Multitemporal Extraction of Pedicularis kansuensis in the Bayinbuluk Grassland Based on UAV Images

ZHANG Jiarong¹,²,³, ZHAO Jin¹,⁴, LI Haining³,⁵, GONG Yanming¹,³, LIU Yanyan¹,³, LIN Jun⁶, LI Kaihui¹,³

¹ Xinjiang Institute of Ecology and Geography, Chinese Academy of Sciences, Urumqi 830011, Xinjiang, China
² University of Chinese Academy of Sciences, Beijing 100049, China
³ Bayinbuluk Alpine Grassland Observation and Research Station of Xinjiang, Bayinbuluk 841314, Xinjiang, China
⁴ Xinjiang Key Laboratory of RS & GIS Application, Urumqi 830011, Xinjiang, China
⁵ College of Grassland Science, Xinjiang Agricultural University, Urumqi 830052, Xinjiang, China
⁶ Center for Grassland Biological Disaster Prevention and Control of Xinjiang Uygur Autonomous Region, Urumqi 830000, Xinjiang, China

Abstract

Invasive plants have significantly impacted global ecosystem functions and biodiversity. Existing research has primarily focused on single-temporal-phase monitoring and classification of vegetation, with relatively few studies addressing continuous multitemporal monitoring of critical phenological stages, particularly during early phenophases. This study examined Pedicularis kansuensis, an invasive plant in the Bayinbuluk grassland of Xinjiang, using UAV-based multispectral remote sensing data and machine learning algorithms to extract its spatial distribution across key phenological stages (emergence, initial flowering, peak flowering, and senescence). The results demonstrated that: (1) The spatial distribution patterns during early growth stages (emergence and initial flowering) exhibited high spatial coincidence rates with the peak flowering stage, and the random forest algorithm could effectively map the early-stage distribution of P. kansuensis; (2) The spatial distribution of P. kansuensis showed significant inter-annual variation, with spatial overlap between years below 15%; (3) During the growing season (excluding senescence), the normalized difference vegetation index calculated from the 555 nm and 720 nm bands showed the highest importance, followed by the visible green band. These findings demonstrate the feasibility of using UAV multispectral remote sensing technology for monitoring early phenological stages of P. kansuensis, providing technical support for early warning and control efforts.

Keywords: Pedicularis kansuensis; unmanned aerial vehicle; multispectral data; random forest; multitemporal; key phenological stages

1. Introduction

Invasive plants have altered plant communities and ecosystem functions worldwide, posing severe threats to native community structure and biodiversity. Under global climate change, controlling invasive plants is crucial for maintaining grassland ecosystem stability and biodiversity. The spectral and structural differences between invasive and native plants during critical phenological stages create opportunities for remote sensing-based distribution mapping. If invasive plant distribution maps can be generated before flowering or seed setting, they can provide essential data for proactive control measures, significantly improving management effectiveness and reducing damage to native forage species.

Pedicularis kansuensis, a hemiparasitic invasive plant, has been expanding across Xinjiang, Tibet, and Qinghai provinces in China. Its rapid spread significantly impacts grassland ecosystems and threatens local livestock development. As an annual or biennial herb with strong reproductive capacity and parasitic characteristics, it has rapidly invaded the Bayinbuluk grassland in the Tianshan Mountains, posing a serious threat to native forage grasses and altering the original community ecological balance. Previous control measures have focused primarily on the peak flowering stage, employing manual removal, corolla removal, or mowing. However, the plant's flowering period lasts over 40 days, during which it produces numerous seeds, and mowing or removal often disperses seeds, reducing control effectiveness and promoting germination in subsequent years. While existing remote sensing and UAV-based identification studies have concentrated on the peak flowering stage, the ability to identify P. kansuensis before peak flowering remains an urgent challenge for effective management.

The rapid development and widespread application of unmanned aerial vehicle (UAV) technology have provided new solutions for vegetation remote sensing monitoring. Compared with satellite remote sensing, UAVs offer higher spatial resolution data with minimal disturbance to the lower canopy and flexible flight planning based on local weather conditions. This study used UAV multispectral remote sensing data and machine learning algorithms to conduct multitemporal monitoring of P. kansuensis during key phenological stages in 2023 and 2024. Based on acquired multitemporal multispectral UAV data, we utilized machine learning algorithms to extract and identify P. kansuensis and other land cover types, explored the feasibility of extraction at each critical phenological stage, and analyzed inter-annual spatial distribution changes to provide scientific support for management strategies targeting P. kansuensis expansion.

1.1 Study Area

The study area (Fig. 1) is located in the Bayinbuluk grassland, Hejing County, Bayingolin Mongol Autonomous Prefecture, in the central Tianshan Mountains of Xinjiang (42°18′–43°34′N, 82°27′–86°17′E). This grassland is situated in a high-altitude intermountain basin with an average elevation of 2,500 m. The region has an average annual precipitation of 265.7 mm and mean annual temperature of -4.8°C. The natural vegetation community consists primarily of perennial forage grasses, including Stipa purpurea, Festuca kryloviana, Koeleria cristata, Leymus tianschanicus, Elymus nutans, and Potentilla multifida. Pedicularis kansuensis is mainly distributed in the alpine grasslands around the basin periphery and sporadically in wetland areas. The growing season for P. kansuensis in Bayinbuluk typically spans approximately 110 days, beginning to green up in mid-May, reaching peak flowering in late July, and entering senescence by late August.

This study was conducted in the Dayouerdus Basin of Bayinbuluk grassland, where three sample plots were established for UAV observations and ground surveys. Each plot measured approximately 100 m × 100 m, with three 1 m × 1 m quadrats established in each plot. Ground surveys were conducted concurrently with each UAV sampling campaign to confirm plant community composition. Detailed species composition and basic information for each plot are provided in Table 1.

[FIGURE:1] Location of ground sampling plots and UAV RGB imagery in the Pedicularis kansuensis study area

[TABLE:1] Information of study sites

1.2 Data Sources and Preprocessing

The study employed a DJI Matrice 300 RTK UAV equipped with a Changguang MS600 Pro multispectral camera system for data acquisition. This system integrates high-precision positioning with multispectral imaging technology to capture fine-scale temporal data on vegetation growth status. The MS600 Pro camera features six spectral channels: three visible bands—blue (center wavelength 450 nm, bandwidth 35 nm), green (555 nm, 22 nm), and red (660 nm, 27 nm); two red-edge bands (720 nm, 10 nm; 750 nm, 10 nm); and one near-infrared band (840 nm, 30 nm).

UAV data collection was conducted on multiple dates in June, July, August, and September 2023, with specific sampling dates shown in Table 2. To minimize illumination effects, data acquisition was performed between 10:00–14:00 local time under clear, cloudless conditions. Flight parameters were set as follows: altitude 80 m, forward overlap 80%, side overlap 70%, and expanded margin 10 m. Standard gray panels were photographed before and after each flight for radiometric correction, with panels placed horizontally on open ground and the multispectral camera hovering vertically 80 cm above the panel center to ensure shadow-free capture.

Multispectral data preprocessing involved: (1) Using Yusense Map software for band registration and radiometric calibration, converting raw data to reflectance using standard gray panel reflectance data; (2) Using Pix4D software for precise geometric registration of reflectance data, feature point extraction and matching, dense point cloud generation, and orthomosaic production through texture mapping and orthorectification; (3) Using ArcGIS software for spatial alignment and pixel matching across different temporal images from the same plot to ensure spatial consistency; (4) Using ENVI software for region-of-interest cropping.

[TABLE:2] Sampling dates of UAV data

1.3 Feature Selection and Sample Collection

The feature set comprised single-band spectral features and normalized difference vegetation indices. Single-band spectral features utilized reflectance data from the six MS600 Pro bands. Considering that normalized difference vegetation indices can effectively eliminate reflectance effects from cloud shadows and terrain relief, we calculated all possible normalized spectral index combinations to remove background interference. For clarity, these were designated as NDji, where j and i represent specific band numbers, with the formula: NDji = (ρj - ρi)/(ρj + ρi), where ρj and ρi represent reflectance of the j-th and i-th multispectral bands, respectively.

Sample data were collected based on field quadrat surveys combined with visual interpretation of UAV imagery. The main land cover types in the study area included P. kansuensis, perennial herbaceous plants, and bare land (Plot 1 also contained a river channel). Given that the research objective focused on P. kansuensis, samples were divided into two classes: P. kansuensis and other land cover types. Detailed sample information is provided in Table 3. Samples were split into training and validation sets at a 7:3 ratio for machine learning algorithm training and accuracy assessment.

[TABLE:3] Sample selection in the study area

1.4 Machine Learning Methods

Support Vector Machine (SVM) is a machine learning algorithm based on statistical learning theory that finds an optimal hyperplane to separate different classes by maximizing the margin between them. SVM offers advantages including small sample handling, strong noise resistance, and suitability for high-dimensional data analysis, maximizing model generalization capability.

Random Forest (RF) is an ensemble machine learning algorithm based on Classification and Regression Trees (CART). It resamples the training dataset into multiple subsets, trains a classification tree on each subset, and combines predictions through voting. RF excels at handling high-dimensional features, assessing feature importance, and resisting overfitting.

This study used grid search to optimize key hyperparameters for both algorithms. For SVM, the penalty parameter C was tested in the range [0.1, 1, 10] with kernel functions including linear, radial basis function (RBF), and polynomial. For RF, preliminary experiments showed model performance stabilized with 200 trees, so this was fixed while optimizing remaining hyperparameters via grid search: minimum leaf node samples [3, 5, 7, 10], tree depth [6, 8, 10, 15], and feature selection methods including "sqrt" and "log2". Both algorithms employed 10-fold cross-validation to determine optimal hyperparameter combinations.

Permutation importance was used to evaluate feature contributions to model predictions. This method establishes a baseline model, then randomly shuffles values of individual features while keeping others constant, measuring performance degradation to assess importance. Unlike Gini impurity, permutation importance accounts for feature interactions and is unaffected by feature scale or type, providing more accurate assessment of actual contributions. To obtain stable results, each feature was permuted 10 times, with average importance scores calculated, normalized for comparability, and used to identify key features for P. kansuensis identification across different periods.

2 Results

2.1 Classification Accuracy Evaluation

Comparison of model accuracy across different monitoring periods revealed that both SVM and RF performed similarly, with RF slightly outperforming SVM in most periods. Accuracy for P. kansuensis showed consistent temporal trends across all plots: lower accuracy during early growth stages in mid-June, peaking during peak flowering in late July to late August, then gradually declining.

The highest accuracy occurred in mid-August (peak flowering), with RF achieving an F1-score of 99.58% and Kappa coefficient of 0.99. The initial flowering stage (late June to early July) also showed high accuracy, with Plot 1 peaking in early July (F1-score: 97.78%, Kappa: 0.96) and Plot 2 peaking in late July. Model performance for identifying P. kansuensis was slightly lower than for other land cover types, with producer's accuracy lower than user's accuracy, indicating more P. kansuensis samples were misclassified as other land cover types than vice versa.

[TABLE:4] Classification accuracy of site 1
[TABLE:5] Classification accuracy of site 2
[TABLE:6] Classification accuracy of site 3

2.2 Spatial Distribution Changes of Pedicularis kansuensis

Spatial distribution maps for 2023 and 2024 revealed significant spatial differences in P. kansuensis distribution (Fig. 2). Plot 1 showed relatively small inter-annual area changes, while Plots 2 and 3 exhibited substantial variation. In Plot 2, distribution areas were 650.69 m² in 2023 and 636.88 m² in 2024, with an overlap of only 79.10 m² (12.16% and 12.42% of respective annual areas). Plot 3 showed the most dramatic changes: distribution area reached 8,256.87 m² in 2023, decreased to 7,479.11 m² in 2024 (a reduction of 777.76 m²), but with the lowest overlap area (271.57 m²) among all plots, representing just 1.98% of the 2024 area.

[FIGURE:2] Spatial distribution of P. kansuensis in the study area in 2023 and 2024
[TABLE:7] Spatial transition matrix of P. kansuensis during 2023-2024

2.3 Feature Importance

Feature importance calculated using the random forest algorithm showed fluctuating temporal patterns. The normalized difference index based on 555 nm and 720 nm bands was most important during most periods, with relative importance increasing initially, peaking in mid-August, then declining (Fig. 3). In mid-September (late growing season), the near-infrared band (840 nm) showed notably increased importance. The green band (555 nm) became most important in late September. The red-edge band (720 nm) showed highest relative importance in mid-June and mid-July.

[FIGURE:3] Temporal changes in relative importance of features

3 Discussion

This study demonstrated multitemporal monitoring and classification of P. kansuensis using UAV multispectral data and machine learning algorithms. Classification accuracy during peak flowering (mid-August) was significantly higher than other phenological stages, with Kappa coefficients ≥0.96, confirming that peak flowering is the most distinguishable stage for P. kansuensis identification. This aligns with previous studies selecting peak flowering or distinct phenological stages for invasive plant monitoring. During peak flowering, plants reach heights >30 cm with distinctive rose-colored corollas, facilitating UAV multispectral identification. The high accuracy during initial flowering (late June to early July) suggests potential for early-stage monitoring, with F1-scores >85% and producer's accuracy >80%. However, lower accuracy during emergence (mid-June) reflects the plant's small size (1–3 cm height), creating mixed pixels. Senescence stage (mid-September) showed lowest accuracy due to spectral similarity with withered grasses.

Spatial overlap analysis revealed that early growth stages could detect the year's maximum distribution area, with high spatial coincidence to peak flowering (Fig. 4), enabling early-stage distribution mapping. As plants grew, distribution areas expanded gradually from the emergence base, peaking during late peak flowering.

[FIGURE:4] Spatial overlap of P. kansuensis between different stages and peak flowering stage (mid-August 2023) in site 1

Feature importance rankings varied temporally, providing reference for vegetation monitoring methods. During the growing season (excluding senescence), the normalized difference index using 720 nm and 555 nm bands was most important, followed by the green band. This differs from some studies showing red-edge (750 nm) as most important, likely due to differences in classification targets, feature selection, and sensor bands. Our MS600 Pro camera's red-edge band at 720 nm, combined with the green band, enhanced spectral contrast and improved classification. The green band's importance reflects P. kansuensis' distinct spectral characteristics in the 550–680 nm range throughout the growing season.

Significant inter-annual spatial distribution differences were observed, consistent with the "passenger" invasion mechanism where P. kansuensis occupies vacant niches created by long-term grazing. Decomposition of litter in invaded communities may inhibit growth in subsequent years, forcing the plant to seek new vacant niches. Additionally, P. kansuensis distribution is constrained by soil moisture, thriving only within specific moisture ranges, which aligns with its seed germination requirements.

UAVs offer clear advantages for small-area monitoring but have higher costs for large-scale data acquisition compared to satellite remote sensing, which can quickly and economically obtain broad coverage. Future research should: (1) Develop temporal sequence models using multitemporal remote sensing data to leverage phenological spectral characteristics and avoid spectral confusion; (2) Integrate environmental factors (soil moisture, temperature, topography) based on P. kansuensis growth habits to build more robust monitoring models; (3) Further investigate spectral feature differences between P. kansuensis and easily confused land cover types to clarify error sources.

4 Conclusions

This study used UAV multitemporal multispectral remote sensing data to monitor different phenological stages of the invasive plant Pedicularis kansuensis, employing machine learning algorithms to map its spatial distribution throughout the growing season. The main conclusions are:

1) Random forest slightly outperformed support vector machine, with model accuracy varying across growth stages: peak flowering (late July to late August) > initial flowering (late June to early July) > emergence (mid-June) > senescence (mid-September). Spatial distribution during early growth stages showed high overlap with peak flowering, with consistent key features, demonstrating that random forest can effectively map early-stage P. kansuensis distribution.

2) All three plots exhibited significant inter-annual variation in P. kansuensis spatial distribution, with <15% spatial overlap between years, creating uncertainty for management.

3) During the growing season (excluding senescence), the most important feature for distinguishing P. kansuensis from other species was the normalized difference index calculated from 555 nm and 720 nm bands, followed by the green band. Feature importance changed markedly during senescence, with near-infrared band importance increasing and showing distinct patterns across different background environments.

These results demonstrate the feasibility of UAV multispectral remote sensing for early phenological stage monitoring of P. kansuensis, providing technical support for early warning and precision control of this invasive species.

References

[1] Wang C, Jiang K, Liu J, et al. Moderate and heavy Solidago canadensis L. invasion are associated with decreased taxonomic diversity but increased functional diversity of plant communities in East China[J]. Ecological Engineering, 2018, 112: 55-64.

[2] Adeli Maidi, Liu Yanyan, Guli Nuer, et al. Poisonous plants in natural grasslands in Urumqi County and their control measures[J]. Arid Zone Research, 2013, 30(6): 1044-1048.

[3] Ju Ruiting, Li Hui, Shih Chengjen, et al. Progress of biological invasions research in China over the last decade[J]. Biodiversity Science, 2012, 20(5): 581-611.

[4] Weisberg P J, Dilts T E, Greenberg J A, et al. Phenology based classification of invasive annual grasses to the species level[J]. Remote Sensing of Environment, 2021, 263: 112568.

[5] Liu X, Liu H, Datta P, et al. Mapping an invasive plant Spartina alterniflora by combining an ensemble one class classification algorithm with a phenological NDVI time series analysis approach in middle coast of Jiangsu, China[J]. Remote Sensing, 2020, 12(24): 4139.

[6] Sun C, Li J, Liu Y, et al. Plant species classification in salt marshes using phenological parameters derived from Sentinel-2 pixel differential time series[J]. Remote Sensing of Environment, 2021, 256: 112320.

[7] Graenzig T, Fassnacht F E, Kleinschmit B, et al. Mapping the fractional coverage of the invasive shrub Ulex europaeus with multitemporal Sentinel-2 imagery utilizing UAV orthoimages and a new spatial optimization approach[J]. International Journal of Applied Earth Observation and Geoinformation, 2021, 96: 102281.

[8] Valente J, Hiremath S, Ariza Sentis M, et al. Mapping of Rumex obtusifolius in nature conservation areas using very high resolution UAV imagery and deep learning[J]. International Journal of Applied Earth Observation and Geoinformation, 2022, 112: 102864.

[9] Wijesingha J, Astor T, Schulze Brueninghoff D, et al. Mapping invasive object Lupinus polyphyllus Lindl. in semi natural grasslands using object-based image analysis of UAV borne images[J]. PFG Journal of Photogrammetry, Remote Sensing and Geoinformation Science, 2022, 13: 1047670.

[10] Kupková L, Červená L, Poková M, et al. Towards reliable monitoring of grass species in nature conservation: Evaluation of the potential of UAV and PlanetScope multi temporal data in the Central European tundra[J]. Remote Sensing of Environment, 2023, 294: 113631.

[11] Knapp L S P, Coyle D R, Dey D C, et al. Invasive plant management in eastern North American forests: A systematic review[J]. Forest Ecology and Management, 2023, 550: 121517.

[12] Wang D, Cui B, Duan S, et al. Moving north in China: The habitat of Pedicularis kansuensis in the context of climate change[J]. Science of the Total Environment, 2019, 697: 133979.

[13] Bao Gensheng, Wang Hongsheng. Allelopathic effects of Pedicularis kansuensis on several graminaceous grass species on alpine meadow[J]. Chinese Journal of Grassland, 2011, 33(2): 88-94.

[14] Li H, Gong Y, Fang F, et al. Effects of nutrient addition on Pedicularis kansuensis invasion of alpine grassland[J]. Atmosphere, 2023, 14(2): 367.

[15] Sui X, Kuss P, Li W, et al. Identity and distribution of weedy Pedicularis kansuensis Maxim. (Orobanchaceae) in Tianshan Mountains of Xinjiang: Morphological, anatomical and molecular evidence[J]. Journal of Arid Land, 2016, 8(3): 453-461.

[16] Li W, Huang L, Yang L, et al. Phenotypic plasticity drives the successful expansion of the invasive plant Pedicularis kansuensis in Bayanbulak, China[J]. Diversity, 2023, 15(3): 313.

[17] Wang W, Tang J, Zhang N, et al. Spatiotemporal pattern of invasion by Pedicularis kansuensis in the Bayinbuluke Land, China, during 2019-2021: An analysis based on PlanetScope and Sentinel-2 data[J]. Remote Sensing, 2023, 15(18): 4383.

[18] Wang W, Tang J, Zhang N, et al. Automated detection method to extract Pedicularis kansuensis based on UAV images[J]. Drones, 2022, 6(12): 413.

[19] Zhao J, Li K, Zhang J, et al. Mapping invasive species Pedicularis kansuensis and background grassland using UAV and machine learning algorithms[J]. Drones, 2024, 8(11): 639.

[20] Yang X, Smith A M, Bourchier R S, et al. Mapping flowering leafy spurge infestations in a heterogeneous landscape using unmanned aerial vehicle Red Green Blue images and a hybrid classification method[J]. International Journal of Remote Sensing, 2021, 42(23): 8854-8872.

[21] Valero Jorge A, Zayas R G D, Matos Pupo F, et al. Mapping and monitoring of the invasive species Dichrostachys cinerea (Marabú) in central Cuba using Landsat imagery and machine learning (1994-2022)[J]. Remote Sensing, 2024, 16(5): 798.

[22] Xing F, An R, Guo X, et al. Mapping invasive noxious weed species in the alpine grassland ecosystems using very high spatial resolution UAV hyperspectral imagery and a novel deep learning model[J]. GIScience & Remote Sensing, 2024, 61(1): 2327146.

[23] Innangi M, Marzialetti F, Di Febbraro M, et al. Coastal dune invaders: Integrative mapping of Carpobrotus sp. pl. (Aizoaceae) using UAVs[J]. Remote Sensing, 2023, 15(2): 503.

[24] Vetter V M S, Kreyling J, Dengler J, et al. Invader presence disrupts the stabilizing effect of species richness in plant community recovery after drought[J]. Global Change Biology, 2020, 26(6): 3555-3565.

[25] Liu Yanyan. Ecological Factors of Pedicularis kansuensis Maxim. Expansion in Bayanbulak Grassland[D]. Urumqi: Xinjiang University, 2018.

[26] Liu Y, Li W, Sui X, et al. An exotic plant successfully invaded as a passenger driven by light availability[J]. Frontiers in Plant Science, 2020, 88(5): 391-406.

[27] Li Haining, Liu Yanyan, Gong Yanming, et al. Effects of Pedicularis kansuensis invasion on litter decomposition in Bayanbulak alpine steppe[J]. Grassland and Turf, 2023, 43(1): 1-11.

[28] Hu J, Li K, Deng C, et al. Seed germination ecology of semiparasitic weed Pedicularis kansuensis in alpine grasslands[J]. Plants, 2022, 11(13): 1777.

[29] Cortes C, Vapnik V. Support vector networks[J]. Machine Learning, 1995, 20(3): 273-297.

[30] Breiman L. Random forests[J]. Machine Learning, 2001, 45(1): 5-32.

[31] Chen X, Ishwaran H. Random forests for genomic data analysis[J]. Genomics, 2012, 99(6): 323-329.

[32] Strobl C, Boulesteix A L, Zeileis A, et al. Bias in random forest variable importance measures: illustrations, sources and a solution[J]. BMC Bioinformatics, 2007, 8: 25.

[33] Boulesteix A L, Bender A, Lorenzo Bermejo J, et al. Random forest Gini importance favours SNPs with large minor allele frequency: Impact, sources and recommendations[J]. Briefings in Bioinformatics, 2012, 13(3): 292-304.

[34] Dronova I, Spotswood E N, Suding K N. Opportunities and constraints in characterizing landscape distribution of an invasive grass from very high resolution multi spectral imagery[J]. Frontiers in Plant Science, 2017, 8: 890.

[35] Zhan Guoqi, Yang Guodong, Wang Fengyan, et al. The random forest classification of wetland from GF-2 imagery based on the optimized feature space[J]. Journal of Geo-information Science, 2018, 20(10): 1520-1528.

[36] Sun Fafu, Lai Ning, Geng Qinglong, et al. Estimation of nitrogen contentration in winter wheat leaves based on hyperspectral images of UAV[J]. Arid Zone Research, 2024, 41(6): 1069-1078.