Identification of Ecological Restoration Approaches and Effects Based on the OO-CCDC Algorithm in an Ecologically Fragile Region

Abstract

A full understanding of the patterns, trends, and strategies for long-term ecosystem changes helps decision-makers evaluate the effectiveness of ecological restoration projects. This study identified the ecological restoration approaches on planted forest, natural forest, and natural grassland protection during 2000–2022 based on a developed object-oriented continuous change detection and classification (OO-CCDC) method. Taking the Loess hilly region in the southern Ningxia Hui Autonomous Region, China as a case study, we assessed the ecological effects after protecting forest or grassland automatically and continuously by highlighting the location and change time of positive or negative effects. The results showed that the accuracy of ecological restoration approaches extraction was 90.73%, and the accuracies of the ecological restoration effects were 86.1% in time and 84.4% in space. A detailed evaluation from 2000 to 2022 demonstrated that positive effects peaked in 2013 (1262.69 km²), while the highest negative effects were observed in 2017 (54.54 km²). In total, 94.39% of the planted forests, 99.56% of the natural forest protection, and 62.36% of the grassland protection were in a stable pattern, and 35.37% of the natural grassland displayed positive effects, indicating a proactive role for forest management and ecological restoration in an ecologically fragile region. The negative effects accounted for a small proportion, only 2.41% of the planted forests concentrated in Pengyang County and 2.62% of the natural grassland protection mainly distributed around the farmland in the central-eastern part of the study area. By highlighting regions with positive effects as acceptable references and regions with negative effects as essential conservation objects, this study provides valuable insights for evaluating the effectiveness of the integrated ecological restoration pattern and determining the configuration of ecological restoration measures.

1. Introduction

The restoration and reconstruction of ecosystems are the main approaches to improving ecosystem services and achieving sustainable development goals [1,2]. A series of ecological programs have been launched in the Loess Plateau region since the 1980s, including artificial and natural restoration programs, such as the Three Northern Protected Forest Program (1978), the Return of Cropland to Forest Program (1999), and the Natural Forest Protection Project (2000) [3,4,5]. The ecological restoration approaches employed in these projects typically include large-scale afforestation, returning cultivated land to forests or grasslands, prohibiting grazing, and creating grassland enclosures. The Chinese government [6,7] and other studies [8,9] have reported that these programs have significantly improved the ecological environment. The ecological impact of separate approaches and the composite pattern of multiple approaches on the ecological environment are largely unknown. Continuous spatiotemporal monitoring is essential to identify different ecological restoration approaches and their impacts. If land surface processes can be measured as continuous fields along the spatial and temporal scales, it would be useful to address some fundamental questions on where and how various ecological restoration approaches are carried out. This is of prime importance to evaluate the ecological services and the effectiveness of ecological restoration measures in fragile areas, which provide the strategic goals of managing plantation forests.

With the development of remote sensing technology, satellite imagery has become an important data resource for analyzing the effects of ecological restoration [10]. Generally, the strategy of evaluation of ecological restoration effects is based on the classification algorithm [9]. Among those, the extreme gradient boost (XGBoost) algorithm and dominant species–physiognomy–ecological (DSPE) system work well [11,12]. The XGBoost algorithm successfully mapped the spatio-temporal distribution of ecological function in a terrestrial landscape using a set of classifiers derived from EO Sentinel-2 satellite image mosaics, whereas the DSPE system classified the ecological communities from remote sensing images. The Landsat time series analysis has been developed in ecological restoration effect assessments recently. Landsat time series change detection algorithms are designed to characterize, classify, and detect the vegetation change. The Landsat-based detection of trends in disturbance and recovery [13], and continuous change detection and classification (CCDC) [14], is widely used. Existing methods assessed the overall quality and functionality of the entire regional ecosystem [15]; however, the evaluation of ecological restoration effects is not linked to specific ecological restoration approaches and does not distinguish between natural recovery and human-induced restoration [9].

There is a challenge to identifying ecological restoration approaches and assessing their effectiveness. Current change detection algorithms can accurately detect the time of changes and land cover types before and after changes. Some current studies have achieved ecological restoration assessment by considering the pixel as the basic unit in change detection algorithms [16]. However, the distribution of geographical features is spatially continuous, and pixel-based analyses ignore spatial connections and context information [17]. In addition, severe fragmentation in the Loess Plateau has led to the coexistence of small patches of different types of ecosystems within a larger patch of a specific ecosystem, such as a few trees or cultivated land scattered over a large grassland [18]. The diversity of topography leads to high landscape heterogeneity in the region [19], presenting difficulties in identifying ecosystem restoration patterns. Therefore, developing an object-oriented time-series change detection method is urgently required to identify the long-term characteristics of ecosystem restoration patterns at the landscape scale in ecologically fragile areas. Most studies have analyzed the effects of ecological restoration at a regional scale based on time-series analysis methods, whereas less attention has focused on ecological restoration approaches and the effects of separate ecological restoration approaches on ecological restoration. Evaluating the effects of multiple ecological restoration approaches can help in decision-making for composite ecological restoration patterns.

Therefore, in this study, taking a typical Loess hilly region, namely, the southern region of Ningxia Hui Autonomous Region, China, as a case study, we identified ecosystem restoration approaches and evaluated their effects. Furthermore, this study aimed to provide an insightful view (where and how) of ecological restoration approaches in a typical Loess hilly region by proposing a novel object-oriented continuous change detection (OO-CCDC) method. Specifically, we aimed (1) to develop an automatic methodology for identifying multiple ecological restoration approaches and (2) to evaluate the ecosystem restoration approaches through continuous detection.

2. Study Area

The semiarid and subhumid loess hilly ecoregion (35°11′–36°31′N, 105°19′–106°57′E) in southern Ningxia is located in the northwest Loess Plateau, with an area of 15,526 km². As one of the three ecological functional zones in Ningxia (Figure 1), this study area has high ecological vulnerability and sensitivity in a long-term period [20]. The study area included Liupan Mountain and Qingshui Valley, which are 1500–2200 m above sea level. The average temperature ranges from 6.7 °C to 8.8 °C, the precipitation ranges from 458.6 mm to 668.2 mm, and the sunshine duration ranges from 2056.9–2384.4 h [21,22]. The soils are mainly black loam and yellow loose loam. Owing to the influence of population growth and historical wars, large-scale deforestation and land reclamation occurred, seriously damaging the regional ecological environment. Since 2000, the Chinese government has implemented a series of ecological restoration policies and measures to address the serious environmental problems in the region [22,23]. The ecological restoration of vegetation in this region is divided into three types: planted forest, natural forest protection, and grassland protection.

3. Materials and Methods

In this study, we proposed an OO-CCDC algorithm that integrates image segmentation and a change detection algorithm to continuously and automatically highlight the effectiveness of different ecological restoration approaches, providing an insightful view of the vegetation change process in fragile areas. The general workflow is shown in Figure 2. First, we proceeded with Landsat composite images from 2000 to 2022 and land use classification in 2022. Second, multi-scale segmentation was performed to generate homogeneous objects using the Landsat time series. The normalized difference vegetation index (NDVI) and enhanced vegetation index (EVI) were calculated, and the median of all pixels in the same object was assigned to the object to reconstruct an object-based time-series dataset. The OO-CCDC algorithm was used to detect changes in the three ecological restoration approaches. The detection accuracies were evaluated at space and time levels. Finally, the map products of the OO-CCDC were used to analyze trends in vegetation using different ecological restoration approaches.

3.1. Data Preprocessing

3.1.1. Landsat Time Series Construction

All available Landsat-7 and Landsat-8 Level-1 products with cloud cover of less than 20% between 2000 and 2022 were synthesized and processed on the GEE platform. The number of images varied from 68 to 263 each year, with 3720 scene data points (Figure 3). In addition, the Fmask algorithm was used to mask shadows, clouds, and snow in the Landsat image [24]. To maintain the consistency of different sensors, the ETM+ and OLI data were uniformly harmonized on the GEE platform [25] (https://developers.google.com/earth-engine/tutorials/community/landsat-etm-to-oli-harmonization, accessed on 23 February 2023). Seasonal composite images for synthesis were carried out in spring (March, April, and May); summer (June, July, and August); autumn (September, October, and November); and winter (December, January, and February). The median synthesis method was employed to generate four images per year. This approach provides a robust way to mitigate the potential impact of outliers or extreme values that might distort the synthesized seasonal data [26]. Moreover, the method is widely utilized for processing and synthesizing remote sensing data [27].

3.1.2. Reference Sample Collection

Reference Sample of Ecological Restoration Approaches

We conducted a field survey of ecological restoration approaches in the study area from June to August 2022 (

Table 1. The numbers and sources of reference pixels for each ecological restoration approach.

Ecological Restoration Approaches	Field Survey Data	Image Data	Total Number
Natural forest protection	623	911	1534
Planted forests restoration	687	641	1328
Natural grassland protection	532	498	1030
Others	446	2063	2509
Total number	2288	4113	6401

), including natural forest protection, natural grassland protection, and planted forests. In total, 2288 samples were collected, comprising 623 natural forest protection samples, 687 planted forest samples, 532 natural grassland protection samples, and 446 others.

To address the challenge of distinguishing between different ecological restoration approaches using the acquired Landsat and high-resolution images, we utilized both GF-2 satellite data and field investigations. A subset of 6401 samples was selected for model training and accuracy assessment, representing 80% of the total samples. Table 1 provides further details on the selection process.

Reference Sample of Ecological Restoration Detection

The sampling points are shown in Figure 4 and

Table 2. The numbers and sources of sampling points for ecological restoration detection.

Ecological Restoration Approaches	Changed Pixels	Unchanged Pixels	Total Number
Natural forest protection	21	39	60
Planted forest	112	103	215
Grassland protection	117	108	225
Total Number	250	250	500

. Field surveys were conducted in restoration areas where ecological restoration projects had been implemented in the eastern, central-southern, and northern regions of the study area from 2007 to 2018 (the ecological restoration project data were provided by the Ningxia Soil and Water Conservation Monitoring Station). Land use/cover type, primary vegetation cover, and ecological restoration approaches were recorded for each site. Owing to limited field surveys in recent years, the interpretation of vegetation change processes was conducted directly from Google Earth and the GF2 and GF1 satellites through visual interpretation. A total of 500 pixels reference sites were collected, consisting of 250 samples from the degraded and recovered regions as well as an additional 250 samples that showed no change. Among them, we collected 21 samples from natural forests, 112 samples from planted forests, and 117 samples from natural grasslands in the degraded and restored areas and selected sample points inside and outside the boundaries of the object patch, with an average patch size normally larger than 6000 m². A total of 150 object reference sites were collected, consisting of 100 samples from the degraded and recovered regions as well as an additional 50 samples that showed no change.

3.2. Classification of Ecological Restoration Approaches

3.2.1. Feature Extraction

The objects of natural forests, planted forests, and natural grasslands were extracted as three different ecological restoration approaches. We extracted a range of diverse features, including spectrum, phenology, texture, and topography, to facilitate the identification of ecological restoration approaches. The normalized difference vegetation index (NDVI) is a commonly used spectral index to reflect vegetation status [28]. This index was calculated by combining visible and near-infrared light reflected in Landsat imagery captured in 2022 [29,30]. The phenological index, known as the normalized differential phenology index (NDPI), relies on the spectral contrast between vegetation and soil or snow. This significantly enhances the monitoring capacity for spring phenology in regions or seasons affected by snow contamination [31]. To obtain texture features, we generated a gray-level co-occurrence matrix (GLCM) by calculating the joint probability distribution of pairs of gray levels derived from median NDVI imagery for 2022 [32,33]. For the terrain features, we utilized digital elevation models (DEMs) obtained from the NASA Jet Propulsion Laboratory (JPL) at a resolution of approximately 1 arcsec. Calculations for aspect, elevation, and slope were performed, and the data were subsequently resampled to an approximate resolution of 30 m [21].

3.2.2. Classification

We extracted three main ecological restoration approaches (natural forests, planted forests, and natural grasslands) by the land use map of 2022 on the GEE platform. The random forest (RF) classifier, as an ensemble classifier, constructs multiple decision trees by utilizing randomly selected subsets of training samples and variables. This approach has led to outstanding classification performance and efficient processing speed [34]. Therefore, we opted for the random forest classifier to extract the ecological restoration approaches. All of the extracted features were utilized as inputs for the RF models to classify ecological restoration approaches. The parameters of the RF model are shown in

Table 3. Random forest (RF) models parameters.

Parameters	Description	Value
Number of Trees	The number of decision trees created	100
Variables Per Split	The number of variables per split	null
Min Leaf Population	The minimum size of a terminal node	1
Bag Fraction	The fraction of input to bag per tree	0.5
Out of Bag Mode	Whether the classifier should run in out-of-bag mode	false
Seed	Random seed	0

(https://code.earthengine.google.com/3570d44f93c0324c6148cf34fd7fe621, accessed on 3 August 2023).

3.3. Estimating Ecological Restoration Based on Object-Oriented Continuous Change Detection

3.3.1. Multiscale Segmentation

Image segmentation is the first step in performing object-oriented change detection. The multi-scale segmentation algorithm effectively solves salt and pepper noise [35,36]. The images are clustered from bottom to top at different scales according to spectral similarity and shape factor to form homogeneous image objects within the scale range. By clustering adjacent pixels or tiny segmentation objects, image segmentation is realized based on region-merging technology to ensure maximum intra-object homogeneity and minimum inter-object homogeneity [37,38,39,40]. The homogeneity criterion (H^l) in the multi-scale segmentation algorithm is defined based on spatial and spectral attributes at different scales [41].

$${\mathrm{H}}^{\mathrm{l}}\left({\mathrm{O}}_{\mathrm{a}}^{\mathrm{l}}\right)={\mathsf{\omega }}_{\mathrm{sp}}^{\mathrm{l}}°{\mathrm{h}}_{\mathrm{sp}}^{\mathrm{l}}\left({\mathrm{O}}_{\mathrm{a}}^{\mathrm{l}}\right)+{\mathsf{\omega }}_{\mathrm{sh}}^{\mathrm{l}}°{\mathrm{h}}_{\mathrm{sp}}^{\mathrm{l}}\left({\mathrm{O}}_{\mathrm{a}}^{\mathrm{l}}\right)$$

(1)

where a is the input image, l is the scale, and the spectral weight $\left({\mathsf{\omega }}_{\mathrm{sp}}^{\mathrm{l}}\right)$ and the shape weight (${\mathsf{\omega }}_{\mathrm{sh}}^{\mathrm{l}}$) can be defined as ${\mathsf{\omega }}_{\mathrm{sp}}^{\mathrm{l}}$ = 1 − ${\mathsf{\omega }}_{\mathrm{sh}}^{\mathrm{l}}$.

We used eCognition to calculate local variance (LV) and the rate of change of local variance (ROC-LV) for determining optimal scale. When a peak appears in the ROC-LV graphs, it is considered as the optimal scale [42]. We calculated LV and ROC-LV at possible suitable scales from 100 to 1000 (

Table A1. Accuracy assessment for the ecological restoration approaches classification in 2022.

	Natural Forests	Planted Forests	Natural Grassland	Others
User accuracy (%)	97.94	86.07	88.64	90.14	Overall accuracy = 90.73%
Producer accuracy (%)	97.94	86.07	88.65	89.24	Kappa = 88.47%

and Figure A1). Compared with the multi-scale segmentation test, we comprehensively considered the image segmentation effect and the number of objects to determine the optimal segmentation scale as 500 and finally complete the multi-scale segmentation. The candidate segmentation parameters of different scales are shown in

Table 4. Comparison of multi-scale segmentation results.

Segmentation Scale	Number of Objects	Description	Splitting Effect
300	270,962	The image is too fragmented, the edge of the cultivated land is blurred, and the segmented object is too small
500	215,414	The division is relatively complete, the details are well preserved, and each object is relatively independent from the others
800	175,420	Partitioned objects are too large, and some objects are divided together

. After achieving the segmentation process using the eCognition software, we uploaded the segmentation results on the GEE platform.

3.3.2. Object-Level Feature Sets Generation

NDVI and EVI [43] describe the growth status of vegetation [44] and have been widely used in ecological change studies. We reconstructed the NDVI and EVI seasonal time series at the object level. The segmented unit was calculated using the median values of the NDVI (M_NDVI) and EVI (M_EVI) for all pixels within each object as the median value of discrete data can better filter out the effects of noise or outliers [27,45].

M_NDVI/M_EVI = median {index₁, index₂, …, index_m}|ol

(2)

where index represents NDVI or EVI, m represents the number of pixel within each object, and ol is the optimal scale (500) based on Equation (1).

We selected three locations—natural forest, planted forest, and natural grass—as examples to show the change during 2000–2022. The Landsat composite images of the locations are shown in Figure 5a, where the vegetation cover of natural forests was high at all times, whereas the planted forest and natural grass changed over time. The NDVI and EVI time series at the object level obtained by image segmentation are displayed in Figure 5b,c, respectively, and reflect the cover change of the three vegetation types.

3.3.3. Characterizing Vegetation Change Processes under Multiple Ecological Restoration Approaches

Using different ecological restoration approaches, the reconstructed object-level MNDVI and MEVI time series were employed to implement the CCDC algorithm for vegetation restoration detection on the GEE platform. The CCDC algorithm used all available observations in the time series [14]. Land-surface changes can be decomposed into three elements: (1) interannual (gradual) changes, (2) intra-annual (seasonal) changes, and (3) abrupt changes (breakpoints). Using the segmented objects as a detection unit, the proposed OO-CCDC algorithm can effectively detect restoration and avoid missed detection or over detection as long as any trend and amplitude of the NDVI or EVI within an object changes. The formula for the OO-CCDC algorithm is as follows:

$${\mathsf{\rho }}_{\left(\mathrm{ix}\right)}={\mathrm{a}}_{1,\mathrm{i}}\cos \left(\mathrm{x}2\mathsf{\pi }/\mathrm{T}\right)+{\mathrm{b}}_{1,\mathrm{i}}\sin \left(\mathrm{x}2\mathsf{\pi }/\mathrm{T}\right)+{\mathrm{c}}_{1,\mathrm{i}}\mathrm{x}+{\mathrm{a}}_{0,\mathrm{i}}$$

(3)

where $\mathrm{i}$ represents the M_NDVI or M_EVI time series calculated by Equation (2); $\mathrm{x}$ represents the first day of the year and the range of values is $1-365;{\mathrm{a}}_{1,\mathrm{i}}$ and ${\mathrm{b}}_{1,\mathrm{i}}$ represent the intra-annual (seasonal) coefficient; ${\mathrm{c}}_{1,\mathrm{i}}$ represents the inter-annual (gradual) coefficient; ${\mathrm{a}}_{0,\mathrm{i}}$ is the error value of the fit; and $\mathrm{T}$ is equal to 365.

The OO-CCDC algorithm can identify all breakpoints within a specified time series and obtain information on breakpoint times and change trends (https://gee-ccdc-tools.readthedocs.io/en/latest/lctutorial/change.html, accessed on 3 August 2023 ).

The effects of different ecological restoration methods on vegetation restoration were assessed using trend analysis. Three types of recovery trends were defined for the proposed method: growth, stability, and decrease. The effectiveness of ecological restoration was determined based on these trends. Therefore, we classified nine ecological restoration effects: positive effects/recovery (increasing), positive effects (stable), and negative effects/degradation (decreasing) (

Table 5. Classification criteria for ecological restoration effects.

Ecological Restoration Approaches	Trend	Effects
Natural forest protection	increasing	Positive
	stable	Positive
	decreasing	Negative
Plant forest	increasing	Positive
	stable	Positive
	decreasing	Negative
Natural grassland protection	increasing	Positive
	stable	Positive
	decreasing	Negative

, Figure 6). According to the detection results of OO-CCDC, we obtained the latest trend using the gradual coefficient (c_1,i) in Equation (3). A decreasing trend in the MNDVI and MEVI time series indicated a decreasing pattern, and vice versa.

According to the detection results of OO-CCDC, we obtained the latest trend using the gradual coefficient (c_1,i) in Equation (3). A decreasing trend in the M_NDVI and M_EVI time series indicated a decreasing pattern, and vice versa. We defined a stable pattern as no breakpoints detected throughout the study period. After determining whether an object was decreasing, increasing, or stable, we overlaid the ecological restoration approaches of the study area in 2022 to determine the restoration patterns of the ecological restoration approaches.

3.3.4. Accuracy Assessment

We evaluated the spatial and temporal accuracy of change detection by the OO-CCDC and CCDC algorithms, respectively. The total 500 pixel-based samples (list in Table 2) were employed for verification at the spatial domain. For each pixel, we visually inspected all Landsat images, high spatial resolution images from Google Earth, and GF 1-2 images to determine and record whether change events had occurred or not. Then, a confusion matrix and an assessment of the overall accuracy, user accuracy, and producer accuracy of the change location by OO-CCDC and CCDC algorithms were produced, respectively, following commonly used methods [46,47]. After assessing the spatial accuracy of the change detection, the temporal accuracy was evaluated for pixels that were spatially identified. For each pixel, we visually interpreted Landsat images, high spatial resolution images from Google Earth, observations of index trajectories, and GF1 and GF2 images to help record the change times for the specific location. Owing to the limitations of the high-resolution images, the exact time of the restoration change was difficult to determine [48]. Therefore, a time delay between the land cover change and the actual occurrence of a restoration commonly exists. Here, the error distribution of the identification of change time for CCDC and OO-CCDC algorithms was calculated following Equation (4):

$$\mathrm{error} =\left|\mathrm{actual} \mathrm{breakpoint} \mathrm{time} - \mathrm{algorithm} \mathrm{identification} \mathrm{of} \mathrm{breakpoint} \mathrm{time}\right|$$

(4)

4. Results

4.1. Accuracy Assessment

The overall accuracy of extraction the classification of ecological restoration approaches results of natural forest, plant forest, and natural grassland for the three different approaches to ecological restoration was 88.73%, and the kappa coefficient was 0.86 (Table A1).

The overall accuracy of the pixel-based evaluation by OO-CCDC algorithm was 84.4%, and the object-based was 89.3%, whereas the overall accuracy by CCDC algorithm was 74.5% (

Table 6. Accuracy assessment for the CCDC according to pixel-based samples.

Category	Real Category/Pixel	Total Number of Pixels	--	User Accuracy/%
--	Changed pixels	Unchanged pixels	Total number of pixels	--
Changed pixels	233	17	250	93.2
Unchanged pixels	115	135	250	54.0
Total number of pixels	348	152	500	--
Producer accuracy/%	66.9	88.8	--	Overall accuracy = 74.5%

,

Table 7. Accuracy assessment for the OO-CCDC according to pixel-based samples.

Category	--	Real Category/pc	--	User Accuracy/%
--	Changed pixels	Unchanged pixels	Total number of pixels	--
Changed pixels	225	25	250	90.0
Unchanged pixels	53	197	250	78.8
Total number of pixels	278	222	500	--
Producer accuracy/%	80.9	88.7	--	Overall accuracy = 84.4%

and

Table 8. Accuracy Assessment for OO-CCDC according to object-based samples.

Category	--	Real Category/pc	--	User Accuracy/%
--	Changed objects	Unchanged objects	Total number of objects	--
Changed objects	89	11	100	89
Unchanged objects	5	45	50	90
Total number of objects	94	56	150	--
Producer accuracy/%	95	80	--	Overall accuracy = 89.3%

). The producer and user accuracies both of the pixel-based and the object-based evaluation by OO-CCDC algorithm were more than 80%, except the user accuracy of the unchanged object was slightly lower (78.8%).

The assessment accuracies of the detection time by CCDC and OO-CCDC algorithms are shown in Figure 7. The overall accuracies of the two algorithms were 78.5% and 86.1%, respectively. We also recorded the error distribution of the detection times for the CCDC and OO-CCDC algorithms. Accurately identified objects accounted for 88.1% of the OO-CCDC algorithm. The error ratios within one year of the OO-CCDC and CCDC algorithms in identifying the occurrence time were 94% and 90%, respectively.

Five typical regions were further studied to compare the results of the two algorithms to better analyze the restoration patterns (Figure 8). According to the latest Google Earth and GF1-2 images, regions “a” and “e” are cropland, and regions “b” and “d” represent Liupanshan Airport and agricultural greenhouses, respectively. The boundary of those surface covers is accurately identified by the OO-CCDC algorithm and partly detected by CCDC. Region “c” was characterized as the restoration pattern of planted forest. The identification by OO-CCDC was consistent with the distribution of ground objects since the segmentation in OO-CCDC improved the internal homogeneity of the detection objects, whereas the detection results of CCDC were not accurate and were significantly more fragmented.

4.2. Spatial Distribution of Multiple Ecological Restoration Approaches

Natural forest protection was mainly concentrated in the Xihua Mountains, Nanhua Mountains, Moon Mountains, and Liupanshan Nature Reserve. The planted forest was mainly distributed in the Loess hilly and gully areas in the east and west of the study area and around the natural forest in the southern region. Natural grassland protection areas were mainly distributed in the arid grasslands of the Loess Hills in the central and northern parts of the study area (Figure 9a).

As shown in Figure 9b, the area of the three ecological restoration approaches was 116,01.09 km², accounting for 60.34% of the study area, which is affected by the arid and semi-arid climate, with natural grassland protection occupying the largest area (5425.10 km²; 28.22%), followed by planted forest (4386.22 km²; 22.81%) and natural forest protection being scarce (1789.77 km²; 9.31%).

4.3. The Ecological Restoration Effect for Multiple Ecological Restoration Approaches

From 2000 to 2022, the ecological restoration effects in the study area were presented in a stable pattern (9110.17 km²; 47.39%), followed by positive effects (2065.52 km²; 10.74%). Only 1.20% of the study area exhibited negative effects (230.35 km²). The area of stable effects in the study region of planted forest occupied the largest proportion (3952.30 km²; 20.56%), followed by natural grassland protection (3376.28 km²; 17.56%) and natural forest protection (1781.59 km²; 9.27%). In addition, negative effects were detected in 0.1% of natural grassland, 2.62% of planted forest, and 2.41% of natural forest (1.80 km²). The positive effects during the study period included 1919.05 km² of natural grassland, 140.37 km² of planted forest, and 6.10 km² of natural forest. In the northern hilly areas of the study area, the natural grassland area increased by 35.38% (Figure 10).

In terms of spatial distribution, for the afforestation approach, 94.39% of the planted forests had stable effects, which were concentrated around the Liupanshan Nature Reserve in the central-eastern and southern parts of the study area and scattered in the arid areas in the central and western parts (Figure 11). In the southeastern part of the study area in Pengyang County, 3.20% of the planted forests displayed positive effects. In contrast, the negative effects of planted forests accounted for 2.41% and were concentrated in the southeastern part of the study area in Pengyang County, with degraded woodlands mainly distributed around newly cultivated terraces.

For natural forest protection, 99.56% of natural forests had stable effects. The positive effects of natural forest accounted for only 0.34%, scattered on the southern edge of the Nanhuashan Nature Reserve; the negative effects of natural forest accounted for only 0.10%, mainly scattered on both sides of the road inside the Liupanshan Nature Reserve; and for natural grassland protection, 62.36% of the grassland area stabilized, mainly in the central and northern hilly areas of the study area. In addition, 35.37% of the grasslands showed positive effects, mainly in the northern part of the study area, and the negative effects of natural grassland only accounted for 2.26% and were mainly distributed around the farmland in the central-eastern part of the study area.

For temporal (Figure 12), the area with positive effects reached the maximum value in 2013, resulting in 1262.69 km², followed by 2019 and 2018, which were 359.49.32 and 223.17 km^2, respectively. Conversely, the maximum negative effects occurred in 2017 (54.54 km²), followed by 2016 (45.72 km²) and 2018 (41.71 km²). For natural forest protection, the positive effects remained consistent, with an average annual recovery area of 0.27 km². The largest occurrence of positive effects was observed in 2013, covering an area of 2.96 km², followed by 2003, with an area of 1.37 km². However, the largest negative effects were recorded in 2003, with an area of 0.40 km². Regarding natural grassland protection, there was a gradual increase in positive effects in the area from 2000 to 2003, reaching its peak in 2013, covering an area of 84.29 km². However, a notable increase in negative effects was observed from 2016 to 2018, with the highest impact occurring in 2018 and affecting an area of 29.09 km². For planted forests, there was a slight increase in the area from 2000 to 2003, reaching its peak in 2013 at 84.29 km². However, the negative effects on planted forests significantly increased from 2016 to 2018. The largest negative effects were concentrated during this period, with the maximum effect occurring in 2018, covering an area of 29.09 km².

5. Discussion

5.1. Efficient Framework for Assessing Effectiveness of Multiple Ecological Restoration Approaches

Ecological restoration is important in China’s ecological protection [22,44,48]. It serves as a primary means to maintain biodiversity, improve ecosystem services, and ensure ecosystem stability [49,50]. From the new perspective of ecological civilization construction, the simultaneous detection of the effects of various ecological restoration approaches is important for formulating regional ecological restoration policies, promoting comprehensive land consolidation and environmental protection, and constructing new patterns of territorial space development and protection [47]. For different restoration approaches, planted forests are thought to have lower diversity, productivity, and ecological stability compared with natural forests during the long-term recovery process [51,52]. Especially, a significant distinction of ecosystem service existed in the process of natural forest loss and plantation expansion [53]. Therefore, it is necessary to analyze the multiple effects of ecological restoration [54]. The evaluation of ecological restoration effectiveness could be divided into two aspects. On one hand, the strategy of evaluation of multiple ecological restoration effects is based on the land use classification in interval of several years (i.e., every five years). After that, the land use transition matrix was widely used to assess changes in forest and, consequently, evaluate ecosystem quality and service capacity. However, the change process of land use types is ignored [55]. On the other hand, time-series approaches have recently increased research interest in landscape dynamics analysis and ecological restoration effect assessments. There is an increasing need to develop efficient automated/semi-automated methods for detecting the ecological effects of ecological restoration approaches. Existing methods that deal with time-series datasets are usually pixel based and cannot solve the high landscape heterogeneity in the ecological restoration region [56]. Most important, the specific ecological restoration approaches were not considered when evaluating the ecological effects, especially when failing to distinguish natural forests protection and planted forests restoration [9]. In this study, we developed a flexible framework for continuously characterizing ecological restoration effects by combining multi-scale segmentation with the CCDC algorithm with high spatial and temporal accuracy. We specifically targeted the approaches of ecological restoration, extracting natural recovery and artificial restoration, thereby enabling a precise evaluation of the ecological restoration effects of various approaches. This holds significant importance for designing ecological restoration engineering measures effectively.

5.2. Comprehensive Evaluations of the OO-CCDC Algorithm

Theoretically, owing to the severe fragmentation and high heterogeneity of patches in the Loess Plateau, pixel-based detection algorithms may not accurately analyze the restoration patterns of natural forest protection, natural grassland protection, and planted forests [57]. This indicates that the OO-CCDC algorithm proposed in this study can be employed to identify ecological restoration patterns in areas with small inconsistent patches. Specifically, after optimal segmentation in the OO-CCDC algorithm, the internal homogeneity of each object and the heterogeneity between objects were solved using the segmentation quality evaluation function [58,59]. The optimal segmentation scale layer improves the detection accuracy of vegetation restoration.

From the algorithm recognition results, the recovery areas identified by the OO-CCDC and CCDC algorithms were more northern than southern in terms of spatial distribution, and the disturbance was mainly distributed in the eastern part of the study area. Most of the recovery and perturbation occurred after 2010; however, the region recognized by the CCDC was significantly larger than that recognized by the OO-CCDC. This is because the pixels that meet the breakpoint detection conditions of CCDC are recognized by the algorithm, whereas OO-CCDC requires that the median pixels in an object meet the breakpoint detection conditions to be recognized. Therefore, OO-CCDC avoids over detection caused by outliers, and from the results of ecological restoration pattern identification, the object-based segmentation results can better maintain data integrity. In the actual restoration process, grasslands and forests are spatially continuous, and this method can effectively reduce the mutual interference between different vegetation types.

5.3. Spatiotemporal Patterns of Ecological Restoration Effects

The results of the ecological restoration patterns showed that the large-scale restoration of planted forests and natural forest protection occurred in 2003 and was closely related to the implementation of the policy. The study area was identified as a demonstration area for returning farmland to forests and grasslands in 2000 and was formally launched in 2002, insisting on the synergy between migration and ecological restoration during the implementation process [60]. During the implementation process, migrants were relocated in conjunction with ecological restoration [60]. Ecological migrants were used to reforest the relocated land, cultivate and maintain it, and restore woodland and grassland vegetation according to local conditions. Under the policy of returning farmland to forest, the planted forest area has increased significantly [61], which is consistent with other studies [21,48].

From the perspective of the extent and area of restoration, natural grassland protection was the main ecological restoration mode in the study area. Researchers believe the best method for restoring large grassland areas is ‘natural restoration’, which relies entirely on natural succession to restore decreased natural grassland protection without human intervention [57]. The large-scale grassland protection restoration in 2012 was due to the promulgation of “the Regulations on the Prohibition of Grazing and Fencing in the Ningxia Hui Autonomous Region” in 2011, which was the first local regulation in China to ban the grazing and fencing of grasslands [48]. In response to this policy, Ningxia implemented a series of ecological restoration projects, such as a ban on grazing on natural grasslands and ecological migration [59]. Consequently, the large-scale grassland restoration in 2012 proved that natural grassland protection restoration achieved remarkable results. Overall, the ecological restoration patterns showed a steady improvement in planted forests, natural grassland protection coverage, and the optimization of the ecological environment. The proposed method utilizes Landsat time series data to identify large-scale vegetation restoration and degradation in a long-term and automatic manner. This algorithm captures the type of multiple ecological restoration approaches. However, this method had several limitations. Specifically: (1) In arid and semi-arid areas with low original vegetation coverage, the temporal trajectory of some pixels undergoing ecological damage or restoration is not significantly changed, leading to the omission of decreasing pattern identification [62]. (2) This study focused on identifying natural or artificial ecological restoration approaches, and attention to detailed plantation species will better guide forestation planning and design regarding the selection of optimal restoration patterns [6]. Additional data sources and improved methods should be developed in future studies [63].

6. Conclusions

In this study, we evaluated the ecological effectiveness of multiple restoration approaches by developing an OO-CCDC method. We detected and separated multiple ecological restoration effects (positive, negative) of three ecological restoration approaches, including natural forest protection, planted forest restoration, and natural grassland protection. Through the application of the OO-CCDC method to the Loess hilly region of southern Ningxia during 2000–2022, the results showed that: (1) the accuracy of ecological restoration approaches extraction was 88.73%, and the accuracies of ecological restoration effects were 86.4% in time and 84.4% in space based on the OO-CCDC method. (2) The detailed evaluation from 2000 to 2022 demonstrated that positive effects peaked in 2013 (1262.69 km²), while the highest negative effects were observed in 2017 (54.54 km²). (3) In total, 94.39% of the planted forests, 99.56% of the natural forest protection, and 62.36% of the grassland protection had a stable effect, and 35.37% of the natural grassland protection, 3.20% of the planted forests, and 0.34% of the natural forest protection displayed positive effects, indicating a proactive role for forest management and ecological restoration in an ecologically fragile region. (4) Only 2.62% of the grassland protection, 2.41% of the planted forests, and 0.1% of the natural forest protection had negative effects. The planted forests and the natural grassland protection with negative effects mainly concentrated in Pengyang County and around the farmland in the central-eastern part of the study area, respectively. By highlighting regions with positive effects as acceptable reference and regions with negative effects as essential conservation objects, this study holds significant importance for designing ecological restoration engineering measures effectively.