Hyperspectral Remote Sensing Combined with Ground Vegetation Surveys for the Study of the Age of Rodent Mounds

Abstract

Background: Rodents severely damage the ecological environment of grasslands, and rodent mounds of different ages require distinct management strategies. Understanding the age of these mounds aids in formulating targeted restoration measures, which can enhance grassland productivity and biodiversity. Current surveys of rodent mounds rely on ground exposure and mound height to determine their age, which is time-consuming and labor-intensive. Remote sensing methods can quickly and easily identify the distribution of rodent mounds. Existing remote sensing images use ground exposure and mound height for identification but do not distinguish between mounds of different ages, such as one-year-old and two-year-old mounds. According to the existing literature, rodent mounds of different ages exhibit significant differences in vegetation structure, soil background, and plant diversity. Utilizing a combination of vegetation indices and hyperspectral data to determine the age of rodent mounds aims to provide a better method for extracting rodent hazard information. This experiment investigates and analyzes the age, distribution, and vegetation characteristics of rodent mounds, including total coverage, height, biomass, and diversity indices such as Patrick, Shannon–Wiener, and Pielou. Spectral data of rodent mounds of different ages were collected using an Analytical Spectral Devices field spectrometer. Correlation analysis was conducted between vegetation characteristics and spectral vegetation indices to select key indices, including NDVI₆₇₀, NDVI₇₀₅, EVI, TCARI, Ant, and SR. Multiple stepwise regression and Random Forest (RF) inversion models were established using vegetation indices, and the most suitable model was selected through comparison. Random Forest modeling was conducted to classify plateau zokor rat mounds of different ages, using both vegetation characteristic indicators and vegetation indices for comparison. The rodent mound classification models established using vegetation characteristic indicators and vegetation indices through Random Forest could distinguish rodent mounds of different ages, with out-of-bag error rates of 36.96% and 21.74%, respectively. The model using vegetation indices performed better. Conclusions: (1) Rodent mounds play a crucial ecological role in alpine meadow ecosystems by enhancing plant diversity, biomass, and the stability and vitality of the ecosystem. (2) The vegetation indices SR and TCARI are the most influential in classifying rodent mounds. (3) Incorporating vegetation indices into Random Forest modeling facilitates a precise and robust remote sensing interpretation of rodent mound ages, which is instrumental for devising targeted restoration strategies.

1. Introduction

Rodent infestations on grasslands have become a significant obstacle to the sustainable development of grassland animal husbandry and have had a profound impact on grassland ecosystems. These infestations lead to a decline in the yield of high-quality forage grasses, reduced areas available for livestock, soil erosion, and grassland desertification, among other issues [1]. Effective control of rodent populations has become a critical component of grassland ecosystem protection. The plateau zokor (Myospalax baileyi), a major rodent pest in alpine meadows, influences the structure and function of the meadow ecosystem through changes in its population size. Rodent mounds of different ages exhibit considerable differences in vegetation appearance and plant composition [2]. The coverage of plant communities on rodent mounds gradually increases over time from their formation, with clear differences in plant community morphology at various stages. New mounds are almost completely exposed, while older mounds, formed over several years, are dominated by 1- and 2-year-old plants. As vegetation succession progresses on rodent mounds, 1- and 2-year-old plants and perennial grasses together form the dominant species of the mound community [3]. Niu et al. [4] showed that the annual successional coupling of rodent mounds at different stages is crucial for species coexistence and promotes the renewal of alpine meadows. Furthermore, rodent mounds of different ages require different management strategies. The impact of newer and older mounds on grassland ecosystems varies. Understanding the age of these mounds can help develop targeted restoration measures, which can enhance both grassland productivity and biodiversity [5].

Traditional methods for determining the age of rodent mounds often rely on assessing soil properties, vegetation cover, and species composition, which is both time-consuming and labor-intensive [6]. Yang Y.B. et al. [7] conducted a study classifying plateau zokor mounds in subalpine meadows based on vegetation cover, abundance, plant species composition, and mound height. Zhang Qian et al. [8] analyzed various factors such as the total cover, importance value, richness, evenness, and dominant species of plant communities in rodent mounds of different ages, discovering significant variations. Hyperspectral remote sensing data, which provide detailed spectral information, support large-scale vegetation monitoring and offer the benefits of simplicity, speed, high sensitivity, and non-destructive analysis [9,10]. Tao Zhang et al. [11] employed a UAV-based hyperspectral remote sensing platform for data collection, enhancing the efficiency of grassland rodent surveys. Thus, remote sensing methods can efficiently and conveniently identify rodent mound distributions. However, existing remote sensing imagery often relies on assessing bare ground and mound height to confirm the presence of rodent mounds. Gao X. et al. [12] successfully applied UAV hyperspectral remote sensing technology to identify rodent burrows in the desert grasslands of Inner Mongolia. Nevertheless, there is a notable lack of studies that utilize hyperspectral analysis for determining the age of rodent mounds. The current literature indicates that rodent mounds of varying ages exhibit significant differences in vegetation structure, soil background, and plant diversity [5]. Hyperspectral remote sensing, with its high resolution, extensive spectral bands, and continuous data characteristics, accommodates both spectral continuity and distinguishability [13], making it a robust tool for vegetation and biodiversity classification [14]. Numerous studies have employed spectral data for classifying vegetation cover and identifying characteristic plant species. For instance, Wen Tong et al. [15] analyzed the original spectral data of 36 common plants in the alpine grasslands of the Sanjiangyuan region, showing that the spectra align with the typical characteristics of green plants. Due to differences in plant morphology, the main spectral distinctions were found in the visible light range. Feng Haiying et al. [16] demonstrated that the red-edge region (680–760 nm) is highly sensitive to vegetation cover and that the first-order derivative of the red-edge spectrum correlates strongly with vegetation cover, making the red-edge slope (k) a useful parameter for estimating vegetation cover. Miao Chunli et al. [17] indicated that the RF machine learning modeling method is effective for estimating aboveground biomass in alpine meadows, and UAV-based hyperspectral remote sensing can achieve high-precision monitoring of aboveground biomass in these environments. As the succession stage of rodent mounds progresses, changes are observed in the surface species count, richness index, Shannon–Wiener index, and evenness index [3]. Thus, hyperspectral remote sensing data can be utilized to analyze the vegetation and soil characteristics of zokor rat mounds. By integrating spectral indices with vegetation characteristic indices, the age of rodent mounds can be effectively determined.

Hence, this study conducts a survey of the vegetation conditions of plateau zokor rat mounds in an alpine meadow, analyzing the correlation between the vegetation characteristics of the Zokor mounds and the vegetation spectral indices. It also performs spectral classification of plateau zokor mounds of different ages, with the aim of accurately and stably determining the age of plateau zokor mounds using remote sensing. This approach is expected to enhance the precision and stability of remote sensing assessments for the management of these mounds.

2. Research Methodology

2.1. Overview of the Study Area

The study area is located in Kajiado Township, Hezuo City, Gannan Tibetan Autonomous Prefecture, Gansu Province (Figure 1), at the transitional zone between the Qinghai–Tibet Plateau, the Loess Plateau, and the Longnan Mountains. The topography is characterized by higher elevations in the north and south and lower elevations in the middle, with the city center situated in the central part of the Hezuo Basin. The average altitude is 2788 m. The region has a cold and humid climate, with a long cold season and a short warm season. The annual average temperature ranges from −0.5 °C to 3.5 °C, with the highest extreme temperature at 28 °C and the lowest at −23 °C. The average annual precipitation is 545 mm, mainly concentrated from July to September. The region has an average frost-free period of 48 days, with ample sunlight throughout the year and high solar energy utilization rates. The terrain is divided into three geomorphological zones: the high mountain area, the mountain plateau area, and the hilly mountainous area.

The Gannan grasslands, situated in the upper reaches of the Yangtze and Yellow Rivers, serve as a vital natural ecological barrier and water conservation area, playing an indispensable role in safeguarding the ecological security of these river systems [18,19]. This region is also a key area for animal husbandry and the production of agricultural and livestock products in Gansu Province. Research has shown that rodent damage poses a serious threat to the sustainable management of these grasslands [20]. Statistics indicate that rodent infestations in Gannan have affected nearly 2000 hectares, accounting for 50% of the total usable grassland area [21]. These infestations have significantly hindered the healthy development of the livestock industry in Gannan. Therefore, controlling rodent populations and reducing the impact of infestations are critical issues for the economic development of Gannan Prefecture [22]. Managing the population of plateau zokors, including distinguishing their ages, is essential for effective control and disaster prevention [23].

2.2. Research Methodology Flowchart

The collected spectral data were preprocessed to remove interference and calculate spectral indices. A correlation analysis was conducted on the collected vegetation characteristic indicators and the calculated spectral indices to identify vegetation indices with strong relationships. Subsequently, multivariate stepwise regression models and Random Forest models were established using these vegetation indices for comparison. The models were compared by assessing their R-squared (R²) and Root Mean Square Error (RMSE) to select the most suitable one. Following this, classification models for zokor mounds were developed using vegetation characteristic indices and selected vegetation indices, which were then validated and compared (Figure 2).

2.3. Field Investigation

2.3.1. Sample Plot Selection and Ground Survey

Three plots were established in the study area. Within each plot, the ages of the mounds were determined based on soil properties, vegetation coverage, and species composition, and they were categorized into four successional stages (

Table 1. Age division of rat mounds.

Type	Age of the Mound	Character of Mound	State of Vegetation
1	Mounds through the first winter	Less compaction soil	Total cover is more than 20%, and annual plants account for more than 90% of the total
2	Mounds through the second winter	More compaction soil	Total cover is more than 60%, and perennial Gramineae plants account for less than 15% of the total
3	Mounds through the third winter	Compaction soil	Total cover is more than 75%, and annual plants account for less than 10% of the total
4	Mounds through the fourth winter or more	Compaction soil	The total cover is more than 85%
5	Natural vegetation	Compaction soil	The total cover is more than 90%

): one-year mounds (ZM0), two-year mounds (ZM1), three-year mounds (ZM2), and multi-year mounds (ZMM), following the classification method outlined by Yang Y.B [7]. In each successional stage, 20 mounds were selected. The mounds were approximately elliptical in shape, with long and short axes ranging from 45 cm × 45 cm to 65 cm × 68 cm. The plot size for the mounds was 50 cm × 50 cm [8]. A total of 80 mounds were selected, with the adjacent normal grassland vegetation in each successional stage serving as the control group (CK). Aboveground biomass, height, coverage, and frequency were investigated. Additionally, a portable field spectrometer was used to measure the spectral data of both the mounds and the CK group grassland vegetation.

2.3.2. Spectral Data Acquisition

Plant community spectral data were collected using a portable FieldSpec^® 4 Hi-Res ASD ground spectrometer from ASD (Analytical Spectral Devices, Inc., Boulder, CO, USA) in 2020 during the bloom period. The spectrometer’s wavelength range is 350–2500 nm, covering the entire visible light spectrum (380–780 nm) and part of the near-infrared range (780–2500 nm). The spectral resolution wavelengths were 3 nm for the 350–1000 nm range and 10 nm for the 1000–2500 nm range, with sampling intervals of 1.4 nm and 2.5 nm, respectively, and a resampling interval of 1 nm for data output (

Table 2. Instrument parameters and requirements.

Instrument Name	Spectral Band	Test Time	Weather Requirements	Assay Requirements
Field Spec® 4 Hi-Res ASD	300~2500 nm	10:00~14:00	Dry, windless, clear and cloudless	Whiteboard correction is carried out in time after measurement

).

2.4. Data Processing

The following vegetation indices were calculated: the importance value (IV) [23], Patrick’s richness index [24] (hereafter simplified as PK), the Shannon diversity index [25] (hereafter simplified as H), and Pielou’s evenness index [26] (hereafter simplified as P).

IV = (relative cover + relative height + relative biomass)/3

The richness index (Patrick) is the total number of species occurring within the sample plot.

$$\mathrm{H}=-{\sum }_{\mathrm{i}=1}^{\mathrm{S}}\mathrm{I}\mathrm{n}\mathrm{P}\mathrm{i}\left(\mathrm{P}\mathrm{i}\right)$$

$$\mathrm{P}=(-{\sum }_{\mathrm{i}=1}^{\mathrm{S}}\mathrm{I}\mathrm{n}\mathrm{P}\mathrm{i}\mathrm{P}\mathrm{i})/\mathrm{I}\mathrm{n}\mathrm{S}$$

where S is the number of species in the sample.

Pi is the relative importance value of species i.

The plant spectral data were preprocessed using ViewSpec Pro software (Version 6.20) on the spectrometer to remove data with large spectral differences within the same plant feature. GABS was calculated using Equation (3), as the spectral curves were greatly affected by external environmental noise [27].

$$\mathrm{R}\mathrm{E}\mathrm{F}(\mathrm{\lambda })={\displaystyle \frac{{\mathrm{R}}_{\mathrm{\lambda }-2}+{\mathrm{R}}_{\mathrm{\lambda }-1}+{\mathrm{R}}_{\mathrm{\lambda }}+{\mathrm{R}}_{\mathrm{\lambda }+1}+{\mathrm{R}}_{\mathrm{\lambda }+2}}{5}}$$

(1)

$$\mathrm{A}\mathrm{B}\mathrm{S}\left(\mathrm{\lambda }\right)=\lg (1/{\mathrm{R}\mathrm{E}\mathrm{F}}_{\mathrm{\lambda }})$$

(2)

$$\mathrm{G}\mathrm{R}\mathrm{E}\mathrm{F}\left(\mathrm{\lambda }\right)=\left[{\mathrm{R}\mathrm{E}\mathrm{F}}_{\mathrm{\lambda }+1}-{\mathrm{R}\mathrm{E}\mathrm{F}}_{\mathrm{\lambda }-1}\right]/\left[\left(\mathrm{\lambda }+1\right)-\left(\mathrm{\lambda }-1\right)\right]$$

(3)

$$\mathrm{G}\mathrm{A}\mathrm{B}\mathrm{S}\left(\mathrm{\lambda }\right)=\left[{\mathrm{A}\mathrm{B}\mathrm{S}}_{\mathrm{\lambda }+1}-{\mathrm{A}\mathrm{B}\mathrm{S}}_{\mathrm{\lambda }-1}\right]/\left[\left(\mathrm{\lambda }+1\right)-\left(\mathrm{\lambda }-1\right)\right]$$

(4)

where R is the original reflectivity and λ is the wavelength (nm).

The acquired REF, ABS, GREF, and GABS data were comparatively analyzed to find the characteristic wavelength bands.

2.5. Vegetation Index Selection

Spectral measurements are susceptible to interference from external factors such as soil background, atmospheric aerosols, and vegetation canopy, which can cause biases in the measurement results. However, the RVI and VARI can effectively correct for the effects of atmospheric aerosols and eliminate some radiometric errors. NDVI₆₇₀ and NDVI₇₀₅ are highly sensitive to canopy structure; the PSRI can be used to detect and monitor vegetation health; the GNDVI is more sensitive to vegetation canopy greenness; the PRI is sensitive to plant carotenoids; the NDWI detects canopy moisture; the OSAVI eliminates some soil background effects; and the VARI corrects most macro-aerosol effects. Selecting these vegetation indices helps improve spectral accuracy and eliminate radiometric errors [28,29] (

Table 3. Spectral index.

Vegetation Indices	Name	Calculation Formula
GI	Greenness vegetation index	R554/R667
ARVI	Atmospheric impedance vegetation index	(R810 − (2R680 − R480))/(R810 + (2R680 − R480)
VARI	Visual barometric impedance index	(R555 − R680)/(R555 + R680 − R480)
NDVI705	Normalized vegetation index 705	(R750 − R705)/(R750 + R705)
MSR705	Improved red-edge ratio vegetation index	(R705 − R445)/(R705 + R445)
NDVI670	Normalized vegetation index 670	(R800/R670)/(R800 + R670)
CI	Chlorophyll index	(R750/R705) − 1
PSRI	Vegetation attenuation index	(R680 − R500)/R750
RGI	Relative green index	R690/R550
EVI	Enhanced vegetation index	(2.5 × (R782 − R675)/(R782 + 6R675 − 7.5R445 + 1)
RVI	Vegetation index	R800/R760
SAVI	Soil-adjusted vegetation index	1.5 × ((R800 − R760)/(R800 + R760 + 0.5))
MCARI	Chlorophyll absorption correction index	((R700 − R670) − 0.2 × (R700 − R550)) × (R700/R670)
TCARI	Improved chlorophyll absorption vegetation index	3 × ((R700 − R670) − 0.2 × (R700 − R550)) × (R700/R670)
OSAVI	Soil-regulated vegetation index	1.16 × ((R800 − R670)/(R800 + R670 + 0.16))
CIrededge	Red-edge chlorophyll index	(R700 − R800)/(R690 − R720) − 1
PRI	Photochemical vegetation index	(R570 − R531)/(R570 + R531)
CRI	Carotenoid reflex index	(1/R510) − (1/R550)
ANT	Anthocyanin reflection index	((1/R550) − (1/R700))
SR	The NIR/R ratio is called the simple ratio	R900/R680

).

2.6. Data Analysis

The analysis was conducted using R (Version 4.3.3, Vienna, Austria) to assess the correlation between vegetation characteristic indicators and vegetation indices. Subsequently, SPSS software (Version 28.0, Armonk, NY, USA) was used to perform a multiple stepwise regression analysis on the selected vegetation indices. The spectral variables were then subjected to Random Forest inversion modeling using the R language package, and the model was tested with a test set. The Random Forest (RF) is a classification method based on K decision trees and integrates these K decision trees to obtain a combined classification result. The classification result is determined by a majority in the classification results from each decision tree [30,31,32].

3. Results and Analysis

3.1. Vegetation Characteristics Survey of Rodent Mounds Across Different Successional Stages

The results of the survey on the cover, height, and biomass of the sample plots in different years of rodent mound succession (

Table 4. Plant and cover survey of sample plots.

Vegetation Composition	Plant Statistics
	ZM0	ZM1	ZM2	ZMM
Coverage/%	21.62	60.54	72.12	88.73
1, 2-year-old plants	whitefly Lap pula myosotis V. Wolf	Whitefly Lap pula myosotis V. Wolf	Whitefly Lap pula myosotis V. Wolf	Whitefly Lap pula myosotis V. Wolf
	microporous grass (botany) Microula sikkimensis Hems	microporous grass (botany) Microula sikkimensis Hems	chard (Beta vulgaris), a foliage beet Potentilla chinensis Ser	microporous grass (botany) Microula sikkimensis Hems
	chard (Beta vulgaris), a foliage beet Potentilla chinensis Ser	chard (Beta vulgaris), a foliage beet Potentilla chinensis Ser	Rice hip celery Ammi majus L.	chard (Beta vulgaris), a foliage beet Potentilla chinensis Ser
		Amaranth Artemisia Portulaca oleracea L.		Amaranth Artemisia Portulaca oleracea L.
		Rice hip celery Ammi majus L.		Rice hip celery Ammi majus L.
				Plantain Plantago asiatica L.
Perennial herbaceous plant	Gentian Gentiana scabra Bunge	Gentian Gentiana scabra Bunge	Gentian Gentiana scabra Bunge	Gentian Gentiana scabra Bunge
	Buttercup Ranunculus japonicus Thunb	Buttercup Ranunculus japonicus Thunb	Buttercup Ranunculus japonicus Thunb	Buttercup Ranunculus japonicus Thunb
	Lris Iris tectorum Maxim	Lris Iris tectorum Maxim	Lris Iris tectorum Maxim	Lris Iris tectorum Maxim
	Flat beans Melissitus ruthenicus (L.) Peschkoua	Echinococcus beans Oxytropis DC	Echinococcus beans Oxytropis DC	Echinococcus beans Oxytropis DC
	Beaded tooth Polygonum viviparum Linn	Flat beans Melissitus ruthenicus Peschkoua	Artemisia cold Artemisia frigida Willd. Sp. Pl	Artemisia cold Artemisia frigida Willd. Sp. Pl.
	Gentiana G. macrophylla	Beaded tooth Polygonum viviparum Linn	Flat beans Melissitus ruthenicus Peschkoua	Flat beans Melissitus ruthenicus (L.) Peschkoua
	Siberian butterfly Polygonum sibiricum Laxm	Gentiana G. macrophylla	Beaded tooth Polygonum viviparum Linn	Beaded tooth Polygonum viviparum Linn
		Siberian butterfly Polygonum sibiricum Laxm	Gentiana G. macrophylla	Gentiana G. macrophylla
		Dandelion Taraxacum mongolicum	Siberian butterfly Polygonum sibiricum Laxm	Dandelion Taraxacum mongolicum
		Dwarf grass Kobresia humilis Sergievskaya	Dwarf grass Kobresia humilis Sergievskaya	Dwarf grass Kobresia humilis Sergievskaya
		Arrowroots Carumcarvi	Arrowroots Carumcarvi	Arrowroots Carumcarvi
		Anemone Anemone catharsis Kitag	Anemone Anemone catharsis Kitag	Anemone Anemone catharsis Kitag
			Oriental strawberry Fragaria orientalis Losinsk
Gramineae		Rough shadow grass Cleistogenes squarrosa Keng	Draped grass Elymus dahuricus Turcz	Precocious grass Poa Annual
			Precocious grass Poa Annual
Perennial weeds	Ask Jing Equisetum arvense L.	Ask Jing Equisetum arvense L.	Ask Jing Equisetum arvense L.	Ask Jing Equisetum arvense L.

) clearly show differences in the vegetation species on the surface of rodent mounds of different ages.

As the age of rodent mounds increases, the aboveground biomass (Pk) continuously increases, with the lowest value at ZM0 and significant differences between CK and ZMM, ZMM and ZM1, and ZM1 and ZM0 (p < 0.05) (Figure 3A). Similarly, the aboveground height (H) of the mounds continuously increases, with the lowest value at ZM0 and significant differences between CK and ZMM, ZMM and ZM1, and ZM1 and ZM0 (p < 0.05) (Figure 3B). The aboveground phytomass (P) of the mounds first increases, then decreases, and finally increases again, with the lowest value at ZM0 and a significant difference between CK and ZM2 (p < 0.05) (Figure 3C). The total coverage (T) of the mounds continuously increases, with the lowest value at ZM0 and significant differences between CK and ZMM, ZMM and ZM2, ZM2 and ZM1, ZMM and ZM1, and ZM1 and ZM0 (p < 0.05) (Figure 3D). Biomass (B) first increases, then decreases, and finally increases again, with the lowest value at ZM0 and a significant difference between CK and ZMM (p < 0.05) (Figure 3E). These differences in ground vegetation characteristics across different ages of rodent mounds reveal discernible patterns that can help distinguish the ages of the mounds.

3.2. Original Spectra of Rodent Mounds of Different Ages

The original spectral reflectance of rodent mounds at different ages (Figure 4) indicates that, at around 550 nm, all the other mounds except for ZM0 exhibit varying degrees of the “green peak” phenomenon, with the intensity decreasing in the order ZMM > ZM2 > CK > ZM1. A trough appears near 675 nm, and a plant-specific “red-edge” phenomenon occurs near 760 nm, where the reflectance of all mounds except that of ZM0 rises sharply. Although the overall trend of this red-edge phenomenon does not differ significantly among the treatments, variations in reflectance (REF) can still be observed. Within the near-red light wavelength range (750–970 nm), REF gradually increases, and the REF for the five treatments remains consistently high, with the highest-to-lowest order being ZM1, CK, ZMM, ZM2, and ZM0. Subsequently, the REF for ZM1 and CK decreases, and the first trough in the near-infrared band appears at around 970 nm for ZM1 and CK. At 1150 and 1250 nm, ZM1 and CK show a second trough in the near-infrared band, while ZM0, ZM2, and ZMM show their first trough. These findings suggest that the reflectance of the original spectrum contains errors when determining the age of the rodent mounds and therefore cannot be directly used for classification. Spectral indices need to be calculated to more accurately assess the succession stages of the rodent mounds, which have significant implications for ecological restoration and land management.

3.3. Hyperspectral Classification of Rodent Mounds of Different Ages

3.3.1. Relevance Analysis

Using Pearson’s correlation coefficient method to analyze the correlation between vegetation characteristic indicators and spectral indices (Figure 5), it was found that the spectral indices NDVI₆₇₀, NDVI₇₀₅, EVI, TCARI, and SR are highly significantly correlated (p < 0.01) with the vegetation characteristic indicators Pk, H, B, and T. The spectral indices EVI and SR are highly significantly correlated (p < 0.01) with the vegetation characteristic indicator P, and the spectral indices NDVI₆₇₀, NDVI₇₀₅, Ant, and TCARI are significantly correlated (p < 0.05) with the vegetation characteristic indicator P. After this analysis, spectral indices with strong correlations relevant to this study were selected for subsequent modeling: NDVI₆₇₀, NDVI₇₀₅, EVI, TCARI, Ant, and SR (hereafter referred to as “vegetation indices”).

3.3.2. Hyperspectral Classification Model Comparison

Multiple Stepwise Regression Model

We carried out modeling utilizing highly correlated vegetation indices (

Table 5. Multiple stepwise regression modeling.

Vegetation Indicators	Modeling	R2	RMSE
Patrick	Y = 20.039X(NDVI705) − 18.139X(NDVI670) + 24.190X(TCARI) − 0.187X(SR) + 9.635X(EVI) − 0.976X(Ant) + 5.844	0.388	1.888
Shannon–Wiener	Y = 12.559X(NDVI705) − 9.603X(NDVI670) + 17.918X(TCARI) − 0.021X(SR) − 1.480X(EVI) + 0.029X(Ant) + 1.967	0.482	0.284
Pielou	Y = 3.652X(NDVI705) − 2.769X(NDVI670) + 4.790X(TCARI) + 0.001X(SR) − 0.780X(EVI) + 0.048X(Ant) + 1.09	0.342	0.049
Total coverage	Y = 405.711X(NDVI705) − 319.057X(NDVI670) + 2668.599X(TCARI) + 4.102X(SR) − 424.667X(EVI) − 17.812X(Ant) + 71.715	0.704	18.663
Biomass	Y = 682.292X(NDVI705) − 518.034X(NDVI670) + 1285.846X(TCARI) + 5.645X(SR) − 314.295X(EVI) − 8.066X(Ant) + 71.183	0.477	24.151

).

To verify the accuracy of the model, a precision assessment was performed on the model (Figure 6), yielding the following results: Pk (R² = 0.388, RMSE = 1.888), H (R² = 0.482, RMSE = 0.284), P (R² = 0.338, RMSE = 0.049), T (R² = 0.704, RMSE = 18.663), and B (R² = 0.477, RMSE = 24.151). These results indicate that the multiple stepwise regression model, established using vegetation indices, has good predictive capability for vegetation characteristic indicators, with T demonstrating the optimal predictive ability among the indicators.

Random Forest Model

All vegetation indices were subjected to Random Forest inversion modeling (

Table 6. Random Forest model parameters.

Vegetation Indicators	The Mean of Squared Residuals	%Var Explained	R2	RMSE
Patrick	2.9786	46.6	0.4617	1.7258
Shannon–Wiener	0.0776	47.23	0.474	0.2787
Pielou	0.0035	−9.06	0.069	0.0598
Total coverage	339.9304	64.72	0.6485	18.4372
Biomass	475.728	51.33	0.5192	21.8112

), with the number of trees (ntree) set to 600 and the number of variables to try at each split (mytry) set to 3. The Random Forest models for T, B, Pk, and H had high variable explanatory power, meeting the model accuracy standards, with R-squared values of 64.72%, 51.33%, 46.6%, and 47.23%, respectively. However, the Random Forest model for the Pielou index had a poor R-squared value of −9.06%, which did not meet model accuracy standards and was therefore excluded from further consideration.

To assess the accuracy of the model, the established Random Forest model was validated using a test set (Figure 7). The linear regression model validation results for the predicted versus actual values are as follows: Pk (R² = 0.537, RMSE = 0.823), H (R² = 0.459, RMSE = 0.143), T (R² = 0.635, RMSE = 24.287), and B (R² = 0.948, RMSE = 15.173). These results indicate that the Random Forest model built using vegetation indices has good predictive capability for vegetation characteristic indicators, with T showing the best predictive ability for the vegetation characteristic indicators.

Combining the analysis of the two models, it was found that the effect of using Random Forest modeling is relatively better. This comparison suggests that the Random Forest approach outperforms others in terms of predictive accuracy for the vegetation indices, making it a suitable method for modeling vegetation characteristic indicators.

The variable importance results from the Random Forest model (Figure 8) are as follows: In section A, the importance ranking from highest to lowest is TCARI, EVI, SR, NDVI₇₀₅, NDVI₆₇₀, and Ant, with the TCARI being the most important variable and having an Incremental Mean Squared Error (IncMSE) value of 17.79. In section B, the importance ranking from highest to lowest is TCARI, EVI, NDVI₆₇₀, SR, Ant, and NDVI₇₀₅, with the TCARI being the most important variable and having an IncMSE value of 16.64. In section C, the importance ranking from highest to lowest is TCARI, EVI, SR, NDVI₆₇₀, NDVI₇₀₅, and Ant, with the TCARI being the most important variable and having an IncMSE value of 18.93. In section D, the importance ranking from highest to lowest is SR, NDVI₆₇₀, NDVI₇₀₅, EVI, TCARI, and Ant, with SR being the most important variable and having an IncMSE value of 15.48. The analysis indicates that the TCARI plays a significant role in the inversion of Pk, H, and T, while the SR is crucial for the inversion of B. Ant performed the poorest in inversion and was subsequently excluded from further modeling. The above demonstrates the importance of the vegetation indices SR and TCARI in modeling vegetation characteristic indicators, which can better detect the age of rodent mounds.

3.3.3. Rodent Mound Classification

Classification Based on Vegetation Feature Indicators

Using vegetation characteristic indicators to establish a Random Forest (RF) classification model, the error rate for the number of trees (ntree) indicates that when ntree is set to 600, the model’s internal error rate stabilizes. Thus, ntree is chosen to be 600. The variable importance plot of the RF model (Figure 9) shows that the out-of-bag (OOB) error rate is 36.96%. The RF model’s importance indicators, ranked from highest to lowest, are B, T, H, and Pk. The Gini coefficients of the RF model, ranked from highest to lowest, are B, T, H, and Pk. B is the variable with the highest importance indicators and Gini coefficients.

Utilizing vegetation characteristic indicators for modeling and classifying existing rodent mound data (

Table 7. Classification and error rates of aboveground plants in rat mounds in different years.

	ZM0	ZM1	ZM2	ZMM	CK	Error Rate
ZM0	4	4	0	0	0	0.5
ZM1	2	13	0	1	0	0.1875
ZM2	0	2	0	3	0	1
ZMM	0	1	3	4	0	0.5
CK	0	0	0	1	8	0.110

), the misclassification rates are as follows: ZM1 was incorrectly identified as ZM0 at a rate of 50%; ZM0 and ZMM were misidentified as ZM1 at a rate of 18.75%; ZM0, ZM1, and ZMM were misidentified as ZM2 at a rate of 100%; ZM1 and ZM2 were misidentified as ZMM at a rate of 50%; and ZMM was misidentified as CK at a rate of 11.10%.

Classification Using Vegetation Indices

A Random Forest (RF) classification model was established using vegetation indices. The error rate for the number of trees (ntree) indicated that when ntree is set to 600, the model’s internal error rate stabilizes, so ntree was set to 600. The variable importance plot of the RF model (Figure 10) shows that the out-of-bag (OOB) error rate is 21.74%. The RF model’s variable importance indices, ranked from highest to lowest, are NDVI₆₇₀, SR (simple ratio), TCARI (transformed chlorophyll absorption in reflectance index), EVI (enhanced vegetation index), and NDVI₇₀₅. The Gini coefficients of the RF model, ranked from highest to lowest, are TCARI, NDVI₆₇₀, SR, EVI, and NDVI₇₀₅. Both the SR and TCARI are identified as the variables with the highest importance indices and Gini coefficients.

When using vegetation indices for classification modeling and testing on the existing rodent mound data (

Table 8. Classification and error rates of rat mound spectra in different years.

	ZM0	ZM1	ZM2	ZMM	CK	Error Rate
ZM0	4	4	0	0	0	0.5
ZM1	1	14	1	0	0	0.125
ZM2	0	1	3	1	0	0.4
ZMM	0	0	2	5	1	0.375
CK	0	0	0	0	9	0

), the following misclassifications were observed: ZM1 was incorrectly identified as ZM0 at a rate of 50.0%; ZM0 and ZM2 were misidentified as ZM1 at a rate of 12.5%; ZM1 and ZMM was misidentified as ZM2 at a rate of 40.0%; ZM2 and CK were misidentified as ZMM at a rate of 37.5%; and CK had no misclassifications.

After comparing the models and validating them with actual measured data, it was found that the classification model built using vegetation indices (with an error rate of 21.74%) performed better than the classification model built solely using vegetation characteristic indicators (with an error rate of 36.96%).

4. Discussion

4.1. Vegetation Characteristics of Rodent Mounds of Different Ages

Niu et al. [4] demonstrated that as the age of rodent mounds increases, there is a notable enhancement in species diversity and biomass within alpine meadow plant communities. Hu [33] and Xiang [5] discovered that while rodent mounds do not substantially disrupt the overall structure and function of plant communities, they induce localized alterations in community composition. These modifications stimulate the renewal and vitality of plant communities. This study illustrates that, as rodent mounds mature, indices such as Pielou’s evenness, Patrick’s diversity, plant height, biomass, and total vegetation cover all show an upward trend. This trend suggests that the formation and persistence of rodent mounds positively influence vegetation recovery. The biomass of perennial grasses, sedges, and persistent weeds rises with the natural recovery age of rodent mounds, likely due to the fact that rodent mounds offer new habitats and resources, fostering growth and biomass accumulation in these plants [4]. Ye et al. [34] have shown that as rodent mounds age, both the quantity and variety of aboveground vegetation on the mounds steadily increase, particularly the species richness of Poaceae and perennial weeds [35]. Over time, the composition of communities and the degree of species co-occurrence also grow [36], indicating that the development and persistence of rodent mounds enhance the complexity and stability of plant communities. Thus, rodent mounds hold significant ecological implications for alpine meadow ecosystems, promoting increased diversity and biomass within plant communities and contributing to the maintenance and enhancement of ecosystem stability and vitality.

4.2. Original Spectral Characteristics of Rodent Mounds of Different Ages

Green plants are uniquely identifiable from other substrates by their distinct spectral reflectance characteristics, which are intricately linked to their growth, development, and overall health [37]. Specifically, the chlorophyll in plants exhibits a pronounced sensitivity to spectral responses in the visible light spectrum, absorbing light primarily at blue (450 nm) and red (650 nm) wavelengths while displaying an absorption peak around 550 nm [38]. In this study, rodent mounds at various stages of succession reveal differing spectral reflectance patterns. Notably, ZM0, characterized by low vegetation coverage, presents a spectral profile resembling that of soil, while the remaining mounds demonstrate varying degrees of the “green peak” phenomenon at around 550 nm. Additionally, close to 760 nm, all stages of succession, except for ZM0, exhibit the plant-specific “red-edge” phenomenon. Although the overall trend of this red-edge phenomenon does not show significant differences, variations in reflectance (REF) are observed across different treatments. Within the near-infrared wavelength range (750–970 nm), the REF exhibits a gradual increase, maintaining high levels across all five succession stages, in line with spectral findings on five common alpine plants reported by Liu et al. [39]. However, the reflectance of the raw spectrum may introduce errors when determining the age of the rodent mounds and therefore cannot be directly used for age classification. This highlights the necessity of calculating spectral indices to more accurately assess the succession stages of rodent mounds, which is of considerable importance for ecological restoration and land management practices.

4.3. Hyperspectral Classification of Rodent Mounds of Different Ages

A comprehensive comparison of classification models for rodent mounds indicates that the Random Forest algorithm is particularly effective, owing to its robust performance with hyperspectral data and vegetation indices. Zhang et al. [40] demonstrated that the Random Forest model significantly outperformed the multivariate regression model in simulating variations in the Normalized Difference Vegetation Index (NDVI) and understanding the underlying driving forces on the northern slope of the Tianshan Mountains. Within the Random Forest framework, importance analysis has shown that the Transformed Chlorophyll Absorption in Reflectance Index (TCARI) plays a vital role in estimating key parameters such as Patrick’s index, Height, and Total coverage. This significance is related to how factors like plant height and density influence the TCARI [41], which in turn affects the corresponding values of Patrick’s index, height, and Total coverage [42]. Moreover, the simple ratio (SR) is crucial for estimating biomass, as it is sensitive to surface vegetation density; typically, sparser vegetation results in a higher SR [31]. Consequently, the selected vegetation indices, the SR and TCARI, can effectively distinguish the ages of rodent mounds, reflecting the impact of aboveground vegetation and biomass on vegetation coverage [43]. The variations observed between rodent mounds of different ages can primarily be attributed to differences in vegetation coverage [44]. This underscores the importance of utilizing these indices in ecological assessments and rodent mound age classification. The classification model for rodent mound ages established using vegetation characteristic indicators with Random Forest exhibited an out-of-bag (OOB) error rate of 39.13%, with significant classification errors for the ZM2 stage. This was due to non-significant differences in the indices of Pielou’s index, Patrick’s index, Height, and Biomass between the ZM2 stage and the ZM1 and ZMM stages, leading to misclassifications. Another Random Forest model, established using vegetation indices, demonstrated an improved OOB error rate of 21.74%, with the highest classification error for ZM0. This error is attributed to the minimal difference in aboveground plant species and cover between the ZM0 and ZM1 stages. The OOB data provide an effective means of cross-validation, helping to verify the model’s stability and reliability [45]. Validation with the actual detected data indicates that modeling with vegetation indices yields better classification results, capable of distinguishing between rodent mounds of different ages.

5. Conclusions

(1)

The formation and presence of rodent mounds hold significant ecological importance for alpine meadow ecosystems. They not only promote increased plant community diversity and biomass but also contribute to the maintenance and enhancement of ecosystem stability and vitality.

(2)

The vegetation indices SR (simple ratio) and TCARI (transformed chlorophyll absorption in reflectance index) were identified as having the most significant impact on the classification of rodent mounds.

(3)

By employing vegetation indices in Random Forest modeling to analyze the age of rodent mounds, the precision of remote sensing interpretation of rodent mound ages is enhanced. This improves both the accuracy and stability of classification, facilitating faster analysis for determining the ages of rodent mounds and enabling the formulation of more targeted restoration measures.