4 Discussion

4.1 Feasibility of Vegetation Species Identification

The comparative accuracy of species identification algorithms by remote sensing at different scaled has not been extensively studied (Camarretta et al., 2020). This study utilized pixel-based supervised classification to identify plant species by training and validating machine learning algorithms. SVM outperformed maximum likelihood and neural network classifiers in fragmented land plots with mixed vegetation types, exhibiting higher accuracy with limited training samples and minimal Hughes phenomenon. However, pixel-based supervised classification may not fully utilize the textural and structural information capability of remote sensing imagery, leading to ”salt and pepper” and ”spectral confusion” phenomena, particularly for species with similar spectral characteristics. Future studies may introduce texture features and spatial information (Camarretta et al., 2020), such as LiDAR and digital elevation models, to address these limitations (Shokirov et al., 2023).
Monitoring and evaluation are crucial steps in restoration projects, emphasizing the implementation of practical indicators instead of broadly defined and often ambiguous monitoring standards (Australian Government, 2016; Nilsson et al., 2016; Young et al., 2022). The transition from conceptual ideas to on-ground implementation aims at cost-effective and successful restoration (Evju et al., 2020). This study highlights the benefits of low-altitude UAV multispectral data for cost-effective, easy-to-use, and high-efficiency ecological monitoring at a fine scale, thereby improving the ease and accuracy of monitoring efforts.
Firstly, the accuracy of low-altitude UAV imagery surpasses that of satellite remote-sensing data (Johansen et al., 2019). Satellite remote sensing covers a wide spatial extent and diverse landscape types, making it difficult to capture detailed vegetation information (Peng et al., 2021), leading to less precise classification. This is especially true in arid and semi-arid areas where sparse vegetation, a mixture of herbaceous shrubs, and different soil reflectance values complicate accurate vegetation monitoring (Rossi et al., 2022).
Secondly, UAV multispectral data is more cost-effective than hyperspectral sensors and LiDAR for many applications (de Almeida et al., 2021; Haneda et al., 2023). Hyperspectral data offers a multitude of spectral bands that enhance diversity detection potential; however, this might result in a decrease in classification accuracy in high spectral dimensions (Gholizadeh et al., 2018). Despite its advantages in characterizing vegetation vertical structure (Fernandez-Guisuraga et al., 2022), the widespread application of LiDAR is significantly limited due to its high expenses. Therefore, practitioners need to weigh both the economic costs of LiDAR and the time costs associated with hyperspectral data analysis. In light of these considerations, integrating UAV multispectral data emerges as the optimal approach for monitoring ecological restoration in mining areas. The cost-effectiveness of UAV multispectral data, compared to hyperspectral sensors and LiDAR, makes it a favorable choice for various applications
The present study has some limitations regarding species classification, which may be improved by combining deep learning and spectral features to increase accuracy and efficiency (Saadeldin et al., 2022) such as the integration of vegetation height and canopy extraction and species classification. Further in-depth research could also explore the effects of lower flight altitudes (e.g. 10-50m) and adjusted flight attitudes (e.g. oblique photogrammetry) for more accurate monitoring of low-growing herbaceous or recently emerged seedlings.

4.2 Feasibility of UAV-based Vegetation Diversity Monitoring

Field measurements are essential for validating the accuracy of results. Figure 11 shows the relative importance values of vegetation species obtained through field surveys. Poplar, Artemisia, and Salix were ranked as the top three species by importance value, consistent with results from UAV imagery. Poplar dominated the tree layer, while Artemisia dominated the shrub layer. Notably, Artemisia had the largest relative abundance and relative frequency, values of 0.48 and 0.94, respectively, indicating it is the most abundant and widely distributed species in the sampled area and therefore a key dominant species.