	Low	88.34	79.06	98.50	91.47	79.23	98.11
	Medium	79.54	86.57	97.04	91.73	82.26	96.37
	High	41.34	88.04	94.50	89.27	83.35	94.08
UNet	Low	91.34	88.36	99.28	95.94	87.75	98.77
	Medium	84.35	92.27	97.79	93.10	86.19	97.79
	High	45.43	91.16	96.76	84.00	89.23	95.44

).

4. Discussion

4.1. The Influence of Species Participation on Spectral Characteristics

Hyperspectral remote sensing data are rich in information and capable of precisely capturing subtle changes in plant spectra. Consequently, combining these data with deep learning networks enhances a model’s ability to differentiate subtle differences in plant involvement levels and improves classification accuracy. Research indicates that both S. transiliense and C. arenarius exhibit similar trends of reflectance absorption valleys and reflection peaks across three involvement levels, aligning with the typical spectral response characteristics of green plants to light. Specifically, they display chlorophyll absorption valleys in the visible light band and high reflection peaks in the near-infrared band due to leaf internal structure [26]. When S. transiliense is at a medium involvement level within the community, it exhibits higher reflectance and near-infrared band reflectance increases at high involvement levels. This may relate to increased aboveground biomass, leading to more leaf layers and enhanced multiple scattering effects in the near-infrared range [27,28]. Furthermore, when S. transiliense involvement exceeds 0.3, plant canopy coverage is relatively high, reducing soil background interference and highlighting its spectral characteristics. This aligns with Yang [29], who found that soil background impact on plant canopy reflectance depends on soil reflectance and vegetation coverage, with soil influence decreasing as vegetation coverage increases. In contrast, C. arenarius shows gradually decreasing reflectance as involvement levels increase, potentially due to spectral reflectance saturation reducing photosynthetic efficiency. In spring, when C. arenarius plants are small and densely distributed, spectral reflectance does not significantly respond to increases in biomass or leaf area index. The high leaf overlap creates a complex canopy structure, increasing shaded areas and reducing overall reflectance [30,31].

4.2. Selection of Characteristic Bands

Correlation analysis revealed that both plant species demonstrated relatively low correlation coefficients between the blue band and the red edge, as well as between the red band and the near-infrared band, across all three levels of participation. Studies indicate that the blue and red bands are influenced by chlorophyll absorption and scattering, whereas the red edge and near-infrared bands primarily reflect leaf structure, moisture content, and vegetation biomass [32,33]. These spectral bands are widely recognized for their ability to capture plant structural characteristics [34]. Therefore, the low correlation between these bands amplifies the differences between the characteristics of S. transiliense and C. arenarius under different participation levels. Different plants and distribution conditions require suitable band combinations to enhance recognition. By calculating the OIF values of band combinations, it was found that combinations of red and blue bands with red edge and near-infrared bands exhibit higher OIF values. These combinations can capture the unique and complementary multidimensional information of plant reflectance spectra, thereby reducing information redundancy [35] and offering advantages in identifying S. transiliense desert grasslands. The optimal band combination for S. transiliense at low, medium, and high participation levels is consistently RREN. This suggests that it maintains relatively consistent physiological characteristics across different participation levels, resulting in the information content of these bands being persistently optimal. Its survival strategy relies more on photosynthetic efficiency and canopy expansion, so the combination of red, red edge, and near-infrared bands can stably capture its key features. This highlights the universality and stability of S. transiliense as a perennial herbaceous plant and its dominant position in stable desert ecosystems. C. arenarius, however, has the optimal band combination of RREN at low and high participation levels, but BREN under medium participation. This may be because, at a participation level between 0.3 and 0.6, the spectral characteristics of the blue light band of the plant change more significantly and have stronger complementary information with other bands. This also reflects the specificity and instability of C. arenarius as an annual herb.

4.3. The Influence of Species Participation on the Accuracy of Identification Models

The DeepLabv3p, PSPNet, and UNet deep learning models demonstrate strong generalization ability in accurately segmenting plants of varying sizes and shapes, extracting effective features from complex backgrounds, and adapting to different species and environments [36,37,38]. Regarding the overall recognition of S. transiliense as the primary target, the OA and Kappa coefficients of the models increase with higher participation levels. The UNet model consistently achieves the highest accuracy across all participation levels, with OA values of 96.39%, 96.81%, and 98.20% and Kappa coefficients of 0.92, 0.92, and 0.94. This indicates that increased target plant participation enhances the separability of deep learning models and improves the precision of plant edge feature identification. When C. arenarius is the primary recognition object, all models maintain OA values above 91% and Kappa coefficients above 0.79, with UNet again performing optimally (OA values of 98.16%, 95.98%, and 93.56% and Kappa coefficients of 0.92, 0.88, and 0.85). However, accuracy decreases with increasing participation, suggesting that in small-scale, highly mixed grassland communities, spectral mixing effects may intensify, reducing the model’s ability to capture boundary features and thereby affecting discriminative power [39]. Under both recognition conditions, all deep learning models demonstrate high classification accuracy, confirming that hyperspectral images provide ample spectral information to support deep learning models across datasets of varying complexity. This underscores the feasibility of this method for desert grassland recognition. Notably, the UNet model achieves exceptional accuracy, aligning with Liu et al. [40], who reported UNet’s high performance in predicting desert grasslands, typical steppes, and meadow steppes, with the accuracy reaching 97.35%. Similarly, Guo et al. [41] also found that the UNet-based model effectively distinguishes maize and wheat seedlings from field weeds, achieving an accuracy rate of 99.3%.

From the perspective of identifying different objects, the UNet model demonstrates the most stable performance when focusing on S. transiliense as the primary target, particularly in distinguishing between S. transiliense, C. arenarius, and bare land. Under high participation conditions, the PA and UA values for S. transiliense and bare land reach as high as 92.09%, 92.09% and 97.94%, 97.94%, while the accuracy for C. arenarius remains stable at over 85%. In contrast, when C. arenarius is the primary recognition target, the model’s PA and UA values for S. transiliense and bare land exhibit a downward trend. Especially under high participation conditions, the model’s ability to identify S. transiliense significantly decreases, with the PA value dropping to a minimum of 39.91%, although the UA value remains above 84%. This indicates a higher rate of missed detections but a relatively low rate of false positives, with most S. transiliense pixels accurately identified. This phenomenon may be related to the “masking effect” of C. arenarius on the dominant background in spectral characteristics. In mixed pixels, C. arenarius, due to its high participation and distinct spectral features, dominates the observed spectrum, reducing the model’s sensitivity to S. transiliense. The resulting prediction probability distribution is biased toward the dominant class, leading to a decrease in the recognition rate of the secondary class and an increase in missed detections [42,43]. Bare land, as a background category, maintains an identification accuracy of over 94% across all models, indicating that its spectral features are stable and easy to classify, with minimal influence from vegetation participation. From the perspective of model performance, UNet consistently achieves high identification accuracy under various plant participation conditions. Its structural use of skip connections enables the better integration of multi-scale information, adapting to target differences in the complex background of desert grasslands. This finding is supported by other ecological remote sensing studies, such as Wang et al. [44], who confirmed the significant potential of the UNet model in accurately identifying the endangered species Populus euphratica in deserts with small training datasets.

5. Conclusions

This research demonstrates the applicability of combining hyperspectral remote sensing with deep learning methods for the high-precision identification of S. transiliense desert grassland plants, highlighting the significant influence of species participation on model performance and providing a scientific basis for real-time desert grassland monitoring. Species participation not only affects spectral information acquisition but also impacts model training and feature extraction effectiveness. For S. transiliense, increased participation within the community enhances the model’s ability to capture key features accurately. Conversely, With the increasing participation of C. arenarius in the community, more challenging to identify due to the spectral saturation and masking effects. A higher species participation level does not inherently guarantee higher information extraction efficiency, as reflectance properties and environmental factors must also be considered.

In this study, S. transiliense exhibits the universal and stable characteristics of a perennial herb, while C. arenarius demonstrates the specific and unstable traits of an annual herb. Bare land maintains a high recognition accuracy across all models and remains largely unaffected by species participation. The UNet model shows superior robustness and generalization in identifying desert grassland plants, offering a clear direction for selecting remote sensing classification models for sericite–Artemisia desert grassland. Due to the uneven distribution of grassland plants, some classes have a relatively small sample size when the samples are divided based on plant participation. This imbalance in the samples may also affect the generalization ability of the recognition model. Future improvements in model recognition within complex vegetation communities can be achieved through expanded datasets, optimized network structures, and multi-source data integration, thereby advancing hyperspectral remote sensing applications in ecological monitoring and biodiversity conservation.