Monitoring Vegetation Dynamics in Desertification Restoration Areas of Wuzhumuqin Grassland Ecosystem

Abstract

The desertified ecological restoration vegetation of Wuzhumuqin grassland plays an important role in the ecological restoration and protection of the region. However, there are few studies on the monitoring of the changes in ecological restoration vegetation in grassland sandy land in the past. In order to improve the low efficiency of ecological restoration vegetation monitoring, this study used Gaofen-6 (GF-6) remote sensing data to calculate the kernel Normalized Difference Vegetation Index (kNDVI) and vegetation coverage of ecological restoration vegetation and analyze their spatial and temporal trends. At the same time, a transform three-branch network structure based on deep learning is proposed to extract visual features. The kernel Normalized Difference Vegetation Index-position-temporal awareness transformer (kNDVI-PT-Former) model monitoring method based on two-phase remote sensing image features combined with kNDVI for spatio-temporal feature extraction can accurately obtain the vegetation changes in desertification ecological restoration in Wuzhumuqin grassland. The results show that the kNDVI of the study area shows an increasing trend from 2019 to 2024. The kNDVI value is 0.4086 in 2019 and 0.4927 in 2024. From the perspective of the change trend of vegetation coverage, the overall vegetation coverage of the Wuzhumuqin desertification restoration study area showed a gradual increase trend from 2019 to 2024, and the vegetation coverage increased by 19% in 2024 compared with 2019. The transformation of vegetation coverage from low level to high level in the study area is more prominent. Based on the self-built monitoring dataset of more than 5.2 million pairs of grassland vegetation changes, through model comparison and analysis, the kNDVI-PT-Former model obtains that the Class Pixel Accuracy (CPA) is 0.7295, the Intersection over Union (IoU) is 0.7228, and the overall monitoring accuracy of the model is improved by 11%. Furthermore, the stability of the model’s performance was confirmed through evaluation with five-fold cross-validation.

1. Introduction

Grasslands, constituting approximately 20% of the world’s terrestrial area, represent one of the most extensive global ecosystems [1]. These ecosystems possess dual economic and natural attributes [2]. However, exacerbated by overgrazing and climate change, grassland desertification has become increasingly severe [3]. Through persistent restoration efforts, vegetation diversity and coverage in desertified grassland areas have significantly improved, highlighting the critical need for robust monitoring of vegetation growth dynamics and coverage changes. Remote sensing technology has emerged as a powerful tool for grassland ecological monitoring due to its large-scale, rapid, and accurate acquisition of high spatio-temporal-resolution vegetation change information [4]. Compared to traditional methods, remote sensing offers superior spatial advantages, serving not only as a primary means for regional vegetation coverage assessment but also enabling global-scale estimations. Numerous vegetation indices have been developed, among which the Normalized Difference Vegetation Index (NDVI), proposed by Rouse et al. in 1974 [5], remains one of the most widely used indicators for monitoring global vegetation dynamics [6,7,8]. NDVI effectively captures spatial variations in vegetation across major climate zones and serves as a reliable proxy for surface vegetation coverage and growth status. To further enhance monitoring accuracy, Gustau Camps-Valls et al. [9] introduced a nonlinear generalization of NDVI. In 2021, Camps-Valls’ team proposed the kernel Normalized Difference Vegetation Index (kNDVI), demonstrating its superior precision through comparative experiments with NDVI and NIRv. Subsequent studies by Wang et al. [10] validated kNDVI’s robust performance in vegetation monitoring and emphasized its utility for terrestrial ecosystem assessments. Feng et al. [11] developed a comprehensive kNDVI dataset to characterize spatio-temporal vegetation coverage changes and climate-driven factors in the Yellow River Basin from 2001 to 2020. Unlike conventional RGB imagery, remote sensing data contain multiple spectral bands, offering richer information for analysis. Training semantic segmentation models requires large volumes of high-quality labeled data, with quantity and quality directly influencing model stability and adaptability [12]. For instance, Liu et al. [13] successfully extracted urban green space distribution from GF-2 imagery using DeepLabV3+ semantic segmentation networks. Comparative studies by Boston et al. [14] demonstrated that U-Net-based semantic segmentation outperformed random forests in Landsat multi-spectral land cover classification, achieving higher overall accuracy and F1 scores. Similarly, Nazila et al. [15] integrated artificial intelligence and deep learning for advanced vegetation canopy assessment and monitoring.

With the development of artificial intelligence technology, deep learning algorithm has been gradually applied to the field of vegetation change monitoring. However, existing studies still face challenges in accurately capturing vegetation dynamics in complex desertified grasslands. Given the limitations of traditional NDVI in sparse vegetation areas and the lack of deep learning models for grassland desertification monitoring, the core research questions of this study are

• How can the kernel Normalized Difference Vegetation Index (kNDVI) improve the accuracy of vegetation coverage estimation in desertified grasslands?
• Can a deep learning model integrating spatio-temporal features and kNDVI (kNDVI-PT-Former) enhance the precision of vegetation change monitoring in complex ecological restoration areas?
• What are the spatio-temporal dynamics of vegetation coverage in Wuzhumuqin Sandy Land from 2019 to 2024, and how do they reflect the effectiveness of ecological restoration?

At present, existing studies on grassland vegetation monitoring primarily rely on NDVI, which suffers from low accuracy (≤65% in sparse vegetation areas) due to soil background interference [16]. Although kNDVI has shown promise in nonlinear vegetation modeling, its application to grassland desertification restoration areas remains underexplored. Additionally, most deep learning models for remote sensing change detection (e.g., U-Net, PT-Former) neglect spectral index features like kNDVI, leading to suboptimal performance in capturing subtle vegetation dynamics [17]. This study focuses on long-term time-series monitoring during the peak vegetation growing season (August) to capture the core phenological signals of ecological restoration. These gaps are addressed by integrating kNDVI with a Transformer-based architecture, enabling the high-precision monitoring of ecological restoration effects.

2. Materials and Methods

This study utilizes multi-source remote sensing data, with the specific procedures illustrated in Figure 1, which provides a detailed overview of the workflow and the interrelationships among the various components of the research process.

2.1. Study Areas

Wuzhumuqin Sandy Land is distributed in the northeast of Xilingol League. It is one of the five major sandy lands in Inner Mongolia, with a total area of 3231.5 km². Among them, the Wuzhumuqin Sandy Land located in the Xiwuzhumuqin Banner traverses the central and western parts of Xiwuzhumuqin Banner. It is 150 km long from east to west and 10–15 km wide from north to south. It belongs to the narrow Changsha belt, with a total area of 2253.7 km², accounting for 10.03% of the total land area of the whole banner. Xiwuzhumuqin is located in the central part of the Mongolian Plateau in Inner Mongolia, the northern branch of the Greater Khingan Mountains [18], with longitude 116°21′~119°31′, latitude 43°52′~45°23′, and altitude 835~1957 m. It is a typical temperate arid and semi-arid continental climate region. In spring, it is dry and windy. In summer, although there is rain, it is unevenly distributed. In autumn, it is cool but frost and snow come early. Winter is cold and long. The average annual temperature of the whole banner is about 0~1 °C, the average annual precipitation is 345.8 mm, and it decreases from southeast to northwest, thus forming the difference between the eastern humid and the western arid [19]. In this study, part of the sandy land in Xiwuzhumuqin Banner was selected as the study area for monitoring vegetation changes in ecological restoration (Figure 2).

The soil properties in the study area have significant regional characteristics. Most of them are fixed sandy soil types. Most of them are severely desertified forest and grassland in the area. The species type is single, and almost loses the ability of automatic adjustment and recovery. The ecological environment is relatively bad, and it is gradually evolving to complete sandy land, and it is vulnerable to human activities and other disasters. Furthermore, water scarcity and severe sandstorm erosion of grasslands have rendered the ecological environment extremely fragile, prone to dust storms that seriously threaten the surrounding grasslands.

2.2. Data Source and Processing

The data used in this paper is GF-6 remote sensing satellite image data. GF-6 is an important part of China’s “High Resolution Earth Observation System.” It was successfully launched in June 2018. Its multi-spectral camera has a 2 m panchromatic resolution and an 8 m multi-spectral resolution, and has a wide-format imaging capability (up to 90 km). GF-6 image data is mainly used in agriculture, forestry, environmental protection, and other fields, which can provide rich data support for ground object classification, vegetation coverage monitoring, land use analysis and other research. The GF-6 image data used in this study is provided by the China Resource Satellite Application Center. The data covers the research area, including panchromatic band and multi-spectral band, which can meet the research needs. For the data of local vegetation growth season (August) with a cloud cover of 0, the time span is a multi-temporal image from 2019 to 2024.

Table 1. Spectral characteristics of bands.

Sensor	Parameters
GF-6	Spectral coverage (µm)	Panchromatic	0.45–0.90
		Blue (B1)	0.45–0.52
		Green (B2)	0.52–0.60
		Red (B3)	0.63–0.69
		Near-Infrared (B4)	0.76–0.90
	Spatial Resolution (m)	Panchromatic	2
		Multi-spectral	8

presents the band information required for computing kNDVI.

Then, the ENVI software (version 5.3) was used to preprocess the multi-spectral image and panchromatic image of the GF-6 remote sensing image, including radiometric calibration, atmospheric correction, ortho-rectification, image fusion, and image registration, to ensure the accuracy and reliability of the analysis results. Then, the mask extraction was performed on the preprocessed image to obtain the final study area.

2.3. Study Methods

The Kernel Normalized Difference Vegetation Index (kNDVI) is a kind of data based on kernel (machine learning) function. By introducing kernel technology, it can provide more robust and accurate vegetation information at different scales and with nonlinear changes [20].

2.3.1. kNDVI Calculation Method

kNDVI is an advanced index for remote sensing estimation of vegetation coverage, which aims to overcome the limitations of traditional NDVI under complex surface conditions. The calculation formula of NDVI is

$$NDVI={\displaystyle \frac{NIR-red}{NIR+red}}$$

(1)

In the formula, NIR and red represent the reflectance of near-infrared and red bands, respectively [21]. Although NDVI is widely used because of its simplicity and ease of use, it is easily affected by soil spectral characteristics when the soil background is significant, resulting in inaccurate estimation of vegetation coverage. In order to solve this problem, kNDVI introduces the influence of soil reflectance and uses the following formula:

$$kNDVI=\tan h\left({\left({\displaystyle \frac{NIR-red}{2\sigma }}\right)}^2\right)$$

(2)

In the formula, NIR represents near-infrared; red represents the red band; tan h represents hyperbolic tangent function; and $\sigma$ represents the length scale parameter specified in each specific application, which is the sensitivity of the index to sparse or dense vegetation [22], and is linearly proportional to the mean of near-infrared and red reflectance using remote sensing data images; when $\tau =0.5$, the value is a good balance between accuracy and simplicity [23]. Then a reasonable choice is to take the average value $\sigma =0.5\left(NIR+red\right)$. In this case, the formula of kNDVI is as follows:

$$kNDVI=\tan h\left(NDV{I}^2\right)$$

(3)

The improvement of kNDVI enables it to provide more accurate vegetation coverage estimation in a soil–vegetation mixed environment, especially in arid or sparse vegetation areas.

2.3.2. Vegetation Coverage Calculation Method

In this study, an improved dimidiate pixel model [24] was used to calculate vegetation coverage. According to the principle of the pixel dichotomy model, the NDVI value of a pixel can be expressed as the information contributed by the green vegetation part and the information contributed by the bare soil part. In this study, the vegetation coverage was calculated by introducing kNDVI value instead of NDVI value. Its calculation formula is

$$f={\displaystyle \frac{kNDVI-kNDV{I}_{soil}}{kNDV{I}_{veg}-kNDV{I}_{soil}}}$$

(4)

In the formula, kNDVI is the vegetation coverage value of the study area; kNDVI_soil is the kNDVI value of bare soil or unvegetated area; and kNDVI_veg is the kNDVI value of the pixel completely covered by vegetation, that is, the kNDVI value of the pure vegetation pixel [25,26]. In order to better obtain the growth characteristics of vegetation in the study area of sandy land restoration in Xiwuzhumuqin from 2019 to 2024, this study selected the most prosperous vegetation in August and analyzed its coverage. The monthly average value of vegetation coverage in August was calculated and obtained by the Google Earth Engine (GEE) platform. Subsequently, in the ArcGIS environment, the vegetation coverage is divided into five grades according to equidistant distance: low grade (0~20%), lower grade (20~40%), medium grade (40~60%), higher grade (60~80%), and high grade (80~100%) [27].

2.3.3. Trend Analysis

In order to study the spatial difference in vegetation change rate in the Xiwuzhumuqin sandy land restoration ecological area, the spatial variation in average vegetation coverage from 2019 to 2024 was simulated and analyzed by using the linear regression trend analysis method [28]. The linear regression trend analysis method can simulate the change trend of each grid and reflect the spatial characteristics of the change trend of vegetation cover in different periods. The calculation formula is

$${\theta }_{slope}={\displaystyle \frac{n\times {\displaystyle {\displaystyle \sum _{i=1}^n}i\times f_i-\left({\displaystyle {\displaystyle \sum _{i=1}^n}i}\right)\left({\displaystyle {\displaystyle \sum _{i=1}^n}{f}_i}\right)}}{n\times {\displaystyle {\displaystyle \sum _{i=1}^n}{i}^2}-\left({\displaystyle {\displaystyle \sum _{i=1}^n}i}\right)}}$$

(5)

In the formula, θ_slope is the slope of the change trend, n is the monitoring time (n = 6), and f_i is the vegetation coverage in the ith year. θ_slope > 0 indicated that the vegetation coverage showed an increasing trend. θ_slope < 0 indicates that the vegetation coverage is decreasing.

2.4. Eco-Restoration Vegetation Change Dataset

Due to the lack of public datasets for monitoring vegetation changes in grassland desertification ecological restoration, and we have a large number of self-built databases, including more than 5.2 million pairs of remote sensing image data, forming a dataset that integrates spatial and statistical information. The dataset can not only reveal the effect of ecological restoration, but also provide decision support for policy planning and long-term monitoring, provide important scientific support for desertification control and ecological restoration, and promote the sustainable management of land resources.

The self-built dataset consists of 5.2 million pairs of remote sensing image chips (size: 256 × 256 pixels), covering grassland scenes with varying levels of degradation (mild, moderate, severe) during the summer season. Due to geographic and climatic factors, grassland vegetation in northern China (Inner Mongolia Autonomous Region) reaches its peak growth only during summer, particularly in August each year. Vegetation conditions in other seasons are relatively poor and therefore excluded from the study. The annotation process was conducted by three remote sensing experts through visual interpretation, followed by double-blind cross-validation to ensure consistency. Discrepancies were resolved through consensus, and inter-annotator agreement was maintained at ≥95%. The dataset includes three channels: bi-temporal RGB images and a kNDVI difference map. The labels indicate vegetation change (1) or no change (0). Data variety encompasses the following: (1) Spatial diversity: images from different zones of Wuzhumuqin Sandy Land (northwest/south/central regions). (2) Temporal diversity: images acquired from 2019 to 2024, including growing season (August) observations. (3) Feature diversity: integration of spectral (kNDVI), spatial (texture), and temporal (multi-phase) features. The dataset is available upon request from the corresponding author, subject to a data use agreement that complies with local environmental protection regulations. Detailed metadata and labeling protocols can be provided to ensure reproducibility.

2.5. Build a Monitoring Model

The dataset was divided into training (70%), validation (20%), and testing (10%) subsets. The training set was utilized for kNDVI-PT-Former model parameter learning, the validation set served as a reference for hyperparameter optimization and early stopping, and the testing set provided an unbiased assessment of the final model performance. A time-series stratified sampling strategy was employed to ensure that each subset contained data from all study periods (2019–2024). This approach maintained the temporal representativeness of the data and prevented overfitting to specific seasons or years. Prior to cropping, radiometric calibration and atmospheric correction of the two-period remote sensing images were performed using ENVI 5.3. After splitting, data augmentation techniques including random cropping (size: 256 × 256 pixels), rotation (±90°), and horizontal flipping were applied to the training and validation sets to enhance the model’s generalization capability.

Next, the optimization parameter settings will be carried out. The Adam optimizer was selected due to its adaptability in handling sparse gradients and non-stationary objectives. Its default β1 and β2 values (0.9 and 0.999, respectively) were retained, and the weight decay was set to 1 × 10⁻⁵ to prevent overfitting. The initial learning rate was set to 1 × 10⁻⁴, and a step decay strategy was applied, reducing the learning rate by a factor of 0.9 every 10 epochs. This approach balanced the speed of convergence and the risk of overshooting the optimal solution. A batch size of 32 was chosen to optimize the utilization of GPU memory while ensuring sufficient statistical diversity within each batch update. The model was trained for a maximum of 100 epochs, with early stopping triggered when the validation loss did not decrease for 5 consecutive epochs. This strategy avoided overfitting and ensured efficient training.

The hyperparameter tuning was conducted using random search, an efficient strategy for exploring the hyperparameter space. This approach randomly sampled 50 combinations from predefined search ranges, substantially reducing computational costs compared to grid search. The key hyperparameters and their ranges are listed in

Table 2. Hyperparameter search space for model optimization.

Hyperparameter	Search Range
Learning Rate	[1 × 10−5, 1 × 10−3]
Batch Size	[16, 64]
Dropout Rate	[0.1, 0.5]

:

The mean cross-entropy loss on the validation set was employed as the primary metric to evaluate hyperparameter combinations. The configuration yielding the minimum validation loss was selected as the optimal setup for the final model, ensuring robust generalization to unseen data. (Note: These settings were implemented using Python’s scikit-learn library with 5-fold cross-validation on the validation subset to ensure robustness).

2.5.1. kNDVI-PT-Former Model Construction

With the development of deep learning, remote sensing (RS) image change detection (CD) methods have made significant progress. However, many convolutional neural network (CNN)-based methods are limited in capturing long-distance dependencies due to receptive domain constraints. Relying on the self-attention mechanism, Transformer effectively implements global information modeling and is widely used in CD tasks. Nevertheless, the Transformer-based CD method still has problems such as pseudo-variation and incomplete edges, which is due to the lack of position and time correlation in dual-temporal RS images. In order to solve this problem, a position-temporal awareness transformer (PT-Former) model [29] has been proposed, which models the relationship between position and time in dual-temporal images. Specifically, the Siamese network connected to the position-aware embedding module (PEM) is used as a feature encoder to extract the features of the changed region. Then, a temporal difference perception module (TDPM) is designed to capture the cross-temporal shift and enhance the difference perception ability during the cross-temporal interaction. At the same time, the context information of the ground object is aggregated by the fusion block, and the spatial relationship is reconstructed under the guidance of the bi-temporal features. However, the PT-Former model performs poorly in extracting the characteristics of ecological restoration vegetation changes in complex environments. Therefore, this study proposes the kNDVI-PT-Former algorithm model.

The kNDVI-PT-Former model is based on the PT-Former model. PT-Former uses three channels of dual-phase RS images as input (including label data input channel). The improved model adds an input channel. The kNDVI calculated by the two images is differentiated to obtain the changed kNDVI. The changed kNDVI is used as a new input channel to obtain the kNDVI-PT-Former model. By extracting the dual-temporal features and the feature information of kNDVI, after the spatial resolution reduction and semantic extraction, the feature map is input into PEM to generate position embedding, and the features are embedded by PEM and flat operation. In order to capture cross-temporal shift and enhance the difference perception ability during cross-temporal interaction, dual-temporal features and kNDVI features are input into TDPM. The features continue to pass through the Transformer cross-encoder with cross-attention, and the features are reshaped. At the same time, the two-phase RS feature map from the residual stem is down-sampled and fused by the pre-fusion block. Next, the features extracted from the Transformer encoder are aggregated by the post-fusion blocks. In this process, the extracted dual-phase RS images are first fused, and then the fused image is fused with the feature map of kNDVI. The fusion process is based on the feature map of kNDVI. Then, the fused features are input into the Transformer decoder. Then, in the same steps as the original model, in order to enhance the features of the changed region, the features extracted by the Transformer decoder and the post-fusion block are connected. Finally, in order to infer the change map of the observed scene, the classifier head is introduced to generate the change map. The schematic diagram of the kNDVI-PT-Former model is shown in Figure 3.

A. Fusion Block

The fusion of dual-temporal features and the fusion with kNDVI features are crucial in CD, and complete change information is formed under the guidance of each feature. At the same time, in order to integrate the context information of the ground object and reconstruct the spatial relationship of the ground object, a fusion block is designed to fuse the dual-temporal features and then fuse with the kNDVI features. Firstly, the size of F₁ and F₂ is H × W × C flattened and labeled as T₁ and T₂, and the size of F_k is H × W × C flattened and labeled as T_k, both of which are N × C, where N = HW. Then, T₁ and T₂ are connected, and the global information is extracted by using layer normalization and introducing self-attention, and the output is marked as T₁₂. Then, T₁₂ and T_k are also connected and normalized by layer. In order to obtain the change in the relationship between the two-time phase marker and the kNDVI marker, the global context information is also extracted by self-attention. The fusion block schematic diagram Figure 4 is as follows:

The kNDVI-PT-Former model adds feature extraction of kNDVI, and the extraction method is similar to the extraction method of T₁ and T₂ in the PT-Former model. Then, kNDVI features are added to the fusion module for fusion. In the fusion process, advanced dual-phase features are fused. The calculation formulas are as follows.

$$T_1=Flatten(F_1),T_2=Flatten(F_2)$$

(6)

$$T_{12}=SA(LN(C\mathrm{oncat}(T_1,T_2)))$$

(7)

where Flatten (·) represents the flattening operation.

After the fusion of dual-phase features is completed, it is fused with kNDVI features, and the fusion process is based on kNDVI features. By enhancing the ability to perceive spatial changes, local and global location information is extracted, respectively, and then these two parts are fused. The calculation formulas are as follows.

$$T_{12}=Flatten(F_{12}),T_k=Flatten(F_k)$$

(8)

$$T_{12,k}=SA(LN(C\mathrm{oncat}(T_{12},T_k)))$$

(9)

2.5.2. Model Training and Parameters

The model is implemented based on the PyTorch framework and runs on a server equipped with dual NVIDIA A100 Gpus (Nvidia, Santa Clara, CA, USA). The training adopts 5-fold cross-validation and avoids overfitting through the early stop method. The average training time is approximately 12 h per model. The technical parameters and training configuration are shown in

Table 3. Technical parameters and configuration.

Parameters/Configuration	kNDVI-PT-Former
Framework	PyTorch 2.0 + CUDA 12.1
Hardware	NVIDIA A100 40 GB GPU × 2
Optimizer	AdamW (learning rate = 5 × 10−5) weight decay = 0.01)
Batch size	32
Training rounds	200 epochs (early stop method patience = 10)
Loss function	binary cross-entropy + Dice loss
Evaluation index	CPA, IoU, F1-score

.

3. Results and Discussion

3.1. kNDVI Analysis

By comparing the kNDVI of the study area in 2019 and 2024, it was found that the kNDVI gradually increased after the implementation of the restoration measures, indicating that the vegetation growth was improved. The average value of kNDVI in the study area in 2019 was 0.4086, and the kNDVI vegetation index in 2024 reached 0.4927. The kNDVI value in 2024 was significantly higher than that in 2019, indicating that the ecological restoration measures were gradually effective. From the spatial distribution of kNDVI changes in the study area of Xiwuzhumuqin sandy land restoration from 2019 to 2024 (Figure 5), kNDVI increased significantly, and the increased areas were mainly distributed in the northwest, south, and central regions of the study area. The eastern region has little change, and there is a sporadic increase in distribution. Among them, the most obvious increase was in the northwest and south of the study area, while the kNDVI in the southwest of the study area showed a patchy decreasing trend.

3.2. Spatio-Temporal Variation Characteristics of Vegetation Coverage

The spatial distribution of vegetation coverage in the study area of desertification restoration in Xiwuzhumuqin Banner from 2019 to 2024 is shown in Figure 6, and the statistical results of area and proportion of each grade are shown in

Table 4. Changes in the area and proportion of vegetation coverage grade in 2019 and 2024.

Year	Parameter	kNDVI Vegetation Index Classification
		Low	Relatively Low	Medium	Relatively High	High
2019	Area (km2)	82.96	134.27	113.69	70.07	52.42
	Percentage (%)	17.98	29.1	24.64	16.92	11.36
2024	Area (km2)	58.78	70.69	114.1	124.95	92.88
	Percentage (%)	12.74	15.32	24.73	27.08	20.13
Comparison of Results	Area (km2)	−24.18	−63.58	0.41	54.88	40.46
	Percentage (%)	−5.24	−13.78	0.09	10.16	8.77

. It can be seen from Figure 6 that the vegetation coverage of the study is gradually increasing from 2019 to 2024. The vegetation coverage in most areas is on the rise, and the spatial distribution shows a gradual increasing trend from south to north. The high value areas of vegetation coverage are mostly distributed in the northern, central, southern, and some western regions of the study area. Among them, the southern and northern regions increased most significantly, the central region showed a slow sporadic patchy increase trend, while the southwest region of the study area showed a patchy decrease trend. This is because most of the area is dominated by bare soil and sandy land, with sparse vegetation and serious desertification.

According to Table 3, the area of high-grade and higher-grade vegetation coverage in the study area increased by 40.46 km² and 54.88 km², respectively, from 2019 to 2024. The area of medium-grade vegetation coverage fluctuated a little, while the area of low-grade and lower-grade vegetation coverage decreased by 24.18 km² and 63.58 km², respectively. In general, from 2019 to 2024, the vegetation coverage in the study area of desertification restoration in Xiwuzhumuqin Banner showed a trend from a low level to high level, indicating that the vegetation increased significantly and the vegetation coverage showed an upward trend. Among them, the vegetation coverage in the study area reached 47.21% in 2024 and only 28.28% in 2019. According to the results of vegetation coverage, the vegetation coverage in the study area increased by 19 percentage points in 2024 compared with 2019. This change reflects the feasibility and effectiveness of measures such as ecological restoration and forest closure in improving vegetation growth and coverage in this area.

Using the linear trend analysis method, the change trend of vegetation coverage was analyzed by the grid calculator of ArcGIS 10.8, and the final conclusion was divided into five grades by natural classification: significant increase, slight increase, basically unchanged, slight decrease, and significant decrease [30]. The results (Figure 7) show that from 2019 to 2024, the overall vegetation coverage trend in the study area of desertification restoration in Xiwuzhumuqin Banner increased, and there were significant differences in its spatial distribution and change trend.

Table 5. Trend statistics of vegetation coverage in 2019–2024.

Vegetation Coverage Change Slope Range	Change Degree	Area (km2)	Percentage (%)
−8 < Slope ≤ −5	Significant Decrease	38.99	8.45
−5 < Slope ≤ −2	Slight Decrease	118.4	25.66
−2 < Slope ≤ 0	Basically No Change	109.44	23.72
0 < Slope ≤ 2	Slight Increase	120.47	26.11
2 < Slope ≤ 8	Significant Increase	74.1	16.06

shows that vegetation increased in 42.17% of areas, decreased in 34.11%, and remained essentially unchanged in 23.72%. The area of vegetation increase is mainly concentrated in the north, south, and part of the middle of the study area, with a significant increase of 74.1 km² and a slight increase of 120.47 km². The reduction part is distributed in all regions. In general, the reduction in vegetation is mainly concentrated in the southwest, a small part of the central and northeast regions, and the reduction is more serious. Among them, the area of slight reduction is 118.4 km², and the area of significant reduction is 38.99 km². Most of these reduction areas are located in the core area of desertification and areas with frequent human activities. In addition to being affected by climatic factors, overgrazing in the early stage has caused serious degradation of vegetation cover.

To further clarify the climatic context for vegetation dynamics,

Table 6. Statistics of some meteorological data in 2019 and 2024.

Year	Precipitation (mm)	Mean Air Temperature (°C)	Maximum Air Temperature (°C)	Minimum Air Temperature (°C)	Average Wind Speed (m/s)
2019	313.7	5.3	32.9	−24.1	6.3
2024	327.2	4.6	31.5	−23.3	5.4
August 2019	109.8	20.7	32.6	7.2	2.4
August 2024	128.3	19.6	30.3	6.3	1.6

shows key 2019–2024 meteorological data (annual and August metrics).

Annually, 2024 precipitation (327.2 mm) rose 4.3% vs. 2019 (313.7 mm), boosting water availability. Mean temperature dropped (4.6 °C vs. 5.3 °C), with more extreme minima (−24.1 °C vs. −23.3 °C), potentially stressing herbaceous vegetation. Wind speed fell (5.4 m/s vs. 6.3 m/s), reducing erosion and stabilizing microenvironments.

In August (a critical growth month), 2024 precipitation (128.3 mm) was 16.9% higher than 2019 (109.8 mm), supplementing soil moisture. Cooler mean temps (19.6 °C vs. 20.7 °C) avoided heat stress, and lower wind speed (1.6 m/s vs. 2.4 m/s) cut physical damage and moisture loss, aiding physiology.

These changes—more growing, more season rain, moderated temps, less wind—likely improved conditions for vegetation growth.

3.3. Construction of Datasets

3.3.1. Data Image Source

The GF-6 remote sensing image is used as the data source. The multi-spectral image with a resolution of 8 m and the panchromatic image with a resolution of 2 m are preprocessed, respectively, and then fused and registered to form a multi-spectral image with a resolution of 2 m.

3.3.2. Classification System

The system is shown in

Table 7. Classification system.

ID	Change Type
1	Vegetation
0	Sand (Non-vegetation)

below.

3.3.3. Part of the Dataset Display

This is shown in Figure 8 below.

3.4. Model Monitoring Results and Analysis

3.4.1. Evaluation Criteria

In this study, the random forest model, PT-Former model, and kNDVI-PT-Former model were used to build and compare the models of the data on vegetation change monitoring in grassland desertification ecological restoration. By inputting the dataset of vegetation change characteristics, the results of monitoring vegetation change by the above three models are obtained. Subsequently, the accuracy of the above models is evaluated. The semantic segmentation evaluation index is mainly used to measure the classification accuracy of each pixel in the image and the accuracy of the segmentation area. Common metrics include Pixel Accuracy (PA), Class Pixel Accuracy (CPA), Mean Pixel Accuracy (MPA), Intersection over Union (IoU), Mean IoU (mIoU), etc. [31,32].

This study mainly uses the CPA and the IoU to evaluate the accuracy of the model.

• CPA

Class Pixel Accuracy is a commonly used evaluation index in semantic segmentation tasks. It measures the accuracy of the model’s pixel-level prediction for each category. It refers to the proportion of pixels in each pixel of the image that the model correctly classifies as a specific category of pixels. The formula is as follows.

$${CPA}_i={\displaystyle \frac{T{P}_i}{T{P}_i+F{P}_i+F{N}_i}}$$

(10)

The following are used in the formula. TP_i: The model correctly predicts the number of pixels of class i.

FP_i: The model incorrectly predicts the number of background pixels of class i.

FN_i: The model incorrectly predicts the number of class i pixels as the background class.

2.

IoU

Intersection over Union is a commonly used evaluation index in semantic segmentation, which measures the degree of overlap between prediction results and real annotations. The higher the IoU value, the more accurate the prediction results of the model. The formula is as follows.

$$I\mathrm{o}U={\displaystyle \frac{TP}{TP+FP+FN}}$$

(11)

The following are used in the formula. TP (True Positive): Correctly predict the number of pixels for the target class.

TN (True Negative): Correctly predict the number of pixels for the background class.

FP (False Positive): False prediction is the number of background pixels of the target class.

FN (False Negative): Error prediction is the number of target pixels of the background class.

3.

Loss Function

In machine learning and deep learning, the loss function (or cost function) quantifies the discrepancy between the predicted output and the true target. It serves as the optimization objective during model training, guiding parameter updates via gradient descent or other optimization algorithms. The formulation of the hybrid loss function is given as follows:

$$L_{hybrid}=L_{BCD}+\lambda L_{Dice}$$

(12)

where λ is the coefficient defined in the hybrid loss function.

To explain the concepts of BCE loss and Dice loss, we consider the predicted change map $\widehat{Y}$ and the reference map Y as sets of pixels, represented, respectively, as $\widehat{Y}$ = {${\widehat{y}}_i$, i = 1, 2, …, N} and Y = {$y_i$, i = 1, 2, …, N}. Here, ${\widehat{y}}_i$ denotes the probability of change in the ith pixel, while $y_i$ represents the reference value of the ith pixel. In this context, a value of 0 indicates an unchanged pixel, whereas a value of 1 signifies a changed pixel. The total count of pixels in the change map is denoted by N. The combined objective function of these two loss functions can be formulated as

$$L_{BCD}=-{\displaystyle \frac{1}{N}}{\displaystyle \sum _{i=1}^N{y}_i}lo{g}_2{\widehat{y}}_i+\left(1-y_i\right)lo{g}_2\left(1-{\widehat{y}}_i\right)$$

(13)

$$L_{Dice}=1-{\displaystyle \frac{2{\displaystyle {\sum }_{i=1}^N{y}_i{\widehat{y}}_i}}{{\displaystyle {\sum }_{i=1}^N{y}_i+{\widehat{y}}_i}}}$$

(14)

4.

Cross-Validation

Cross-validation is a common method for evaluating the generalization ability and stability of a model. The core idea is to repeatedly divide the dataset into the training set and the validation set, and avoid the bias caused by a single division through multiple rounds of training. This study uses a 5-fold cross-validation method:

(1)

Data Partitioning

The preprocessed remote sensing dataset (including images and labels) is randomly divided into 5 subsets (Subset 1–5) of similar size, ensuring sample independence and uniform distribution across subsets.

Note: If the data contains spatio-temporal correlation (e.g., multi-temporal images of the same region), spatial stratified sampling is used to avoid allocating samples from the same region to different subsets.

(2)

Multi-Round Training and Validation

Round 1: Train the model using Subsets 2–5, validate with Subset 1, and record validation metrics (e.g., overall accuracy CPA, F1-score).

Round 2: Train with Subsets 1, 3–5, validate with Subset 2, and repeat metric recording.

This process is repeated for 5 rounds to obtain 5 sets of metric results.

(3)

Result Analysis

Calculate the mean (to measure the model’s average performance) and standard deviation (to measure stability) of the metrics.

3.4.2. Random Forest Model

The vegetation change dataset is put into the random forest model for training and testing. This process does not need to input kNDVI, only two remote sensing image data and label data in the dataset are needed. The results of monitoring vegetation changes using the random forest model are as follows:

CPA: 0.5860. It shows that the classification and recognition accuracy of the model for vegetation is low.

IoU: 0.5831. It shows that the model has some errors in the classification boundary of vegetation and non-vegetation (sandy land). The output of the model is shown in Figure 9.

The vegetation changes in the whole study area are shown in Figure 10.

3.4.3. PT-Former Model

The vegetation change dataset is put into the PT-Former model for training and testing. This process does not need to input kNDVI; it only needs to input two remote sensing image datapoints and label datapoints in the dataset. The results of monitoring vegetation changes using the PT-Former model are as follows:

CPA: 0.6194. It shows that the classification and recognition accuracy of the model for vegetation is low.

IoU: 0.6157. It shows that the model has some errors in the classification boundary of vegetation and non-vegetation (sandy land). The output of the model is shown in Figure 11.

The vegetation changes in the whole study area are shown in Figure 12.

3.4.4. kNDVI-PT-Former Model

The vegetation change dataset is put into the kNDVI-PT-Former model for training and testing. This process inputs all the data in the dataset, including two-phase remote sensing image data, kNDVI data and label data in the dataset. The results of monitoring vegetation changes using the kNDVI-PT-Former model are as follows:

CPA: 0.7295. It shows that the accuracy of classification and recognition of vegetation by this model is improved compared with the model before improvement.

IoU: 0.7228. It shows that the model has improved the classification boundary performance of vegetation and non-vegetation (sandy land) compared with the model before improvement. The output of the model is shown in Figure 13.

The vegetation changes in the whole study area are shown in Figure 14.

The stability of the model was evaluated by using 5-fold cross-validation. The dataset was randomly divided into 5 subsets. Each time, 4 subsets were used for training and 1 subset for validation. The average value was taken after repeating the process 5 times. The results showed that the mean CPA of kNDVI-PT-Former was 0.732 ± 0.016 and the IoU was 0.725 ± 0.021. This result indicates that the datapoints are closely clustered around the mean, demonstrating the stable performance of the model.

3.4.5. Comparative Analysis of Models

The random forest model, PT-Former model and kNDVI-PT-Former model were verified and analyzed by the vegetation change monitoring dataset of grassland desertification ecological restoration in the study area of Wuzhumuqin. Random forest was chosen as a representative machine learning method to establish a baseline, while PT-Former—sharing the Transformer architecture—allows direct comparison of kNDVI feature integration. This highlights the synergistic effect of spectral indices (kNDVI) and spatio-temporal modeling. Figure 15 and

Table 8. Model accuracy comparison.

MODEL	CPA	IoU
Random Forest	0.5860	0.5831
PT-Former	0.6194	0.6157
kNDVI-PT-Former	0.7295	0.7228

are the accuracy comparison of the three models. The accuracy of the kNDVI-PT-Former model is the best in these models, followed by the PT-Former model, and the performance of the random forest model is relatively poor. Therefore, the PT-Former model based on deep learning and the kNDVI-PT-Former model can extract complex features from multi-spectral remote sensing data more accurately than the random forest model based on machine learning. In particular, the kNDVI-PT-Former model can effectively identify small vegetation changes, and the overall model monitoring accuracy is improved by 11%. Therefore, the use of the kNDVI-PT-Former model for vegetation change monitoring in complex environments has significant advantages and more accurate monitoring results.

Combined with the area and types of grassland desertification ecological restoration projects of Xiwuzhumuqin Banner Forestry and Grassland Bureau over the years, field investigation and verification were carried out in mid-November 2024, and some ecological restoration projects were selected for monitoring and verification in this study area every year. In the study area, the local relevant departments carried out fencing and artificial grass ecological restoration measures on 2.13 km² of sandy land at the beginning of 2021. As shown in Figure 16, the restoration vegetation is growing well and the ecological restoration effect is remarkable.

At the beginning of 2022, fencing and artificial afforestation ecological restoration measures were carried out on 1.33 km² of sandy land. As shown in Figure 17, the vegetation gradually grew and the ecological restoration effect was obvious.

At the beginning of 2023, artificial afforestation ecological restoration measures were carried out on 0.691 km² of sandy land. As shown in Figure 18, due to the short vegetation, it is difficult to identify the existence of vegetation through remote sensing images, and only part of the vegetation can be monitored.

At the beginning of 2024, artificial afforestation ecological restoration measures were carried out on 0.4 km² sandy land, as shown in Figure 19. The vegetation restored by 2024 is shorter and smaller, and it is difficult to identify vegetation in remote sensing images. Only a small part of vegetation can be monitored. Therefore, it is difficult to monitor the presence of vegetation within 2 years after ecological restoration in grassland desertification areas, especially the use of vegetation such as Caragana korshinskii and saplings for ecological restoration, and the effect is not obvious in the short term.

Taking the result map of vegetation monitoring change in the study area output by the kNDVI-PT-Former model as an example, as shown in Figure 20, the area of ecological restoration projects in recent years is marked in the map, and then the monitoring effect maps of each model in the ecological restoration project area are extracted, respectively, and then the monitoring effect of each model is verified by comparative analysis. Among them, the yellow box indicates the ecological restoration project area in 2021; the blue box indicates the ecological restoration project area in 2022; the purple box and orange box represent the ecological restoration project area in 2023 (only the restoration measures are different); and the light blue box indicates the ecological restoration project area in 2024.

The three model monitoring results of the ecological restoration area in 2021 are extracted, respectively, as shown in Figure 21. (a) is the result of the random forest model output, (b) is the result of the PT-Former model output, and (c) is the result of the kNDVI-PT-Former model output. Comparing the three maps, it can be seen that the kNDVI-PT-Former model has the best monitoring effect.

The three model monitoring results of the ecological restoration area in 2022 are extracted, respectively, as shown in Figure 22. (a) is the result of the random forest model output, (b) is the result of the PT-Former model output, and (c) is the result of the kNDVI-PT-Former model output. Comparing the three maps, it can be seen that the kNDVI-PT-Former model has the best monitoring effect.

The areas repaired in 2023 and 2024 are no longer compared due to the relatively small amount of vegetation that can be monitored. The time series map maintains the same geographical range, projections, and color schemes to ensure spatial comparability. The coordinate system for all the above model output figures is WGS84 (EPSG: 4326). The scale of Figure 9, Figure 10, Figure 11, Figure 12, Figure 13, Figure 14, Figure 15 and Figure 20 is 1:200,000, while the scale of Figure 21 and Figure 22 is 1:20,000.

Through comparative analysis with random forest and PT-Former, the kNDVI-PT-Former model, based on an advanced deep learning architecture, integrates time-series analysis and spatial feature extraction techniques. It can break through the limitations of traditional monitoring methods and achieve high-precision dynamic monitoring of vegetation for grassland and sandy land ecological restoration. In practical applications, this model, with its high-resolution remote sensing image processing capability, can cover vast grassland and sandy areas, quickly identify the vegetation changes at different succession stages, and accurately capture the tiny fluctuations in vegetation coverage and biomass, as well as the evolution trend of community structure. Whether it is the assessment of the vegetation restoration process in the ecological restoration project area or the early warning of potential desertification areas, the kNDVI-PT-Former model can provide detailed and reliable data support. This enables relevant departments and research institutions to promptly grasp the effectiveness of ecological restoration, scientifically adjust strategies for combating desertification, and rationally allocate human and material resources, laying a solid technical foundation for establishing a long-term ecological protection mechanism and promoting the sustainable development of grassland and sandy land ecosystems.

Therefore, in complex environments, the use of kNDVI combined with the Transform model for vegetation change monitoring has significant advantages, which can capture long-term vegetation dynamics and support ecological monitoring and climate change research. In this paper, however, no comparative experiments were conducted using other models. In the future, it may be considered to apply more advanced models to vegetation change monitoring in ecological restoration to enhance model diversity. This study has achieved good results in the monitoring of vegetation change in grassland desertification ecological restoration, but it still needs to be refined and improved in the algorithm of extracting vegetation change characteristics in the network layer, so as to provide ideas for future research on grassland vegetation change monitoring. While the kNDVI-PT-Former model demonstrates consistent performance improvements across cross-validation folds (e.g., 11% CPA gain over PT-Former, Section 3.4.5), the absence of formal statistical tests reflects the study’s focus on methodological innovation rather than exhaustive comparative inference. Meanwhile, although this study is based on single-phase data, cross-seasonal validation may enhance the model’s robustness. In future research, multi-seasonal imagery data—such as those collected during the spring green-up period and the autumn biomass peak period—will be integrated with correlation analysis of climatic factors to enhance the comprehensiveness and accuracy of monitoring. The study also plans to incorporate additional quantitative validation experiments and analyses, with the aim of strengthening the applicability and reliability of the monitoring model across diverse environments and conditions.

However, vegetation change monitoring via remote sensing imagery also has certain limitations. Key limitations include the following aspects: First, optical remote sensing data are significantly perturbed by cloud cover, atmospheric scattering, and climatic influences (e.g., precipitation events), which introduce noise into vegetation index calculations. Furthermore, the compromised temporal continuity of time-series datasets often limits the ability to capture rapid vegetation growth phases during critical phenological windows (e.g., spring budburst or monsoon-driven biomass accumulation). Most critically, low-spatial-resolution remote sensing data (e.g., >10 m) are increasingly inadequate for high-precision vegetation monitoring requirements, while high-resolution datasets (e.g., 0.5 m resolution) remain prohibitively expensive and lack universal accessibility for large-scale applications. Therefore, how to utilize remote sensing techniques for more efficient and precise vegetation change monitoring remains a research topic warranting in-depth investigation. Additionally, this study identifies several commonalities with prior research. First, the integration of multi-source data enriches the dataset and establishes a robust foundation for subsequent research. For example, Fan et al. (2024) achieved effective dynamic vegetation cover monitoring in rare earth mines by utilizing multi-source remote sensing data [23]. Second, model comparisons highlight the superiority of specific frameworks. For instance, Li et al. (2024) evaluated four object detection models—classical SSD, RetinaNet, YOLOv3, and Faster R-CNN—for tree detection and found that the SSD model demonstrated the best balance between inference speed and performance on their dataset, positioning it as a powerful target detection solution [33]. The unique contribution of this study lies in its innovative integration of vegetation indices into deep learning algorithms for vegetation change monitoring, a methodology that outperforms traditional models (random forest and PT-Former) in detection accuracy.

4. Conclusions

From the change in kNDVI in 2019–2024, kNDVI gradually increased, indicating that the growth of vegetation has improved. The average value of kNDVI in the study area in 2019 was 0.4086, and that in 2024 reached 0.4927. The value of kNDVI in 2024 was significantly higher than that in 2019. In general, kNDVI showed an upward trend in the past 6 years, indicating that ecological restoration measures have gradually taken effect.

From the perspective of the change trend of vegetation coverage, the overall vegetation coverage of the desertification restoration research area in Xiwuzhumuqin Banner showed a gradual increase trend from 2019 to 2024, and the vegetation coverage increased by 19 percentage points in 2024 compared with 2019. At the same time, the transformation of vegetation coverage from low-level areas to high-level areas in the study area is more prominent. The vegetation coverage in 23.72% of the entire study area remains basically unchanged, and the area of vegetation improvement accounts for 42.17% of the total area of the study area. It is mainly distributed in the northern, southern, and central regions of the study area. The degraded area accounted for 34.11%, and the degraded parts were distributed in all regions, but the degree of degradation was more serious in the southwest and northeast regions.

The monitoring model of vegetation change in grassland desertification ecological restoration based on deep learning is constructed to improve the monitoring accuracy of vegetation in grassland desertification area, and the area of vegetation increase or decrease can be clearly marked. By comparing the output results of the analysis model, the kNDVI-PT-Former model based on kNDVI constructed in this study performed well in the monitoring of vegetation changes in grassland desertification ecological restoration, and showed excellent performance on the test set. The CPA was 0.7295 and the IoU was 0.7228. At the same time, the field verification of the ecological restoration area was carried out, indicating that the monitoring effect of the model was the best, and the overall monitoring accuracy of the model was improved by 11%. Furthermore, the stability of the model’s performance was confirmed through evaluation with five-fold cross-validation.

The kNDVI-PT-Former model developed in this study enables large-scale, rapid, and more accurate monitoring of vegetation changes in grassland desertification restoration areas, providing technical support for ecological restoration monitoring and assessment of sandy grasslands while offering decision-making references for desertification control and sustainable grassland development.