Long-Term Assessment of Post-Mining Spectral Recovery Patterns: Integrating Disturbance Timing, Land-Surface Transitions, and Benchmark-Relative Spectral Closure

Highlights

What are the main findings?

• A 22-year Landsat-based assessment revealed that spectral recovery pathways characterized only 41.8% of mine-affected pixels, with the remaining area stabilizing as non-vegetated surfaces or persisting under active disturbance.
• Mean benchmark-relative spectral recovery (RSRI) in the vegetated-recovery group reached only 0.309, far below the local stable-vegetation reference, indicating that greening does not imply spectral convergence.

What are the implications of the main findings?

• Vegetation greenness trends alone overestimate spectral convergence toward stable reference conditions, highlighting the need for benchmark-relative metrics in post-mining rehabilitation assessment.
• The RSRI framework, requiring no pre-mining baseline data, offers a transferable complement to conventional greening indices for retrospective mine monitoring using open-access Landsat archives.

Abstract

Single-index greening trends can misrepresent post-mining recovery because they do not show whether disturbed surfaces are converging toward the spectral conditions of nearby stable vegetation. Here, we present a 22-year (2003–2024) Landsat-based assessment of the Nannihu molybdenum mine (Henan, China) by combining LandTrendr-based disturbance and recovery timing from annual NBR series with a benchmark-relative spectral recovery index (RSRI) and five-epoch random forest land-surface classification used as contextual support. The classifier was trained on 2024 samples and transferred to earlier epochs without independent validation at each epoch. Historical class labels should therefore be treated as approximate contextual support. A five-type recovery pathway typology showed that only 41.8% of mine-affected pixels followed vegetated recovery pathways, while 28.2% stabilized as non-vegetated surfaces and 25.0% remained under persistent disturbance. Even the combined vegetation recovery type had a mean RSRI of only 0.309 (SD = 0.143), suggesting that greening alone does not imply close benchmark-relative spectral proximity to the local stable-vegetation reference. Disturbance magnitude was the feature most strongly associated with RSRI variation (XGBoost SHAP mean, |SHAP| = 0.075). The RSRI quantifies benchmark-relative spectral proximity using local stable-vegetation benchmarks, and it does not measure species composition, biomass, or ecosystem function. This site-specific case study indicates that benchmark-relative spectral assessment can complement conventional greening metrics in retrospective mine monitoring using open-access Landsat archives, with field validation the natural next step toward linking these spectral findings to ecological or functional recovery.

1. Introduction

Mining activities represent one of the most intensive forms of land-surface disturbance globally, with open-pit operations altering topography, soil structure, hydrological regimes, and vegetation cover [1,2]. Post-mining vegetation recovery is critical for restoring ecosystem services, including carbon sequestration, erosion control, and biodiversity support [3,4]. However, the transition from bare excavated surfaces to more stable vegetated cover is often slow, nonlinear, and path-dependent, frequently spanning decades rather than years [5,6]. This timescale requires long-term (>10 years) monitoring frameworks capable of capturing the full recovery process, yet such assessments remain rare in the mine remote sensing literature [7].

Satellite remote sensing has become the primary tool for large-scale mine-recovery monitoring, with the Landsat archive (1984–present) providing the temporal depth essential for multi-decadal assessments [8,9]. Early mine remote sensing studies relied predominantly on single-image classification of land-cover types [10], while more recent work has adopted time-series approaches that track vegetation index trajectories over time [11,12]. The LandTrendr algorithm [13], originally developed for forest disturbance and recovery mapping, applies temporal segmentation to annual Landsat series to identify breakpoints associated with disturbance onset, disturbance magnitude, and subsequent recovery trends. Although LandTrendr has been extensively validated in forest ecosystems [14,15], its application to mine landscapes—characterized by abrupt, sustained surface removal rather than gradual canopy change—remains limited [16,17,18].

An unresolved limitation in current mine-recovery assessment is the frequent reliance on a single vegetation index, most commonly the Normalized Difference Vegetation Index (NDVI). While NDVI effectively captures photosynthetically active vegetation greenness, it is relatively insensitive to changes in non-photosynthetic vegetation, bare soil exposure, and surface moisture—all of which are relevant surface signals in mine landscapes undergoing structural surface transitions [19,20]. Existing studies therefore often infer recovery from increasing greenness alone. Yet in mine landscapes, greening may occur without spectral convergence toward stable-vegetation reference conditions. Post-mining surfaces can become greener while remaining spectrally distinct from nearby stable vegetation because of differences in substrate, surface moisture, or vegetation structure. The Normalized Burn Ratio (NBR), which incorporates the shortwave infrared band sensitive to soil and non-photosynthetic vegetation, has demonstrated superior sensitivity to mining-induced surface changes [21,22]. A benchmark-relative measure of spectral proximity addresses this limitation by quantifying how closely a recovering pixel approaches the spectral signature of local stable vegetation, yet it remains underexplored in mine rehabilitation studies [23,24].

A further limitation of existing mine-recovery studies is that they rarely examine three complementary aspects of recovery together: disturbance timing and magnitude, land-surface state transitions, and benchmark-relative spectral proximity to local stable-vegetation conditions. A more interpretable mine-recovery assessment would integrate disturbance history, approximate land-surface transitions as contextual information, and benchmark-relative spectral status. Most studies address only one or two of these aspects [25], leaving recovery pathways harder to interpret jointly. In addition, terrain attributes and interannual climate variability have received limited systematic attention as contextual factors in mine remote sensing studies [26].

To address these gaps, we conducted a 22-year (2003–2024) retrospective remote sensing assessment of the Nannihu molybdenum mine in Henan Province, China. The analysis examined how disturbance timing, land-surface state transitions, and benchmark-relative spectral proximity jointly describe long-term recovery patterns, what pathway types emerge from these patterns, and how terrain attributes and interannual climate variability are associated with observed recovery outcomes. Methodologically, the study uses an operationalized benchmark-relative spectral recovery index (RSRI) within a reproducible framework that combines LandTrendr temporal segmentation with contextual five-epoch land-surface classification. A site-specific recovery pathway typology summarizes the heterogeneous recovery trajectories observed across the mine landscape. Within this case study, the results indicate that increasing vegetation greenness does not necessarily correspond to high benchmark-relative spectral similarity to local stable-vegetation conditions. Throughout this paper, “recovery” refers to spectral behavior relative to the stable-vegetation reference unless it is explicitly qualified as ecological recovery.

2. Materials and Methods

2.1. Study Area and Data Sources

The Nannihu molybdenum mine (33.89°–33.94°N, 111.47°–111.52°E) is located in Luanchuan County, western Henan Province, China (Figure 1). The mine footprint covers approximately 3.98 km², with active open-pit extraction operations since the 1970s and phased rehabilitation efforts targeting waste dumps and decommissioned pits. While the coordinates above delineate the core mine footprint, the analytical domain of the present study extends substantially beyond this boundary, encompassing a buffer extending approximately 5 km beyond the mine perimeter (Figure 1b). This expanded spatial extent was deliberately adopted to (1) characterize the cascading ecological effects of open-pit extraction on the surrounding montane vegetation, including disturbance gradients propagating beyond the immediate excavation boundary; and (2) provide a spatially extensive pool of undisturbed, spectrally stable reference pixels, required for normalization within the relative spectral recovery index (RSRI) framework. The region experiences a temperate continental monsoon climate, with a mean annual temperature of approximately 12 °C and mean annual precipitation of approximately 700 mm, concentrated in the summer growing season (May–September). The terrain is predominantly mountainous, with elevations ranging from approximately 800 to 2200 m above sea level.

The satellite data used consisted of Landsat Collection 2 Level-2 surface reflectance products (Thematic Mapper, Enhanced Thematic Mapper Plus, and Operational Land Imager) for the period 2003–2024 at 30 m spatial resolution, acquired through the Google Earth Engine (GEE) platform [27,28]. Ancillary datasets included the Shuttle Radar Topography Mission (SRTM) 30 m digital elevation model, ERA5-Land monthly temperature reanalysis [29], and CHIRPS v2.0 daily precipitation estimates [30]. Growing season (May–September) potential evapotranspiration (PET) was computed from ERA5-Land temperature and radiation components using the Hargreaves method [31].

2.2. Data Preprocessing and Spectral Index Computation

Annual growing season median composites [32] were generated after applying CFMask-based cloud and cloud-shadow masking [33], retaining only pixels with at least three valid observations per growing season. Sensor harmonization between Landsat TM, ETM+, and OLI was performed using the Chastain et al. coefficients [34] to reduce inter-sensor systematic differences. Topographic correction was applied using the Sun-Canopy-Sensor + C (SCS+C) method [35] with the SRTM DEM to minimize illumination effects in mountainous terrain.

Six spectral indices were computed from the harmonized, topographically corrected composites: (1) Normalized Difference Vegetation Index (NDVI) [36], (2) Normalized Burn Ratio (NBR) [37], (3) Enhanced Vegetation Index (EVI) [38], (4) Soil-Adjusted Vegetation Index (SAVI) [39], (5) Normalized Difference Moisture Index (NDMI) [40], and (6) Bare Soil Index (BSI). The NBR served as the primary index for LandTrendr temporal segmentation, while the benchmark-relative recovery index (RSRI) uses NDVI relative to a local stable-vegetation target, and all six indices contributed to land-surface classification.

2.3. Integrated Analysis Framework

2.3.1. Disturbance and Recovery Timing via LandTrendr

LandTrendr temporal segmentation [13,41] was applied to the annual NBR time series (2003–2024) for all mine-footprint pixels. Key parameter settings included the following: maximum segments = 4, spike threshold = 0.9, vertex count overshoot = 3, recovery threshold = 0.25, best-model proportion = 0.75, and p-value threshold = 0.1 (

Table 1. LandTrendr parameter settings and sensitivity analysis.

Parameter	Value	Description
maxSegments	4	Maximum number of trajectory segments; disturbance year stable (±1 yr for >93% of pixels when varied between 4 and 8)
spikeThreshold	0.9	Threshold for removing spectral spikes prior to fitting
vertexCountOvershoot	3	Allowed vertex count overshoot for best-model selection
recoveryThreshold	0.25	Minimum spectral recovery fraction required for detection
bestModelProportion	0.75	Minimum proportion of best model fit for segment acceptance
pValueThreshold	0.1	Significance level for disturbance detection; ΔNBR magnitude varied by <5% when threshold ranged from 0.01 to 0.10

Most parameters follow the default LandTrendr settings of Kennedy et al. [41]; maxSegments and pValueThreshold were adjusted for this surface-mining setting. Sensitivity was evaluated on a 200-pixel random sample by independently varying maxSegments (4–8) and pValueThreshold (0.01–0.10).

). For each pixel, the algorithm identified the year and magnitude (ΔNBR) of the most significant disturbance and, where applicable, the post-disturbance NBR linear trend slope (proxy for recovery rate) and recovery duration (interval from disturbance breakpoint to onset of the post-disturbance recovery segment, in years). Representative trajectory fits were manually inspected for quality assurance (

Table 2. LandTrendr trajectory fit quality metrics (representative sample, n = 100 pixels).

Metric	Mean	SD	Range
RMSE	0.042	0.018	0.012–0.097
$R^2$	0.84	0.09	0.61–0.96
Number of vertices	3.2	1.1	2–5

Metrics from a stratified random pixel sample. RMSE is in NBR units; $R^2$ and vertex count characterize segmentation goodness of fit.

).

2.3.2. Five-Epoch Land-Surface Classification (Contextual Support)

Random forest (RF) classification [42,43] was performed at five epochs (2003, 2008, 2013, 2018, 2024), using all six spectral indices plus three terrain variables (elevation, slope, aspect) as feature inputs. The primary reporting uses a coarse four-class scheme aggregated from an original six-class fine classification: (C1) Active/Bare Mining, (C2) Vegetated Surface (merging Vegetated Dump and Stable Vegetation), (C3) Water/Tailings, and (C4) Other Non-Forest. The original six-class fine scheme (Active Excavation, Vegetated Dump, Barren Dump/Stabilized, Stable Vegetation, Tailings/Water, Other Non-Forest), with honest F1-score reporting, is retained as an exploratory six-class breakdown only, with its summary accuracy reported in the footnote to

Table 3. Classification accuracy for the 2024 epoch (5-fold stratified cross-validation).

Scheme	OA (%)	Macro F1	C1 F1	C2 F1	C3 F1	C4 F1
Six-class (fine)	69.3	0.65	0.71	0.62	0.58	0.68
Four-class (coarse)	75.8	0.72	0.78	0.74	0.69	0.70

OA: overall accuracy; C1–C4 as defined in text. The four-class OA is obtained by aggregating the six-class predictions into the four reporting classes. Because the classifier was trained on 2024 samples and transferred to earlier epochs, per-epoch independent validation was not performed; epoch-specific maps should be interpreted as approximations consistent with the 2024 signature space. The exploratory six-class breakdown is not a primary validated product; it is provided for interpretive context only.

. Training samples for the 2024 epoch (n = 150–200 per class) were selected via stratified random sampling and manually labeled using high-resolution Google Earth imagery. The classifier trained on 2024 data was then applied to all five epochs under the assumption that spectral–structural class signatures are temporally transferable. The 2024 epoch was chosen for training because it is the only epoch with co-temporal sub-meter reference imagery (Esri WorldImagery) supporting confident manual labeling. Assigning pixel-scale historical labels for 2003–2018 would itself require an unavailable high-resolution reference. Accordingly, the five-epoch classifications are used strictly as a qualitative illustration of land-surface context, and no quantitative inference (area change, transition rate, or recovery proportion) is drawn from the historical class labels or transition counts. To bound this uncertainty, we cross-checked the transferred 2018 classification against finer-resolution Sentinel-2 (10 m) imagery at 130 stratified points (broad-class agreement of 76.2%), and a five-epoch temporal-consistency filter flagged isolated single-epoch class flips in 16.7% of mine-pixel epochs (3–6% in the interior epochs 2008–2018). Comparable independent high-resolution imagery was not available for 2008.

2.3.3. Relative Spectral Recovery Index (RSRI)

The RSRI quantifies pixel-level benchmark-relative spectral recovery as a gap-closure fraction that captures how far a pixel’s 2024 greenness has advanced from its post-disturbance minimum toward a local stable-vegetation reference:

$$\mathrm{RSRI}={\mathrm{clip}}_{[0,1]}\left({\displaystyle \frac{{\mathrm{NDVI}}_{\mathrm{pixel},2024}-{\mathrm{NDVI}}_{\mathrm{pixel},\min }}{{\mathrm{NDVI}}_{\mathrm{ref}}-{\mathrm{NDVI}}_{\mathrm{pixel},\min }}}\right),$$

(1)

where ${\mathrm{NDVI}}_{\mathrm{pixel},\min }$ is the pixel’s minimum annual NDVI over 2003–2024 (its post-disturbance trough) and ${\mathrm{NDVI}}_{\mathrm{ref}}$ is the local stable-vegetation target, taken as the median multi-year mean NDVI of reference pixels computed within elevation–aspect strata to control for the topographic background. NBR is retained for LandTrendr disturbance detection, whereas the benchmark-relative recovery measure uses NDVI. RSRI values approaching 1 indicate spectral proximity to reference conditions, whereas values ≪ 1 indicate persistent divergence. RSRI measures benchmark-relative spectral proximity to the local stable-vegetation envelope, and it does not capture species composition, biomass, habitat quality, or ecosystem function. We benchmark this gap-closure scaling against simple percentile and z-score normalizations in Section 3.3, and report its sensitivity to the reference target and benchmark pixels in

Table 4. RSRI sensitivity to the choice of reference target and to benchmark-pixel selection.

Reference Target	Mean RSRI	Pearson r vs. Adopted
Stratified (adopted)	0.189	1.000
Global median	0.188	0.998
Global mean	0.193	0.998
Global p75	0.182	0.998

RSRI is also robust to benchmark-pixel selection: bootstrapping the reference pool (500 resamples, 50% each) gave pixel-level correlations with median $r=0.998$ with the adopted index.

.

2.3.4. Recovery Pathway Typology

K-means clustering ($k=5$) was applied to three pixel-level variables: (1) LandTrendr-detected disturbance year, (2) disturbance magnitude (ΔNBR), and (3) post-disturbance NBR linear trend slope. The choice of $k=5$ was guided by interpretability, capturing the major pathway shapes (closed recovery, partial recovery, stabilized non-vegetated surface, persistent disturbance, non-forest/water) without over-fragmenting the sample. This typology is presented as a heuristic descriptive tool rather than a uniquely optimal partition. The resulting five types were: type I (Closed Recovery), type II (Partial Recovery), type III (Stabilized Surface), type IV (Persistent Disturbance), and type V (Non-Forest/Water). Given the small sample size of type I (n = 55 pixels) and the substantive similarity of types I and II as vegetation-oriented spectral recovery pathways, these two types were merged into “Combined Vegetation Recovery” (n = 4493 pixels) for primary analysis. The full statistics of the five types are reported in Section 3.3.

2.3.5. Terrain and Climate Contextual Analyses

Pixel-level associations between RSRI and terrain attributes (elevation, slope, aspect–northness, and aspect–eastness) were quantified using Pearson correlation. As a supplementary check, the relative importance of these terrain variables for RSRI was additionally summarised with a random forest regressor [42]. Distance from each pixel centroid to the mine center was included as an additional spatial predictor in the XGBoost SHAP feature importance analysis [44,45,46] (full feature list in

Table A1. XGBoost SHAP full feature importance (all 8 features, ordered by mean |SHAP|).

Rank	Feature	Mean |SHAP|	Relative %	Category
1	Disturbance magnitude (ΔNBR)	0.0745	33.0	Disturbance history
2	Years since disturbance (to 2024)	0.0458	20.3	Disturbance history
3	Elevation (m)	0.0232	10.3	Terrain
4	Distance to mine center (m)	0.0227	10.1	Mine spatial
5	LandTrendr disturbance detected	0.0203	9.0	Disturbance history
6	Northness (cos aspect)	0.0143	6.3	Terrain
7	Slope (°)	0.0139	6.2	Terrain
8	Eastness (sin aspect)	0.0110	4.9	Terrain
	Total	0.2257	100.0

Values from spatial block 5-fold CV training run (n = 5 folds, random_state = 42). Terrain features (ranks 3, 6, 7, 8) collectively contribute ≈ 27.7%; disturbance-related features (ranks 1, 2, 5) collectively contribute ≈ 62.3%. “Mine spatial” denotes an anthropogenic-geometric attribute (Euclidean distance from pixel centroid to mine centre). SHAP values indicate feature association, not causal attribution. Because the recovery index and disturbance magnitude share a spectral lineage, the SHAP associations are interpreted descriptively, not as independent mechanistic effects.

, Appendix A). These associations are reported as exploratory only. Because the RSRI reference target is computed within elevation–aspect strata, topographic background is already partly controlled in the index. Residual RSRI–terrain associations are therefore expected to be weak and are not interpreted as topographic controls. Interannual growing season (May–September) climate variables—precipitation, mean temperature, and potential evapotranspiration (PET)—were computed for the mine region and correlated (Pearson r) with annual mean NBR and BSI. The climate analysis was explicitly framed as contextual background to interpret interannual variability, and no causal attribution was attempted.

2.4. Three-Dimensional Integration Visualization

A three-dimensional scatter plot jointly visualizes three analytical dimensions: disturbance timing (X-axis, LandTrendr disturbance year), final land-surface state (color, 2024 epoch classification), and RSRI value (bubble size). This visualization reveals how recovery outcomes co-vary with when disturbance occurred and what surface class resulted, providing an integrative perspective that no single dimension captures.

3. Results

3.1. Long-Term Spectral Trajectories and Disturbance Timing

LandTrendr temporal segmentation of the annual NBR time series (2003–2024) identified a significant disturbance for 9942 of the 10,756 mine-footprint pixels analyzed. The remaining 814 pixels, which lacked a detected disturbance of sufficient magnitude (predominantly stable, sparsely vegetated, or water surfaces), were assigned to recovery types using the class- and spectral-based fallback rule described in Section 2.3.2. The spatial distribution of LandTrendr-detected disturbance years reveals three major disturbance phases: an early phase centered around 2005–2008, corresponding to mine expansion; a peak disturbance period during 2011–2015, coinciding with intensified extraction; and a declining disturbance frequency after 2016 (Figure 2). The mean disturbance magnitude across all disturbed pixels was ΔNBR = −0.237 (SD = 0.146), with impact magnitudes ranging from −0.024 to −0.682. Disturbance timing and magnitude alone, however, do not reveal whether disturbed surfaces have achieved spectral convergence toward reference conditions. That question is addressed in Section 3.3 through the benchmark-relative RSRI.

Class-level spectral trajectories reveal substantial heterogeneity in recovery patterns across land-surface classes (Figure 3). Pixels classified as Active Excavation in 2024 displayed persistently low NBR values (mean = −0.04, SD = 0.08) throughout the study period, with no discernible recovery trend. Pixels classified as Vegetated Dump showed a characteristic trajectory, with an initial decline during the disturbance phase (mean −0.15 ΔNBR), followed by a partial upward recovery in the post-disturbance segment (median post-disturbance NBR slope $\approx +0.008$ yr⁻¹, positive in 86% of these pixels). Over the full 2003–2024 series, however, the net Theil–Sen trend remained negative (NBR −0.009 yr⁻¹, p < 0.001), and 2024 NBR values stayed below pre-disturbance levels for 87% of these pixels. Stable-Vegetation reference pixels maintained high NBR values (mean = 0.59, SD = 0.03) with minimal interannual fluctuation. The estimated recovery duration—defined as the interval between the LandTrendr disturbance breakpoint and the onset of the fitted post-disturbance recovery segment—ranged from 5 to 17 years (median 11 years), confirming the slow, multi-decadal nature of mine recovery.

3.2. Five-Epoch Land-Surface Composition and Transitions (Contextual Support)

Five-epoch land-surface classification revealed progressive structural changes in the mine landscape composition, serving as contextual support only for interpreting the RSRI spatial patterns. The primary inferential evidence comes from LandTrendr-derived disturbance history and the benchmark-relative RSRI (Figure 4). The Active/Bare Mining class (C1) expanded from 0.81 km² (2003) to 1.45 km² (2013), before declining to 0.92 km² (2024). Conversely, the Vegetated Surface class (C2) contracted from 2.47 km² (2003) to 1.78 km² (2013), then partially recovered to 2.14 km² (2024). The Water/Tailings class (C3) remained relatively stable (0.22–0.31 km²), while Other Non-Forest (C4) fluctuated between 0.43 and 0.61 km². As noted in Section 2.3.2, all epoch-specific class areas are approximate contextualizations derived from a 2024-trained classifier and should not be interpreted as independently validated per-epoch reconstructions.

The five-epoch transition matrix (

Table 5. Top-10 five-epoch land-surface transitions by pixel count (four-class scheme). Counts are approximate and are provided as qualitative context only, because the historical classifications were not independently validated per epoch.

Transition Path	Pixel Count	Fraction (%)
C1 → C2	1847	17.2
C2 → C1	1623	15.1
C2 (persistent)	1412	13.1
C1 (persistent)	1098	10.2
C1 → C1 → C2	876	8.1
C2 → C1 → C2	654	6.1
C1 → C4	521	4.8
C4 → C2	489	4.5
C2 → C4	468	4.4
C3 (persistent)	387	3.6

C1: Active/Bare Mining; C2: Vegetated Surface; C3: Water/Tailings; C4: Other Non-Forest.

) identifies the top-10 land-surface transitions, accounting for approximately 87% of all pixel-level class changes. The most frequently labeled transitions, reported here for qualitative illustration only and not as validated counts, were Active/Bare Mining to Vegetated Surface (C1 → C2) and the reverse C2 → C1, broadly consistent with alternating revegetation and mine-expansion phases, together with stabilization of barren dump into vegetated dump within the C2 class. Zone-level annual spectral trajectories (Figure 5) show how NDVI and NBR evolved for the three main surface zones relative to growing season climate conditions. Because the classifier was trained on 2024 data and transferred to earlier epochs, these transitions represent the best approximation of surface-state evolution consistent with the 2024 signature space rather than independently validated per-epoch changes.

To quantify classification consistency across epochs, the 2024-trained classifier was applied to all five historical epochs and evaluated against the 2024 training labels. The 2024 epoch achieved the highest 5-fold cross-validated overall accuracy (OA = 69.3%, six-class scheme), and the transfer-based signature-space agreement declined to 61.2% by 2003, without independent per-epoch validation (

Table 6. Per-epoch internal consistency of the 2024-trained classifier (agreement with the 2024 signature space; not historical classification accuracy).

Epoch	Signature-Space Agreement (%)	Signature-Space Macro F1
2003	61.2	0.58
2008	64.5	0.61
2013	66.8	0.63
2018	68.1	0.64
2024	69.3	0.65

Accuracy estimated by applying the 2024-trained classifier to each epoch and evaluating against 2024 training labels (5-fold cross-validation). These metrics reflect internal consistency relative to the 2024 spectral–structural signature space; because no independent per-epoch ground-truth validation was performed, they do not represent true historical accuracy.

). These values are internal consistency metrics—not independent historical ground-truth validation—reflecting spectral–structural match relative to the 2024 training domain rather than true historical classification performance.

3.3. Benchmark-Relative Spectral Recovery Status: The Core Finding

The RSRI spatial distribution for 2024 shows the central empirical result of this study, namely, that surface greening does not imply close benchmark-relative spectral proximity to the local stable-vegetation reference (Figure 6). Across all mine-affected pixels, the mean RSRI was 0.178 (SD = 0.156), with values ranging from 0.001 to 0.740. Only 55 pixels (0.51% of mine-affected area) achieved RSRI ≥ 0.65, concentrated in older disturbance zones on vegetated dumps.

Merged-typology RSRI statistics show substantial differences across pathway types (

Table 7. RSRI statistics by merged recovery typology (2024).

Typology	Pixel Count	Mean RSRI	SD	Range
Combined Veg. Recovery (I+II)	4493 (41.8%)	0.309	0.143	0.028–0.740
—Type I (closed)	55 (0.5%)	0.684	0.026	0.651–0.740
—Type II (partial)	4438 (41.3%)	0.305	0.141	0.028–0.598
Stabilized (type III)	3038 (28.2%)	0.085	0.067	0.005–0.306
Persistent Disturbance (type IV)	2686 (25.0%)	0.093	0.119	0.001–0.523
Non-Forest/Water (type V)	539 (5.0%)	N/A	N/A	N/A

Type I and type II merged as “Combined Vegetation Recovery” for primary reporting. RSRI is not defined for Water/Tailings surfaces (spectral endpoint applies to vegetated land only); type V is reported for area completeness.

). The Combined Vegetation Recovery type (types I+II, 4493 pixels) achieved a mean RSRI of 0.309 (SD = 0.143). Type III (Stabilized Surface, 3038 pixels) had a mean RSRI of 0.085 (SD = 0.067), while type IV (Persistent Disturbance, 2686 pixels) showed a mean RSRI of 0.093 (SD = 0.119). The small type I subset within Combined Vegetation Recovery (n = 55) achieved a mean RSRI of 0.684 (SD = 0.026), approaching but not reaching reference conditions (full statistics of five types given in Table 7).

The three-dimensional integration visualization (Figure 7) reveals that pixels disturbed earlier (pre-2010) and classified as Vegetated Surface in 2024 tend to exhibit higher RSRI values, but even these older, vegetated pixels rarely achieve RSRI > 0.5. Pixels disturbed after 2015 and classified as Active/Bare Mining in 2024 cluster near RSRI values below 0.10, consistent with ongoing surface disturbance.

Figure 8 illustrates the positive association between disturbance timing and RSRI, with pixels disturbed in earlier phases (pre-2010) clustering at higher RSRI values than more recently disturbed pixels. The recovery pathway type nonetheless modulates this relationship substantially, indicating that disturbance timing alone does not determine the spectral recovery outcome. The RSRI framework was also robust to the choice of reference target and benchmark pixels (Table 4). Alternative reference targets (global median, mean, and 75th percentile) yielded Pearson correlations of r ≥ 0.998 with the adopted index, and bootstrapping the reference pool (500 resamples, 50% each) gave pixel-level correlations with a median r of 0.998. These results confirm that RSRI rankings and spatial patterns are stable across reasonable benchmark choices.

The stable-vegetation reference set comprises N ≈ 88,900 pixels. Independently of the classification labels, these pixels are spectrally stable across 2003–2024 (pixel-wise temporal coefficient of variation, median ≈ 0.10; a small Theil–Sen trend of +0.004 yr⁻¹, i.e., mild background greening with no abrupt change). Their terrain envelope is an elevation of 960–1670 m (mean: 1365 m) and mean slope ≈ 25°. Spatially, they are distributed across the stable-vegetation areas of the approximately 5 km buffer surrounding the mine footprint (Figure 1b). Because the reference is anchored to this spectrally stable set, the RSRI denominator does not depend on any single-epoch classification.

To benchmark the gap-closure scaling, we compared the adopted index against a simple percentile normalization, $(\mathrm{NDVI}-p_5)/(p_{95}-p_5)$, and a z-score, $(\mathrm{NDVI}-{\mathrm{mean}}_{\mathrm{ref}})/{\mathrm{std}}_{\mathrm{ref}}$, computed from the same reference set. Because the reference distribution is strongly left-skewed (skewness ≈ −2.9) and mine surfaces lie far below it, the percentile normalization saturates essentially all mine pixels at zero (no discrimination) and the z-score places essentially all of them outside [0, 1] (unbounded, non-comparable across sites). The adopted gap-closure formulation is bounded to [0, 1] and retains full discrimination among recovery types (

Table 8. Comparison of normalization schemes for the recovery index.

Variant	Bounded [0, 1]	Saturated/Out of Range	Discriminates Types
Gap-closure (adopted)	Yes	0	Yes
Percentile $(\mathrm{NDVI}-p_5)/(p_{95}-p_5)$	Yes	∼100% saturated at 0	No
z-score $(\mathrm{NDVI}-\mathrm{mean})/\mathrm{std}$	No	∼100% outside [0, 1]	Partial

).

3.4. Terrain and Climate Associations (Contextual Analysis)

The RSRI reference target is computed within strata defined by elevation and aspect, so topographic background is already partly controlled in the index, and any residual associations between RSRI and terrain are weak by design. Pearson correlations between RSRI and terrain attributes were consistently weak: RSRI–elevation: r = −0.118 (p < 0.001), RSRI–slope: r = −0.079 (p < 0.001); and RSRI–northness (cosine of aspect): r = 0.031 (p = 0.047). The XGBoost SHAP analysis (Figure 9) identified disturbance magnitude (ΔNBR, mean |SHAP| = 0.075) and years since disturbance (mean |SHAP| = 0.046) as the two dominant features associated with RSRI variation, collectively accounting for approximately 53% of total SHAP importance. A spatial proximity feature, distance to mine center (rank 4, mean |SHAP| = 0.023, 10.1%), ranked among the top-5 predictors (see Table A1, Appendix A), suggesting that spatial position within the mine was associated with RSRI variation beyond what terrain attributes alone captured. Terrain features ranked lower, with elevation (mean |SHAP| = 0.023, rank 3), northness (mean |SHAP| = 0.014, rank 6), slope (mean |SHAP| = 0.014, rank 7), and eastness (mean |SHAP| = 0.011, rank 8) collectively contributing approximately 27.7% of total SHAP importance (full ranked values in Table A1, Appendix A). The weak RSRI–elevation correlation (Figure 10) indicates that RSRI captures mine-disturbance signals rather than natural vegetation gradients. Because disturbance magnitude and the recovery index share a spectral lineage, the SHAP associations are interpreted descriptively rather than as independent mechanistic effects.

Interannual climate covariation analysis (growing season climate context in Figure 5b) was included as exploratory background context rather than as a test of climatic control. Growing season PET showed a strong positive correlation with annual mean NBR (r = 0.703, p < 0.001), while the precipitation (r = 0.137, p = 0.348) and temperature (r = −0.152, p = 0.283) correlations were weak and non-significant. The PET–BSI correlation was negative and moderate (r = −0.669, p < 0.001). These interannual associations do not adjust for serial autocorrelation, legacy disturbance effects, or management interventions, and are presented strictly as descriptive temporal context.

4. Discussion

4.1. Multi-Pathway Spectral Recovery and the Limits of Greening

The recovery pathway typology reveals that post-mining vegetation recovery at Nannihu is not a single monotonic trajectory but a branching process. The Combined Vegetation Recovery type (41.8%) represents pixels where some level of spectral recovery is observable, the Stabilized type (28.2%) represents pixels that have ceased active disturbance but show minimal vegetation colonization, and the Persistent Disturbance type (25.0%) represents areas of ongoing surface alteration.

The central finding of this study is the substantial discrepancy between NDVI-based greening signals and benchmark-relative spectral proximity to the stable-vegetation reference, which is the key distinction motivating the RSRI framework. While the post-disturbance recovery segments of the Vegetated Dump class show a gradual upward slope (median ≈ +0.008 yr⁻¹ in NBR and +0.011 in NDVI), the multi-decadal series trend is net-negative (NBR: −0.009, NDVI: −0.005 yr⁻¹). Nevertheless, the mean RSRI for the Combined Vegetation Recovery type remained at only 0.309, far from the reference condition (RSRI = 1.0). Here, NDVI serves as the conventional greening comparator, the recovery trajectory analyses report both NDVI and NBR, and RSRI itself is computed as an NDVI gap-closure relative to the stable-vegetation reference. For this class, both indices register gradual improvement within the post-disturbance segment, while the full-series trends are net-negative. This discrepancy arises because NDVI is primarily sensitive to chlorophyll absorption in the red band and near-infrared reflectance, whereas NBR additionally captures shortwave infrared reflectance that is sensitive to bare soil exposure, non-photosynthetic vegetation, and surface moisture [16,37]. Pixels that appear as “greening” in NDVI may still exhibit substantial soil background influence in NBR, indicating that the surface has not achieved the spectral structure of mature, stable vegetation [23]. In restoration ecology, this phenomenon parallels the “green desert” concept, where vegetation establishment does not restore ecosystem structure or function [47,48]. In particular, pixels that green in NDVI yet retain a low RSRI are consistent with early-successional, sparse, or structurally atypical (and possibly non-native) cover that remains spectrally distinct from the mature reference. Confirming species identity, however, requires field data. The RSRI framework captures this distinction by contextualizing pixel-level spectral condition against local stable-vegetation benchmarks rather than against absolute threshold values.

Land-surface transitions provide contextual illustration of how surface states shifted over time, complementing the continuous RSRI time series. The transition C1 → C2 (1847 pixels) is consistent with a gradual revegetation process, but because the classifier was trained on 2024 samples and transferred without per-epoch validation, these counts are approximations rather than validated reconstructions. The exploratory six-class breakdown (not a primary validated product; see Table 3 footnote) is consistent with the interpretation that much of the C1 → C2 transition flows into Vegetated Dump rather than Stable Vegetation, reinforcing the interpretation of partial rather than full structural recovery [25,49]. The primary evidentiary weight therefore rests on the LandTrendr disturbance timing and benchmark-relative RSRI. Transition matrices offer qualitative contextual support only.

4.2. Disturbance History Dominates Recovery Variation; Terrain and Climate Are Contextual

The XGBoost SHAP analysis identifies disturbance magnitude (ΔNBR) and years since disturbance as the features most strongly associated with RSRI variation, with these two features alone accounting for approximately 53% of total SHAP importance. Including the binary LandTrendr-detection feature at rank 5, all three disturbance-related features collectively contribute 62.3% (see Table A1, Appendix A). This finding aligns with the conceptual model that the initial severity of surface alteration sets an upper bound on subsequent spectral recovery [50,51]. Terrain features contributed comparatively little to the SHAP importance. This weak contribution is expected by design, as the RSRI reference target is computed within strata defined by elevation and aspect, leaving much of the topographic background already controlled in the index. The residual associations between RSRI and terrain are correspondingly weak (all $\left|r\right|<0.12$) and do not represent definitive topographic controls. Because disturbance magnitude and the recovery index share a spectral lineage, the SHAP associations are interpreted descriptively rather than as independent mechanistic effects.

The strong interannual PET–NBR correlation (r = 0.703) indicates that annual spectral conditions covaried with evaporative demand at the site scale. This association was stronger than the precipitation–NBR (r = 0.137) or temperature–NBR (r = −0.152) correlations. This exploratory relationship does not establish climatic control over recovery trajectories and is reported only as interannual context alongside the dominant mining-disturbance signal [26,52,53].

4.3. Interpretive Limits and Transferability of the Framework

The main methodological addition of this study is an RSRI that provides a benchmark-relative alternative to absolute NDVI-based recovery thresholds by quantifying spectral proximity to local stable-vegetation reference distributions—without requiring pre-mining baseline data, which is rarely available for older mines. The integrated timing–proximity workflow, with contextual land-surface transition support, represents a transferable descriptive approach applicable to other mine sites using open-access Landsat archives. Empirically, to our knowledge, this study represents one of the longer Landsat-based retrospective assessments reported for a metal-mine case [17,18,54].

Several limitations warrant explicit acknowledgment. This is a single-site case study (N = 1 mine). The observed recovery patterns are site-specific and cannot be generalized without multi-site comparative validation. The 30 m Landsat spatial resolution introduces sub-pixel mixing effects at boundaries between surface classes. Classification uncertainty means pixel-level transitions may reflect classification error rather than genuine surface change, and the 2024 5-fold CV accuracy was 69.3% for the six-class scheme and approximately 76% for the four-class scheme. Most importantly, this study did not include field ecological measurements, and spectral recovery (RSRI) does not directly equate to species composition, biodiversity, biomass, or soil function recovery [3,4]. Field-based ecological ground-truthing is essential for translating spectral indicators into ecological meaning. Ground-based studies at comparable coal mines show that soil properties, plant diversity, and soil-biodiversity-driven multifunctionality are decisive indicators of ecological recovery that spectral indices do not capture [55,56]. For management use, RSRI is best treated as a screening and triage indicator rather than a restoration certificate. Low-to-moderate values can prioritize sites for field assessment, high values warrant but do not replace field confirmation of ecological function, and RSRI should not serve alone as a sign-off criterion for completed restoration. Future work should extend the analysis to additional mine types and climatic settings to test framework transferability [1,7], integrate field ecological data to ground-truth RSRI against vegetation structure, soil properties, and biodiversity metrics, and apply higher-resolution imagery (e.g., PlanetScope 3 m or WorldView 0.5 m) to resolve within-dump heterogeneity masked by Landsat’s 30 m pixel.

5. Conclusions

This 22-year (2003–2024) retrospective remote sensing assessment of the Nannihu molybdenum mine indicates that post-mining recovery followed multiple spectral pathways rather than a single uniform trajectory, with persistent divergence from stable-vegetation reference conditions across the majority of mine-affected pixels. The primary evidentiary pillars are LandTrendr-derived disturbance timing and the benchmark-relative RSRI. Five-epoch land-surface classification and transition matrices provide contextual support, but were derived from a classifier trained on 2024 samples and transferred to earlier epochs without independent per-epoch validation. Consequently, historical class labels and transition counts should be interpreted as approximations consistent with the 2024 spectral–structural signature space rather than independently validated surface-state reconstructions.

The key descriptive findings are as follows. First, spectral recovery pathway typing indicates that 41.8% of mine-affected pixels fall into the combined vegetation-recovery group, whereas 28.2% follow predominantly non-vegetated stabilization pathways, 25.0% remain in persistent-disturbance pathways, and 5.0% are grouped as type V (Non-Forest/Water and atypical surfaces). These pathway labels should be interpreted in light of the study’s contextual classification constraints. Second, vegetation greenness recovery did not imply close benchmark-relative spectral proximity in this case study, with the combined vegetation-recovery group showing a mean RSRI of 0.309 and only 55 pixels (0.51%) reaching RSRI ≥ 0.65. Third, disturbance magnitude had the largest association with spectral recovery variation, with terrain attributes showing weaker associations and climate correlations that were exploratory and non-causal.

The RSRI framework quantifies benchmark-relative spectral proximity using local stable-vegetation benchmarks, and it does not measure species composition, biomass, biodiversity, or ecosystem function. The stable-vegetation benchmark represents a local spectral comparator rather than a prescribed ecological restoration target, and RSRI values indicate proximity to local spectral conditions, not fulfillment of ecological restoration standards. It provides a benchmark-relative alternative to absolute recovery thresholds and has the practical advantage of not requiring pre-mining baseline data. However, this study is explicitly a descriptive, site-specific remote sensing assessment, and field ecological validation remains essential for translating spectral indicators into ecological meaning. More broadly, the results show that post-mining greening should not be interpreted as recovery unless disturbed surfaces are also evaluated for convergence toward local stable-vegetation spectral conditions. The workflow is transferable as a descriptive remote sensing approach, but ecological interpretation still requires field validation and broader multi-site testing. Accordingly, RSRI is most useful for prioritizing sites for field assessment rather than for certifying that restoration is complete.