The study was conducted at Manghatu Village, Gerchaolu Sumu, Zhalute Banner, Tongliao City, Inner Mongolia, China (44°62′24″N, 120°45′14″E) at an altitude of 482 m. The region features a temperate semi-arid continental climate with distinct seasons: dry, windy springs; hot summers with scarce rainfall; cool autumns with significant diurnal temperature variations; and long, cold winters. The climatic parameters described in this section were obtained from the China Meteorological Data Service Center and represent long-term averages based on a 30-year period (1995–2024). The long-term mean temperature is 5.5 °C, and the annual accumulated temperature ranges from 2400 to 2600 °C. Annual sunshine averages approximately 3000 hours, with a frost-free period of 115–130 days. Annual rainfall ranges from 300 to 400 mm, with the majority occurring between June and August, accounting for ~70% of the total precipitation. Annual evaporation exceeds 1800 mm, with an average relative humidity of 49%. The grassland is classified as a temperate mountain grassland, characterized by sandy soil with a pH of 7.6. The dominant species is Agropyron cristatum, with key associated species including Lespedeza davurica, Cleistogenes squarrosa, and Carex duriuscula.

This study investigated the effects of varying grazing intensities on grassland ecosystems, commencing in June 2025. Grazing occurred from 15 June to 15 September, during which the sheep remained in their assigned subplots without interruption. The experimental site covered a total area of 3 ha, divided into four subplots of 0.75 ha each. A randomized block design was used to establish four grazing stocking rate treatments. Each subplot was grazed by 0, 3, 6, or 9 six-month-old rams, corresponding to 0, 4, 8, or 12 sheep units per ha, respectively. These treatments corresponded to no grazing (NG), light grazing (LG), moderate grazing (MG), and severe grazing (SG) (Zhang et al., 2024b). The overview of the experimental site and the distribution of the grazing intensity treatments are shown in Figure 1.

Figure 3 illustrates the framework for monitoring grazing intensity using multi-source UAV imagery. The process mainly involves four steps: (1) data acquisition and preprocessing, (2) construction and fusion of multi-source SIs, (3) AIFS, and (4) classification modeling and accuracy evaluation. Data pre-processing includes normalization, standardization, and outlier removal. Multi-source SIs are derived from multiple features, integrating conventional and self-constructed indices to capture complementary information across data sources. An AIFS method, which combines feature importance assessment with redundancy removal, is employed to select the optimal combination of indices gradually. Finally, the performance of the optimized fused features is evaluated using various classification algorithms to achieve high accuracy and robust monitoring.

The spectral mean and variance curves under different grazing intensities are presented in Figure 4. The grey values of the three visible-light features generally increased with grazing intensity, but decreased under severe grazing. In the multispectral bands, green and red reflectance increased with grazing intensity, whereas the red-edge and near-infrared bands showed decreasing trends. However, under severe grazing, both the red-edge and near-infrared bands exhibited an increasing trend.

ANOVA results for visible and multispectral features revealed significant differences (p < 0.05) across grazing intensities, indicating that grazing intensity has a substantial impact on spectral responses. Tukey’s post-hoc test (Figure 5) revealed that the red ( R_RGB) and blue ( B_RGB) components significantly distinguished between low, medium, and severe grazing intensities, with the most substantial differences observed under extreme grazing conditions. The green ( G_RGB) component demonstrated a limited ability to separate low and medium grazing intensities, but retained significant differentiation under severe grazing conditions. For multispectral bands, the green ( G_MS) and red ( R_MS) bands showed substantial differences between low and severe grazing intensities. The red-edge ( RE_MS) and near-infrared ( NIR_MS) bands were sensitive not only to the transition between low and severe grazing, but also to the transition between medium and severe grazing, highlighting their strong ability to capture variations in vegetation status and grazing conditions. Overall, multispectral features offered more comprehensive discrimination of grazing levels than visible-light features. Specifically, the red-edge and near-infrared bands effectively captured vegetation changes under medium-to-severe grazing, whereas visible-light features were more sensitive to extreme grazing.

In this study, the classification performance of grazing intensity was evaluated using KNN, SVM, and RF based on three feature sets: visible light (RGB), multispectral (Multi), and multi-source fusion (RGB+Multi). The results are presented in Figure 6. Among single-source features, RGB achieved the highest accuracy with KNN, yielding an OA of 78.77% and a Kappa of 70.86%. The Multi-features performed best with RF, achieving an OA of 75.94% and a Kappa of 66.74%. In contrast, the multi-source fusion features markedly improved classification, achieving OA values of 83.49%, 86.79%, and 81.13%, with corresponding Kappa values of 76.62%, 81.28%, and 73.25%. These results demonstrate that integrating visible and multispectral features leverages complementary spectral information to improve discrimination among grazing intensity levels. Overall, the combination of multi-source fusion features with the SVM classifier yielded the highest accuracy and the most consistent results.

Figure 7 presents pixel-level classification results under different combinations of original features and classifiers. As shown in Figure 7a, the experimental area was divided into four grazing intensity levels: no grazing (NG), light grazing (LG), moderate grazing (MG), and severe grazing (SG). Figures 7b–d illustrate the classification maps generated by different feature combinations and classifiers. The classification map based on visible (RGB) features using the KNN classifier (Figure 7b) roughly captures the grazing intensity distribution but shows some confusion between LG and MG regions. When using multispectral (Multi) features with the KNN classifier (Figure 7c), the classification accuracy improves slightly, particularly in distinguishing NG and SG areas. However, misclassifications remain at moderate levels. In contrast, the multi-source fusion (RGB+Multi) combined with the SVM classifier (Figure 7d) produces the most accurate and spatially consistent classification, with clear boundaries between different grazing intensities that closely align with the actual distribution. This demonstrates that multi-source feature fusion effectively enhances spatial discrimination of grazing intensity and mitigates spectral confusion caused by soil–vegetation mixing in sandy grasslands.

The M-statistic was applied to assess the separability of the SIs, with results presented in Figure 8. For most indices, M values exceeded 1, indicating good separability among grazing intensity levels. Traditional visible indices (ExG, ExGR, RGBVI) and red–NIR indices (NDVI, GNDVI, SAVI, MSAVI, EVI2) generally exhibited high separability. Their maximum M values all exceeded 1, while the mean values were mainly within the range of 0.6 to 1.0. In contrast, NDRE, idx3, and several other indices showed poor separability, with maximum M values below 0.4. Notably, among the self-constructed indices, idx6, idx7, and idx14 performed best, with maximum M values exceeding 1.5. In particular, idx7 and idx14 reached extreme values above 2, demonstrating substantial advantages in capturing vegetation responses to grazing intensity and outperforming traditional indices.

Figure 9 presents the classification results of three classifiers (KNN, SVM, and RF) using different combinations of SIs. The classification accuracy of visible indices (RGB) was comparable to that of multispectral indices (Multi), with Multi showing slightly better overall performance. The results indicate that near-infrared and red-edge bands provide stronger discriminative capability for grazing level identification. Integrating RGB and Multi (RGB+Multi) markedly improved accuracy, with OA values of 84.91%, 88.68%, and 82.55% for KNN, SVM, and RF, respectively. Corresponding Kappa coefficients all exceeded 75%, indicating strong complementarity between visible and multispectral features. With the newly constructed indices (New-indices), classification performance further improved: all three classifiers achieved OA values above 87% and Kappa values above 82%. These results demonstrate that the newly proposed indices more effectively capture differences in grazing intensity and enhance class separability. Under the AllFusion condition (RGB+Multi+New-indices), results were comparable to those from RGB+Multi and New-indices. Overall, SVM showed the most consistent performance, with a maximum OA of 88.68% and a Kappa above 83%. Comparative analysis indicates that the newly constructed indices substantially improve discrimination of grazing intensity, outperforming single-band combinations, while multi-source feature fusion further enhances model robustness.

Figure 10 presents pixel-level classification maps generated from different SI sets and classifiers. Figure 10a shows the reference grazing-intensity partition (NG, LG, MG, SG). Figure 10b-f display classification maps produced using different index sets and the best-performing classifier for each set. The map produced from RGB-derived indices with RF (Figure 10b) captures the broad spatial pattern but shows misclassification between adjacent LG and MG patches. Using multispectral indices with SVM (Figure 10c) improves discrimination of NG and SG zones and reduces local speckle. The fused RGB+Multi set under SVM (Figure 10d) increases spatial continuity and sharpens boundaries. Notably, the AllFusion map (Figure 10f), which uses the full index suite, attains the best overall spatial agreement with the reference—reflecting the benefit of combining complementary information across index families. The map based on the newly constructed indices (Figure 10e) also yields high spatial consistency and particularly good detection of moderate grazing, despite using a more compact feature set. Overall, AllFusion and the new-index set both perform strongly: AllFusion slightly surpasses other sets in overall spatial agreement, while the new indices offer nearly comparable accuracy with fewer features, demonstrating a favorable trade-off between completeness and parsimony.

This study applied multiple feature selection methods for full-feature fusion, including random frog (RF) (Li et al., 2012), ReliefF (Robnik-Šikonja and Kononenko, 2003), least-squares mutual information (LSMI) (Suzuki et al., 2009), successive projections algorithm (SPA) (Araújo et al., 2001; Yu et al., 2020), elimination of uninformative variables (UVE) (Centner et al., 1996), competitive adaptive reweighted sampling (CARS) (Li et al., 2009), and AIFS. Each method used its own scoring or importance distribution to determine the selected features automatically. As shown in Table 2, the number and type of features selected by different methods varied slightly but showed overall consistency. Most methods consistently selected traditional vegetation indices (e.g., ExG, GNDVI, NDRE, RGBVI, MSAVI) and several self-constructed indices (e.g., idx7, idx8, idx9, idx14), suggesting that these features have strong discriminatory power for grazing intensity across different algorithms. Among these, idx7, idx9, and idx14 appeared most frequently across the six methods, highlighting their stability and central importance. In contrast, the RF and AIFS methods selected fewer features.

Table 3 shows that various feature selection methods exhibit distinct performances across different classifiers (KNN, SVM, RF). Overall, the AIFS method produced the best classification results. SVM achieved high accuracy, with OA and Kappa values of 92.13% and 88.99%, respectively, and also showed strong performance under KNN and RF. This result demonstrates that the method achieves optimal discrimination while requiring fewer features. In contrast, the AllFusion approach, which utilizes all features, yielded the lowest accuracy. The OA and Kappa values across all three classifiers were significantly lower than those of other methods, indicating that redundant features not only fail to improve performance but may also impair classification effectiveness. Overall, appropriate feature selection substantially improves the accuracy of grazing intensity classification, with the AIFS method providing both efficiency and precision.

The classification model developed in this study was applied to perform pixel-level classification on images generated by fusing visible and multispectral data. To reduce noise, a 3 × 3 filter was used to smooth the classification map (Guan et al., 2025), as shown in Figure 11. The predicted grazing intensities (blue, green, yellow, and red areas) show a clear correspondence with the ground-truth grazing intensity areas (dashed regions representing NG, LG, MG, and SG). Regions with actual grazing intensities of NG, LG, MG, and SG were predominantly predicted as blue, green, yellow, and red, respectively. These results indicate that the model can accurately predict varying grazing intensities, demonstrating robust overall performance. However, minor deviations exist between some local predictions and actual conditions, likely due to factors such as terrain undulations and the grazing behavior of sheep.

The spectral mean curves revealed that visible bands generally increased with grazing intensity but declined under severe grazing conditions. This nonlinear pattern can be explained by the pronounced soil–vegetation hybrid effect typical of sandy grasslands. Light to moderate grazing increases soil exposure, enhancing reflectance in blue, green, and red bands due to the bright sandy background. Under heavy grazing, however, trampling induces surface crusting and litter accumulation, which reduce overall reflectance. In the multispectral region, the green and red bands similarly increased with grazing intensity, while the red-edge and near-infrared (NIR) bands decreased, reflecting reductions in chlorophyll content and canopy integrity. The slight rebound of red-edge and NIR reflectance under severe grazing likely arises from soil high-albedo effects and the persistence of grazing-tolerant species.

These results differ from the monotonic spectral decline patterns reported in humid or dense grasslands (Oliveira et al., 2019; Zhang et al., 2021a), illustrating that vegetation spectral responses in semi-arid grasslands are co-regulated by physiological degradation of plants and enhanced background reflectance from soil. This interactive mechanism underscores the need to account for both vegetation and soil components when interpreting spectral responses in dryland ecosystems (Shi et al., 2021).

Statistical analyses (ANOVA and Tukey’s tests) confirmed that all spectral indices exhibited significant differences among grazing intensities. Traditional vegetation indices (NDVI, GNDVI) effectively captured the decline in vegetation cover and photosynthetic activity between light and heavy grazing but were less sensitive to moderate grazing levels. Indices that integrate red and NIR information (NDRE, MSAVI) showed superior discrimination between low–moderate and moderate–heavy grazing intensities, aligning with previous findings from semi-arid regions (Hernandez et al., 2024; Pádua et al., 2024; Pranga et al., 2024).

The self-developed indices (idx7, idx9, idx14) demonstrated even greater adaptability and stability, reflecting the advantages of region-specific spectral design tailored to sandy grasslands. Feature selection consistently identified ExG, NDRE, MSAVI, and the newly constructed indices (idx7, idx9, idx14) as key predictors, confirming their robustness for grazing intensity monitoring (Xu et al., 2022).

Interestingly, NDRE and idx3 exhibited relatively low M-statistic values (<0.4), suggesting weak separability when evaluated independently. However, both were included in the final feature subset derived through AIFS. This apparent discrepancy highlights the difference between univariate separability and multivariate synergy. While the M-statistic evaluates features in isolation, the AIFS approach optimizes combinations based on their joint contribution to classification accuracy. Thus, even indices with limited standalone performance can provide complementary information that enhances the discriminative power and stability of the final model. The inclusion of correlation constraints further ensures non-redundancy among selected features, reinforcing the unique role of each feature in the optimized subset.

In addition, compared with commonly used feature selection algorithms such as ReliefF and CARS, AIFS also shows practical advantages in computational efficiency. ReliefF requires repeated distance-based sampling across the entire feature space, which becomes increasingly expensive as feature dimensionality grows. CARS relies on iterative Monte-Carlo sampling and partial-least-squares regression, making its computational burden dependent on repeated model fitting. In contrast, AIFS performs a single-pass importance ranking using Random Forests followed by lightweight incremental evaluation with correlation filtering and cross-validation. Because redundant features are removed early during correlation screening, subsequent computations are substantially reduced. As a result, AIFS generally achieves competitive or higher classification accuracy while requiring fewer iterations and fewer full model evaluations than ReliefF and CARS, making it more suitable for UAV-based multi-source datasets where dozens of features must be processed efficiently.

From an ecological perspective, the observed spectral responses reflect key plant physiological mechanisms under grazing disturbance (Xiang et al., 2025a, b). Moderate grazing stimulates compensatory growth and chlorophyll regeneration, leading to increased red-edge reflectance, whereas severe grazing reduces leaf area index (LAI) and internal scattering, lowering NIR reflectance. Concurrently, increased soil exposure elevates visible-band reflectance, generating a nonlinear response in the composite spectral signal. This “hybrid soil–vegetation effect” emphasizes that spectral variability in sandy grasslands arises from coupled changes in vegetation physiology, community composition, and soil optical properties.

Methodologically, this study demonstrates that multi-source UAV imagery combined with AIFS enhances classification performance while reducing redundancy, outperforming both full-feature and single-source approaches. The optimized feature set (idx9, idx8, idx7, NDVI, idx2, idx3, NDRE) achieved the highest classification accuracy and robustness, validating the efficiency of automated, data-driven feature selection in heterogeneous environments.

Despite these advances, several limitations remain. The study was conducted on a controlled grazing trial with limited spatial coverage, which may not fully capture the spatial heterogeneity of sandy grasslands. In particular, the relatively small study area restricts the ability to test the generalizability of the proposed method across broader ecological gradients. The use of single-period UAV imagery restricted the analysis of seasonal dynamics. Additionally, some confusion persisted between light and moderate grazing levels, suggesting the need for features more sensitive to early degradation signals. Furthermore, although UAV-based data acquisition reduces many operational uncertainties, potential sources of error remain, including illumination variability, minor fluctuations in flight altitude, and atmospheric effects. In this study, flights were conducted under clear-sky conditions near solar noon to minimize illumination differences, and radiometric calibration panels along with the UAV’s RTK module were used to ensure consistent reflectance retrieval and high positioning accuracy. These measures help reduce uncertainty, but residual variability may still influence spectral responses.

Future studies should expand monitoring to diverse soil–vegetation types, extend the research area to larger and more heterogeneous sandy grassland regions, incorporate multi-temporal UAV datasets for dynamic assessments. Integrating medium- and high-resolution satellite data (e.g., Sentinel-2, GF-6) with UAV observations would also allow multi-scale validation and facilitate the application of the method to regional monitoring. Moreover, integrating structural information from LiDAR or radar could enhance discrimination under low vegetation cover. Further development of hybrid feature selection and deep learning approaches could also improve model robustness and generalizability.