Spectral Characteristics and Identification of Degraded Alpine Meadow in Qinghai–Tibetan Plateau Based on Hyperspectral Data

Abstract

Grassland degradation poses a significant challenge to achieving the Sustainable Development Goals (SDGs) on the Qinghai–Tibetan Plateau (QTP). Effective monitoring of grassland degradation is essential for ecological restoration. Hyperspectral technology offers efficient and accurate identification of degradation. However, the influence of observation time, data analysis methods and classification techniques on the accuracy of identifying alpine grasslands remains unclear. In this study, the spectral reflectance of degraded alpine meadow, alpine meadow, alpine shrub and Tibetan barley was measured from May to September 2023 using a ground spectrometer in the northeastern QTP. First-order derivatives (FDR) and continuum removal were applied to the spectra, and characteristic parameters and vegetation indices were calculated. Support vector machine (SVM), random forest (RF), artificial neural network (ANN) and decision tree (DT) were then used to compare the classification accuracy between different months, transformation methods and characteristic parameters. The results showed that the spectral reflectance peaked in July, with significant differences in the near infrared (NIR) bands between alpine meadow and degraded alpine meadow. Alpine shrub and Tibetan barley showed greater differences in reflectance compared to other vegetation types, especially in the NIR bands. Data transformations improved reflectance and absorption characteristics in the NIR and visible bands. Indices such as DVI, RVI and NDGI effectively differentiated vegetation types. Optimal accuracy for the identification of degraded alpine meadow in July was achieved using FDR transformations and ANN or SVM for classification. This study provides methodological insights for monitoring grassland degradation on the QTP.

1. Introduction

Grasslands are an ecosystem type that occupies 40% of the Earth’s surface and spans several climatic zones and continents. They are a crucial source of production and sustenance for local populations [1]. In recent decades, however, grassland degradation has emerged as one of the most serious ecological and environmental problems facing humanity, with estimates suggesting that almost half of these areas are degraded to some degree [2,3]. This degradation not only leads to a range of environmental issues, such as biodiversity loss, land degradation, and increased soil erosion, but also affects the survival and development of hundreds of millions of people who depend on grasslands for food, fuel, firewood, and medicine [1]. The Qinghai–Tibetan Plateau in southwest China is the highest and largest plateau in the world. It covers an area of about 2.5 million km², of which about 60% is grassland. Grasslands on the QTP play a vital role in providing food sources for herbivores and a material base for 5 million herders, as well as maintaining the sustainable development of regional ecological services [4,5,6]. Grassland degradation has always been a major challenge for the QTP, especially in the context of global change, with uncertainties in the climate variations and hydrologic cycle of the plateau. Millions of herders face the dual impacts of traditional lifestyles and modernisation, which inevitably leads to a range of consequences including biodiversity loss, increased carbon emissions, and increased rates of soil nutrient loss and erosion [7,8,9]. Despite the implementation of various ecological restoration projects since 2000 and the continuous improvement of ecological protection policies and regulations in recent years, the overall trend of grassland degradation on the plateau has not been halted. Consequently, the degradation of alpine grasslands represents a significant challenge for achieving future sustainable development goals in human activity regions [10,11,12].

In light of the numerous adverse consequences associated with grassland degradation, a considerable body of research has been conducted by scientists to elucidate its underlying causes, assess its impact, and develop strategies for its restoration. Among these studies, the accurate identification of the degree and extent of grassland degradation represents a fundamental and pivotal step [13,14]. Methods for identifying grassland degradation typically include ground surveys and remote sensing inversion [15]. The ground survey method uses assessment criteria such as plant cover, community structure, and soil properties. However, this method is inefficient and unsuitable for large-scale investigations [16]. The rapid development of remote sensing technology provides efficient, and large-scale identification tools for identifying grassland degradation. Multi-spectral remote sensing imagery, which is particularly common in optical remote sensing, relies on information from the visible and near-infrared bands of plants. This technology can provide remote sensing indicators such as vegetation and biomass productivity indices, which represent the productivity and health of vegetation, thereby determining the degradation of regional grasslands [3]. The use of remote sensing to monitor grassland degradation at different spatial and temporal scales has made significant progress in different regions of the world. Remote sensing technology has enabled a deeper and more comprehensive understanding of grassland degradation. Key information, such as the area of degraded grassland in different regions, can provide an important basis for assessing changes in grassland ecosystem services, global change, and the impact of human activities on grasslands [16,17,18].

Nevertheless, the findings of grassland degradation derived from remote sensing have been subjected to scrutiny by some researchers and local communities due to discrepancies between these results and ground observations. It is inaccurate to suggest that remote sensing has replaced traditional, labour-intensive ground surveys. Indeed, the application of multispectral remote sensing observations to grassland degradation remains limited, probably for two main reasons. First, there are ambiguities in the definition of grassland degradation between different perspectives. Ecologists and herbologists define grassland degradation primarily in terms of plant community succession [19], while environmentalists, remote sensing scientists and the government focus on productivity, emphasising vegetation cover and above-ground biomass [20]. Second, the limitations of remote sensing observing instruments pose challenges. Multispectral remote sensing imagery generally has limited spectral resolution, which makes it difficult to address the phenomenon of heterospectrality of the same material [21]. As a result, it is difficult to effectively differentiate between grassland communities and species, particularly in alpine meadows with limited height, small leaf areas, and different species forming layers of tangled grass felt. This further reduces the accuracy of monitoring grassland degradation [18]. Therefore, despite its inefficiencies, ground truthing remains an important method for investigating grassland degradation in specific localities. There is a need for new tools and techniques that balance efficiency and accuracy

Hyperspectral data have large potential for the identification of different features because they use spectral characteristics. The primary difference between hyperspectral data and multispectral data lies in the ultra-high spectral identification range and resolution of hyperspectral data, which theoretically enables the identification of grassland species. This capability has the potential to resolve the current conflicts in efficiency, accuracy, and spatial extent of grassland degradation assessment, thereby facilitating an effective link between ground surveys and remote sensing observations [22]. The current methods employed in this field encompass the utilisation of ground spectrometers, unmanned aerial vehicles (UAVs) outfitted with hyperspectral sensors, and hyperspectral remote sensing imagery. These techniques are employed for the purpose of measuring and identifying grassland species, communities, cover, and biomass. These methods provide critical information on grassland degradation [23,24,25]. Grassland species identification is achieved by the direct spectral reflectance of different species using ground spectrometry, followed by analysis of the differences in spectral curves using methods such as data transformations, feature parameter calculations, and vegetation index calculations to select sensitive bands or features [25,26]. Hyperspectral imagery extends the identification results from point to area by using a priori knowledge of ground spectra. This is often achieved by building a library of feature spectra, which allows the monitoring of grassland species and degradation status over a large area [27].

However, there remain several shortcomings in the current research on the identification of grassland degradation using hyperspectral data. Firstly, there is an insufficient understanding of the seasonal change characteristics of alpine meadows, in particular the lack of comparative analysis of degraded alpine meadow identification results across different months. Secondly, although many data transformation and feature extraction methods exist for high-dimensional data, the differences between these transformations and features in the identification of degraded alpine meadows identification are not well understood. Finally, the differences in the accuracy of degraded alpine meadow identification between different classification methods are not comprehensively known. To fill these gaps, this study poses the following scientific question: do temporal variations, data transformation methods and classification methods affect the accuracy of degraded alpine grasslands identification? The aim of this study is to compare the accuracy of identifying different vegetation types in typical degraded alpine grasslands in different months, using different data transformations and feature parameters, and using different classification methods.

In this study, typical degraded alpine meadows and other vegetation types were selected on the QTP and continuously observed with a spectrometer during the growing seasons. The spectral characteristics and identification accuracy of typical vegetation were analysed and compared using different data transformations and characteristic parameters, as well as different classification methods. The results of this study can provide a reference for the identification of grassland degradation on the QTP.

2. Materials and Methods

2.1. Study Site

The study area is located on the southern slopes of the Qilian Mountains in the northeastern part of the QTP, at an average of 3320 m above sea level and with flat terrain characterized by an average slope of less than 5 degrees (Figure 1). The average annual temperature is −1.7 °C, and the annual precipitation is 590.1 mm, with 80% of the rainfall occurring between May and September [28]. The main vegetation types in the area are alpine shrubs and alpine meadows, with sporadic Tibetan barley (Hordeum vulgare) as supplementary winter fodder for livestock. Alpine shrubs, consisting of Potentilla fruticosa, are found on shady slopes and depressions with sufficient moisture. The alpine meadows, primarily composed of grasses and sedges, have long been used as summer pastures (grazing from June to September), with yaks (Bos grunniens) and Tibetan sheep (Ovis aries) as the main livestock. The long-term implementation of different grazing systems, including public grazing (no control of grazing intensity), communal grazing (typically 2 or 3 families grazing together), and individual family pastures, has led to varying degrees of degradation of the alpine meadows. This degradation has resulted in a succession of native plants, with the dominant species shifting from grasses and sedges to legumes and other toxic species [29].

2.2. Acquisition and Processing of Measured Spectrum

In this study, different vegetation types were measured using a Spectral Evolution RS-8800 ground spectrometer (Spectral Evolution, Haverhill, MA, USA) with a spectral range of 350–2500 nm, a spectral resolution of 3–8 nm, and a sampling interval of 1 nm. Measurements were taken between 10:00 and 14:00 from May to September 2023 under sunny, cloudless conditions with good lighting. The probe, with a field of view of 25°, was positioned vertically downward at a distance of approximately 1 metre from the measured feature.

We selected four sample areas with flat terrain and minimal elevation differences, including alpine meadow (AM), degraded alpine meadow (DAM), alpine shrub (AS) and Tibetan barley (TB), each with a sample area of 50 m × 50 m. The native alpine meadow, which had been under a long-term grazing ban, consisted mainly of grasses and sedges. In each sample area, nine sample plots were arranged in an S-shape, and each sample plot was randomly measured five times. Calibration was carried out before and after each measurement using a white board. After data acquisition, the hyperspectral data were first screened to remove outliers, including invalid data such as obvious deviations or fluctuation anomalies. In order to reduce noise from the instrument itself, the hyperspectral data, after rejecting the outliers, were processed for noise reduction using the Savitzky–Golay (SG) convolutional smoothing method [30]. Ultimately, a total of 156 valid sample data points were acquired.

2.3. Hyperspectral Data Analysis

In order to enhance the reflectance and absorption features in the critical band ranges, we performed first-order derivative transformations and continuum removal on the raw reflectance data.

First-order derivative (FDR) transformation is a commonly used method in spectral analysis [31]. It effectively eliminates baseline and other background interference but introduces noise, with higher-order derivatives leading to lower signal-to-noise ratios. In this study, we applied first-order derivative transformation to the SG-smoothed reflectance data, which reduces noise through a moving window. The calculation formula for this method is as follows:

$$FD{R}_{\lambda i}=\frac{R\left({\lambda }_{i+1}\right)-R\left({\lambda }_{i-1}\right)}{2\Delta \lambda }$$

(1)

where FDR_λi is the FDR value at wavelength λ_i; R(λ_i) is the spectral reflectivity at wavelength λ_i; λ_i is the band wavelength; and Δλ is the interval from wavelength λ_i + 1 to wavelength λ_i.

Continuum removal (CR) enhances the absorption and reflection features of spectral curves, making the differences in the eigenvalues of different spectral curves more pronounced [32]. This method enhances the absorption troughs in the spectral curves by setting the envelope values of the start and end points to 1, while the values between these points range are between 0 and 1. The specific calculation process is as follows:

$$C{R}_i=\frac{R_i}{R_{\mathit{Hi}}}$$

(2)

$$R_{\mathit{Hi}}=R_{\mathit{start}}+K\times ({\mu }_i-{\mu }_{\mathit{start}})$$

(3)

$$K=\frac{R_{\mathit{end}}-R_{\mathit{start}}}{{\mu }_{\mathit{end}}-{\mu }_{\mathit{start}}}$$

(4)

where CR_i is the corresponding envelope value at band i; R_i the spectral reflectance at band i; RH_i is the corresponding “hull” value at band i; K is the “hull” slope between the selected absorption starting point and the end point; R_start and R_end are the reflectivities at the absorption starting point and the endpoint, respectively; and μ_start and μ_end are the corresponding wavelength values.

2.4. Calculation of Characteristic Parameters

Vegetation spectra can be described parametrically using characteristic bands and parameters that quantitatively reflect spectral differences between vegetation types. In this study, we selected eight common spectral characteristic parameters, including six location parameters and two area parameters (

Table 1. Spectral characteristic parameter and its definition.

Parameter Type	Characteristic Parameter	Definition
Position parameter	Red edge amplitude	Maximum first order differential in 680–760 nm
	Red edge position	Band length corresponding to red edge amplitude
	Blue edge amplitude	Maximum first order differential in 490–530 nm
	Blue edge position	Band length corresponding to blue edge amplitude
	Yellow edge amplitude	Maximum first order differential in 560–640 nm
	Yellow edge position	Band length corresponding to yellow edge amplitude
Area parameter	Red edge area	Sum of first order differential in red edge range
	Blue edge area	Sum of first order differential in blue edge range

), and nine vegetation indices (

Table 2. Vegetation indices and its formula.

Parameter Type	Vegetation Index	Abbreviation	Formula
Broad-band vegetation index	Normalized difference vegetation index	NDVI	(Rnir − Rred)/(Rnir + Rred)
	Ratio vegetation index	RVI	Rnir/Rred
	Differential vegetation index	DVI	Rnir − Rred
	Enhanced vegetation index	EVI	2.5(Rnir − Rred)/(Rnir + 6 × Rred − 7.5 × Rb + 1)
	Soil Adjusted Vegetation Index	SAVI	((Rnir − Rred)/(Rnir + Rred + 0.5)) × 1.5
	Normalized Green–Red Difference Index	NDGI	(Rg − Rred)/(Rg + Rred)
Narrow-band vegetation index	Photochemical Reflectance Index	PRI	(R531 − R570)/(R531 + R570)
	Red Edge Normalized Difference Vegetation Index	RENDVI	(R750 − R705)/(R750 + R705)
	Plant Senescence Reflectance Index	PSRI	(R680 − R500)/R750

Note: Nir is the near-infrared band, red is the red band, b is the blue band, and g is the green band, where the number is the specified nanometre band.

).

2.5. Vegetation Type Identification

The minimum misclassification canonical analysis (MMCA) method was employed to select sensitive bands or parameters for reflectance, first-order derivative transformation, continuum removal, spectral characterisation parameters, and vegetation indices for different months [33]. The selection criterion was the proportion of variance explained by different projection directions in the MMCA results. The results indicated that the cumulative proportion of variance explained by the first three projection directions was close to 100% for different methods and months. Consequently, we combined the weight coefficients of each wavelength or feature parameter in the first three projection directions to rank and select the top eight wavelengths or features as variables for identification and classification. This study then used support vector machine (SVM), artificial neural network (ANN), random forest (RF), and decision tree classification (DT) methods to classify and identify the vegetation types in the study area. We randomly selected 70% of the sample data as training datasets and 30% as prediction datasets. Finally, the overall classification accuracy and the kappa coefficient were used to comprehensively evaluate and compare the classification accuracies.

3. Results

3.1. Spectral Reflectance Data Analysis

3.1.1. Original Spectral Reflectance Analysis

In the visible light band (400–700 nm), the reflectance changes of different features from June to September formed a peak and a valley: the green peak (510–560 nm) and the red valley (640–680 nm). Degraded alpine meadows had a higher reflectance than alpine meadows in both the green and red light bands. Alpine shrubs had the lowest reflectance within the green peaks and red valleys in June, August, and September, but had slightly higher reflectance than degraded alpine meadows in July. The reflectance characteristics of Tibetan barley were notably distinct. In August and September, it did not exhibit discernible peaks and valleys in the green and red light bands, and its reflectance was markedly higher than that of the other types. However, Tibetan barley showed the lowest reflectance in June and July. In May, none of the features showed distinct peaks and valleys in the visible light bands. The overall reflectance in descending order was degraded alpine meadows, alpine meadows, Tibetan barley, and alpine shrubs (Figure 2).

In the near-infrared band (750–1400 nm), all features reached their highest reflectance, forming distinct wave peaks. Alpine meadows and degraded alpine meadows had similar reflectance in the NIR band, with an average reflectance of 21–46%, higher than that of alpine shrubs and Tibetan barley. The reflectance of degraded alpine meadows exceeded that of alpine meadows in May, August, and September, with values of 35.2%, 39.91% and 35.5%, respectively, compared to alpine meadows with values of 34.1%, 38.5% and 33.7%. In June, alpine meadows had an average reflectance of 37.1%, while degraded alpine meadows had 34.6%. In July, the reflectance patterns were more complex: from 750–950 nm, alpine meadows had higher reflectance than degraded alpine meadows, but from 1150–1400 nm, degraded alpine meadows had higher reflectance. Between 950–1150 nm, there was little difference between the two. Tibetan barley and alpine shrubs had significantly lower reflectance in the NIR band compared to alpine meadows and degraded alpine meadows. In May, Tibetan barley and alpine shrubs had average reflectance values of 15.3% and 19.3%, respectively. In June and August, the reflectance of Tibetan barley was higher than that of alpine shrubs, with mean values of 29.8% and 35.9%, while alpine shrubs had averages of 20.2% and 27.3%. In July and September, alpine shrubs had higher reflectance than Tibetan barley, with averages of 32.6% and 27.9%, while Tibetan barley had 25.0% and 21.8% (Figure 2).

In the shortwave infrared range (1400–2500 nm), the reflectance of different features formed three peaks and two valleys. The peaks were observed in the ranges of 1550–1750 nm, 1830–1900 nm, and 1900–2400 nm, while the valleys occurred in the ranges of 1400–1550 nm and 1750–1830 nm. In this range, degraded alpine meadows had consistently exhibited higher reflectance than alpine meadows. The reflectance of degraded alpine meadows, Tibetan barley, and alpine shrubs varied across different months. In May and June, alpine meadows and degraded alpine meadows had higher reflectance than Tibetan barley and alpine shrubs. In July, alpine shrubs had higher reflectance than both alpine meadows and Tibetan barley, second only to degraded alpine meadows. In August and September, the reflectance of Tibetan barley exceeded that of alpine meadows and degraded alpine meadows (Figure 2).

3.1.2. First-Order Derivative Reflectance

The first-order derivative of reflectance for different vegetation types was significantly amplified in certain band ranges, especially in the 680–760 nm (the red edge range) and to a lesser extent in the 490–530 nm (the blue edge range). The first-order derivatives of Tibetan barley were significantly lower than those of the other vegetation types, especially in May and September. Alpine shrub, alpine meadow, and degraded alpine meadow showed significant differences in the red and blue edge ranges. In June, alpine shrub had higher derivative values than other types in both ranges, but in the remaining months, its values were lower than those of alpine meadow and degraded alpine meadow. In May, alpine meadows had significantly higher values than degraded alpine meadows in both the red-edge and blue-edge ranges. In June and July, alpine meadows had higher values than degraded alpine meadows in the red-edge range, whereas in August and September, the values of the two types were more similar (Figure 3).

The amplitude of the red and blue edges was greater for alpine meadows and degraded alpine meadows than for other types, except in June. Alpine meadows had the highest red-edge amplitude in May, July, and August, and the highest blue-edge amplitude in May. Degraded alpine meadows had the highest red-edge amplitude in September and the highest blue-edge amplitude from July to September. Tibetan barley showed maximum red-edge and blue-edge amplitudes in June. The red-edge position for different types was around 700 nm, shifting to the left in May and September and to the right in June–August. The position of the blue edge was generally at 524 nm. Degraded alpine meadows had the largest yellow-edge amplitude from May to July, corresponding to a wavelength of about 630 nm, while Tibetan barley had the largest yellow-edge amplitude in August and September, corresponding to a wavelength of 560 nm. The area of the red and blue edges was generally larger for alpine meadows and degraded alpine meadows than for Tibetan barley and alpine shrubs. However, Tibetan barley had the largest red edge area in June and the largest blue edge area in August (

Table 3. Characteristic parameters of different vegetation types from May 2023 to September 2023.

202305	RA	RP	BA	BP	YA	YP	RAR	BAR
Tibetan Barley	0.1106	722	0.05	525	0.024	560	5.89	1.5146
Degraded alpine meadow	0.1717	701	0.0618	525	0.041	631	9.4476	2.1159
Alpine meadow	0.2403	702	0.0661	524	0.0346	630	12.6234	2.0256
Alpine shrub	0.0424	698	0.0152	526	0.0167	632	2.5511	0.5353
202306	RA	RP	BA	BP	YA	YP	RAR	BAR
Tibetan Barley	0.645	729	0.1173	524	−0.0028	629	29.9428	2.4564
Degraded alpine meadow	0.4632	721	0.1033	525	0.0131	630	22.314	2.3889
Alpine meadow	0.5899	723	0.1156	524	0.0093	629	28.3945	2.5924
Alpine shrub	0.2118	720	0.0486	524	0.0093	630	10.1985	1.0991
202307	RA	RP	BA	BP	YA	YP	RAR	BAR
Tibetan Barley	0.5834	731	0.0859	524	−0.0026	629	26.0273	1.7532
Degraded alpine meadow	0.7424	724	0.1239	524	0.0037	629	34.576	2.6645
Alpine meadow	0.8014	727	0.1112	524	0.0017	630	36.4924	2.3085
Alpine shrub	0.5891	721	0.1015	524	0.0024	630	26.4427	2.0736
202308	RA	RP	BA	BP	YA	YP	RAR	BAR
Tibetan Barley	0.3347	697	0.1216	523	0.0603	560	17.1052	3.7304
Degraded alpine meadow	0.682	723	0.124	524	0.0089	629	33.5414	2.7234
Alpine meadow	0.692	724	0.117	524	0.0054	629	33.3821	2.584
Alpine shrub	0.4656	721	0.0746	525	0.005	630	20.8703	1.5272
202309	RA	RP	BA	BP	YA	YP	RAR	BAR
Tibetan Barley	0.0691	692	0.0467	516	0.0442	560	3.6087	1.8427
Degraded alpine meadow	0.4797	702	0.1042	524	0.0225	629	24.0076	2.5229
Alpine meadow	0.461	703	0.0984	524	0.0166	629	24.1003	2.3462
Alpine shrub	0.3918	703	0.0765	524	0.0125	630	17.7581	1.6616

Note: red-edge amplitude (RA), red-edge position (RP), blue-edge amplitude (BA), blue-edge position (BP), yellow-edge amplitude (YA), yellow-edge position (YP), red-edge area (RAR), blue-edge area (BAR).

).

3.1.3. Continuum Removal

The continuum removal showed two absorption valleys in the blue and red light bands and one reflection peak in the green light bands, with deeper absorption in the red light. The highest absorption features in the red light bands were observed in alpine meadows, particularly during the months of May and July–August. Tibetan barley and alpine shrubs exhibited the lowest absorption levels in June and September, respectively. Degraded alpine meadows showed moderate absorption values in the red light bands. Tibetan barley and alpine shrubs usually had higher absorption values in the red light bands, with Tibetan barley peaking in May and August–September and alpine shrubs in June–July. Alpine meadows had the lowest absorbance in the blue light bands in July–September, while Tibetan barley had the highest absorbance during the same period. Degraded alpine meadows had the highest absorption values in May–June, while alpine shrubs were typically in the middle range. The reflectance characteristics of alpine meadows in the green light bands were generally insignificant and lowest in July–September, while Tibetan barley had the highest values in May and August–September. Degraded alpine meadows and alpine shrubs had intermediate reflectance characteristics in the green light bands, peaking only in June and July, respectively (Figure 4).

3.2. Vegetation Indices

All vegetation indices showed an initial increase followed by a decrease over time, peaking in July and reaching a minimum in either May or September. Tibetan barley, as an artificial vegetation type, showed the most pronounced changes, with a rapid increase from June to July and a significant decrease from August to September. In contrast, alpine meadows, degraded alpine meadows, and alpine shrubs showed a more gradual increase in vegetation indices from May to June, followed by slower fluctuations. In general, the vegetation indices of alpine meadows and Tibetan barley were higher, followed by degraded alpine meadows and alpine shrubs. However, these patterns varied between different months and vegetation indices. Tibetan barley showed significant differences from the other types in key months for all vegetation indices, especially in August–September for NDVI, RVI, DVI, EVI, SAVI, NDGI, and RENDVI, and in June-July for PRI and PSRI. Alpine meadows generally had higher vegetation indices than degraded alpine meadows, although the differences were not significant for most indices. In particular, RVI and NDGI showed significant differences between alpine meadows and degraded alpine meadows in June–August and in June. Alpine shrubs did not differ significantly from alpine meadows and degraded alpine meadows in most vegetation indices, but DVI and NDGI could distinguish alpine shrubs from other types relatively well. For example, alpine shrubs were significantly different from other types in all months for DVI and in June for NDGI (Figure 5).

3.3. Identification of Vegetation Types and Accuracy Assessment

The classification results for different feature variables showed that ANN and SVM had the highest classification accuracy. The average overall accuracy of ANN was 87%, 90%, 85%, and 87% for the original bands, first-order derivative, continuum removal, and characteristic parameters, respectively. DT and RF had relatively lower classification accuracies, with averages ranging from 64% to 84% and 72% to 86%, respectively. Among the different data transformations and characteristic parameters, first-order derivatives achieved the highest average overall average accuracy (87.4%), followed by characteristic parameters (85.1%) and continuum removal (84.7%), with original bands having the lowest overall accuracy (76.4%). The average classification accuracies based on original bands, first-order derivatives and characteristic parameters were highest in May, ranging from 85% to 93%, followed by July, and lowest in September, ranging from 70% to 77%. However, the mean classification accuracy based on continuum removal was highest in June (94%), followed by July. The different classification methods performed differently over the months. ANN had the highest classification accuracy based on the original bands, first-order derivatives, and continuum removal in June or July. RF had the lowest classification accuracy based on the original bands in July and the highest based on characteristic parameters in July. SVM and DT also achieved their highest classification accuracy with first-order derivatives in June and feature covariates in July, respectively (Figure 6).

4. Discussion

4.1. Influence of Vegetation Phenology on Spectral Reflectance

Changes in spectral features are associated with phenological stages of plants, including greening, flowering, maturity and wilting. These stages exhibit distinct reflectance spectral features, which have been demonstrated in multi-temporal spectral studies across different regions and vegetation types [34,35,36]. For alpine meadows, greening starts in April, blooming and maturity occur from July to August, and wilting starts in September. As a result, the average spectral reflectance of the different vegetation types is at its highest in July and at its lowest in May [37].

The accuracy of vegetation classification is influenced by the phenological stage of the plants, as their spectral reflectance changes with time. Our study found that the classification accuracy of degraded alpine meadows and other vegetation types was highest in May and July. In May, Tibetan barley was just beginning to germinate, making its reflectance different from other types. In addition, the phenological characteristics of alpine shrubs differed from those of alpine meadows, which had not yet greened. In July, the different species entered their blooming or ripening stage, resulting in more pronounced differences in spectral reflectance. Other studies have also found higher identification accuracy during peak growing seasons, although the specific timing may vary from species to species [34,35,36].

4.2. Spectral Transformation Methods and Characterisation Parameters

The first-order derivative reduces the influence of soil background on plant spectral reflectance, partially eliminating sunlight and atmospheric effects and enhancing the extraction of plant spectral information. Continuum removal effectively highlights the absorption and reflectance features of the spectral curves and significantly distinguishes the eigenvalues of different spectral curves [38]. These methods have been widely used in practical applications. In this study, mathematical transformations of the original spectral data were performed to both eliminate external factors and increase the differences between plant reflectance spectral curves. Our study demonstrates two main effects of data transformation: firstly, an enhancement of plant absorption and reflectance characteristics, and secondly, an improvement in the accuracy of plant type identification. For example, the FDR enhanced the differences between plant types in the red-edge region, and the continuum removal amplified the absorption features in the blue and red light regions. Many studies have confirmed that data transformations, including first and second-order derivatives, continuum removal, and logarithmic transformations, improve detection rates. A study of grassland species identification in Inner Mongolia, China, showed that first-order derivatives and successive removal of raw reflectance improved classification accuracy [24,25]. Dong et al. [39] applied six different mathematical transformations to original alpine meadow spectral data on the QTP and used machine learning approaches for species identification.

Different vegetation indices are effective indicators for reflecting differences and classifying features. They help to distinguish the internal structure of vegetation, establish linear and non-linear combinations of spectral information, and improve the accuracy of remotely sensed spectral data [21]. The findings of this study suggest that there is a discrepancy in the capacity of different vegetation indices to differentiate between degraded alpine meadows and other types of vegetation. In particular, DVI and NDGI showed significant differences between vegetation types and were selected as sensitive features for classification. This may reflect the fact that different vegetation types vary more in the NIR, red and green bands. However, the accuracy of classification based on characteristic parameters, including vegetation indices, was not as high as that of the first-order derivative transformations are more effective in highlighting the spectral variability of different vegetation types.

4.3. Influence of Classification Methods on Identifying Alpine Grassland Vegetation Types

The selection of appropriate classification models is crucial for reliable and efficient species identification and degradation monitoring in alpine grasslands. The application of machine learning methods is becoming increasingly prevalent in studies such as hyperspectral feature identification and aboveground biomass inversion, due to their capacity to process high-dimensional data. Dao, Axiotis and He [34] used random forests to map native and invasive grassland species based on high-resolution hyperspectral imagery, and Huang, Li, Xu, Ma, Li and Liu [23] used machine learning methods such as SVM, RF, and ANN to predict the aboveground biomass of alpine meadows on the Qinghai–Tibetan Plateau. The results of this study demonstrate that both SVM and ANN achieve greater accuracy in detecting degraded alpine meadows when compared to alternative methods, across a range of months and data forms. In hyperspectral remote sensing imagery, convolutional neural networks (CNNs) are more widely used and generally more accurate than traditional machine learning methods such as support vector machines. For example, researchers in Inner Mongolia, China, used an improved CNN model to identify grassland species, achieving up to 99% accuracy [40].

This study did not include spectral measurements at the species or community level for the rapid identification of degraded alpine meadows and other vegetation types. It would be beneficial for future research to address this by measuring the spectral reflectance of key indicator species. Furthermore, the current study is limited to the sample point scale, which limits its applicability to large areas. To extend the results of the study to a broader spatial scale, future research should combine ground-based observations with hyperspectral imagery.

5. Conclusions

In this paper, the spectral characteristics of degraded alpine meadows and other typical vegetation types on the QTP were analysed using hyperspectral measurements. The study examined the influence of observation time, data processing methods, and classification techniques on the accuracy of vegetation identification. The results indicated that spectral characteristics varied among different vegetation types and across months. The spectral curves of alpine meadows and degraded alpine meadows were relatively similar, differing primarily in the near-infrared bands. In contrast, the spectral curves of alpine shrubs and Tibetan barley exhibited greater variability and were significantly lower than those of other vegetation types. The application of first-order derivative and continuum removal methods resulted in an enhancement of the reflectance and absorption features observed in the red and visible bands, respectively. Additionally, indices such as DVI, RVI, and NDGI more effectively distinguished between vegetation types. This study recommends that, in order to accurately identify degraded alpine meadows, a combination of spectral measurements taken in July, first-order derivative transformations of the raw spectral data, and the application of a SVM or ANN for classification should be employed. Further research is required to focus on species-level spectral measurements in order to analyse the spectral characteristics of key grassland degradation indicator species, combined with hyperspectral imagery in order to conduct spatial scale degradation assessments of alpine meadows.