Large-scale remote sensing model enables an integrated monitoring approach for high-resolution tracking pest vole populations

Abstract

Grassland voles pose significant challenges to agriculture and public health due to their population outbreaks. Traditional monitoring methods are labor-intensive and costly, particularly in heterogeneous landscapes. This study integrates Sentinel-2 satellite imagery with field data to develop a predictive framework for monitoring fossorial water vole (Arvicola scherman) populations in northwestern Spain. We present a high-resolution habitat suitability model (97% accuracy) and an Optimized Damage Index that accounts for climatic variability to reliably infer fossorial water vole abundance based on vegetation damage of grasslands and meadows. April and August were identified as optimal monitoring periods, as they coincide with opposing grass conditions and vole activity. Our approach enables early detection of outbreak zones, even in the absence of continuous field surveys, and supports scalable, cost-effective vole management. The framework improves decision-making for vole population control, optimizes resource allocation, and can be adapted to other species or regions. These findings highlight the value of remote sensing for proactive, real-time vole management, enhancing sustainable crop protection strategies.

Introduction

Small mammals play a pivotal role in terrestrial ecosystems across diverse regions. Beyond their contributions to trophic webs and ecosystem services¹, understanding their spatial distributions and population dynamics is crucial for several applied fields. In agriculture, voles contribute positively by incorporating organic matter into the soil, thereby enhancing aeration, infiltration, and providing optimal conditions for seed germination². However, in Western Europe, the homogenization of agricultural landscapes has exacerbated the prevalence and impact of grassland vole species, such as the common vole (Microtus arvalis) and the fossorial water vole (Arvicola scherman, formerly the fossorial form of Arvicola terrestris³. Both species exhibit wave-like dispersal patterns at regional scale^4,5, often reaching densities of 500 individuals per hectare and peaking at 1,000 voles/ha during population outbreaks^4,6. Due to their capacity to produce important damage to crops and their role as reservoirs of zoonotic pathogens, these voles are recognized as major agricultural pests and public health risks across their distribution range^7–9.

Effective and evidence-based management of grassland vole pests offer substantial benefits across multiple sectors of human society. A critical step toward achieving effective management is enhancing the monitoring and predictive capabilities of vole management programs^9,10. Developing advanced models capable of accurately characterizing population outbreaks across extended time series and spatial scales is highly recommended for implementing targeted control strategies and issuing timely outbreak risk alerts^9–11. However, drivers that significantly influence vole ecology should be considered to establish effective monitoring strategies. This is particularly evident in the case of fossorial water voles, whose population dynamics may be variably influenced by climate and landscape features across France, Switzerland, Germany, and Spain^12,13. The amplitude of their multi-annual density fluctuations appears to be strongly influenced by climatic conditions. Thus, fluctuations are well-defined in seasonal environments^6,14,15 but lack predictability in low-altitude areas^13,16. Landscapes dominated by pasturelands facilitate unimpeded dispersal, enabling demographic synchrony and large-scale merging of vole populations^5,17,18. Conversely, heterogeneous and highly variegated landscapes restrict dispersal success, thereby influencing population diffusion¹⁹. Thus, multiannual density fluctuations at local-scale have been observed, such as in sub-populations in the Atlantic NW region of Spain¹⁹.

Conventional ground surveys can be expensive and time-consuming when applied to large-scale monitoring efforts²⁰. Monitoring burrowing mammals presents additional challenges due to their extensive ranges, the labor-intensive nature of field surveys, and the high associated costs^11,21. Furthermore, distinguishing between different cohabiting species often requires specialized technicians, adding another layer of complexity^11,16. For A. scherman, broad-scale tracking must account not only for current habitat occupancy but also for nearby habitats that are feasible for colonization, particularly during population outbreaks^5,12. The unpredictability of fossorial water vole settlement locations underscores the need for extensive surveillance to ensure effective population management²². The limited resolution offered by many available cartographic sources, along with the disparity in criteria used to classify land use, makes it highly advisable to develop cartography that identifies the habitats potentially occupied by fossorial water voles. This task becomes especially difficult in variegated landscapes, which are characterized by a mosaic of diverse land-use plots¹⁹. Management efforts must also involve collaboration with local administrations and farmers to ensure the successful implementation of monitoring and control initiatives^10,11. In addition, societal and governmental support for research is not always assured. Effective control strategies for fossorial water voles therefore presents significant challenges due to the complexity of their ecology and the resources required for large-scale surveys²³.

Remote sensing technology has emerged as a reliable and efficient tool for large-scale monitoring^21,24,25 or disease prediction²⁶. Integrating remote sensing technology into expert diagnostic frameworks allows also for the objective evaluation of infestation damage and the accurate prediction of its likelihood^27–29. However, challenges remain. Despite burrow mounds of fossorial water voles could be identify even as they deteriorate with time³⁰, they can be easily mistaken for those of the mole (Talpa europaea, Talpa aquitania or Talpa occidentalis), and they can be mistaken for cow droppings in grasslands, making aerial or satellite imagery used for detecting other species^21,25,31 unreliable for identifying these signs. Sentinel-2 satellites from the European Space Agency (ESA) have been recognized as a valuable resource for predicting abundance and distribution of rodents^32,33 and other pest species^34,35. These satellites offer a high revisit frequency, finer spatial resolution (10 m for visible and near-infrared bands), and enhanced spectral resolution³⁶. Derived vegetation indices, such as NDVI (Normalized Difference Vegetation Index) and NDRE (Normalized Difference Red Edge Index), have been also used for early detection of water stress, habitat quality, or to stablish a correlation with vegetation phenology and health status^36,37. This tool has been very useful in indirectly studying rodent distribution through their effects on vegetation^33,38,39. In this regard, fossorial water voles consume both epigeal and hypogeal parts of plants of dicotyledons and some poaceae^40,41, so the damage to herbaceous species can be very conspicuous. Vegetation indices derived from Sentinel-2’s red-edge bands are likely capable of discriminating between different degrees of damage intensity caused by this species in grasslands and meadows.

In this study, we leverage Sentinel-2 data to address the challenges of monitoring fossorial water vole populations during a recent large-scale outbreak in Galicia, northwestern Spain⁴². Our objectives are twofold: (1) to develop a predictive model for identifying at high resolution scale current and potential habitats that could be occupy by fossorial water voles during their population diffusion and (2) to combine field surveys with vegetation health assessments to create a damage index, which allow to quantify vole abundance based on vegetation damage severity. This approach integrates field observations and remote sensing data to establish a robust and cost-effective framework for pest management, minimizing labor demands while enabling rapid decision-making in response to vole outbreaks.

Materials and methods

Study area

The study was conducted in an agricultural area covering approximately 1,285 km² in the highlands of the Ancares mountains (Galicia, NW, Spain, 42°45′N, 7°14′W), encompassing eight municipalities: As Nogais, Baleira, Cervantes, Folgoso do Courel, Navia de Suarna, Piedrafita do Cebreiro, Samos, and Triacastela (Fig. 1). The agroecosystem is characterized by a variegated landscape, containing a grained mosaic of hay meadows, livestock grasslands, annual crops, settlements, and semi-natural woody vegetation patches, mainly coniferous and broad-leaved forest and heathland patches. Extensive and nearly subsystem livestock farming is primarily developed in small plots separated by stone walls, minimally-managed hedgerows, or small woodland patches. Some relatively large plots (aprox. 30 ha) are shared by several owners and are mainly intended for mowing meadows.

Fig. 1.

Distribution area of the fossorial water vole, Arvicola scherman, in Spain (green), with the study area highlighted in red. The sampling points are shown within the study area. Map generated with QGIS 3.28 (https://qgis.org/).

Field survey

The estimation of fossorial water vole abundance at patch scale was based on the index described by Giraudoux et al.⁴³. The surface signs of activity, mainly earth mounds, were determined along a line that crossed each patch on its largest dimension. This transect was divided into 10 m intervals and the area observed was 2 m² around the technician. The presence or absence of A. scherman was recorded for each interval. The relative abundance was used as an index (Abundance Index, AI) calculated by the ratio: number of positive intervals/total number of intervals^43,44. Thus, five categories were established: 0 (no presence), 1 (1–25%), 2 (25–50%), 3 (50–75%), and 4 (> 75%). Surveys were conducted biannually (spring and autumn) since autumn 2021 to spring 2024.

Remote sensing data

Top-of-atmosphere (TOA) Level-2A Sentinel-2 scenes were acquired using the Google Earth Engine (GEE) API. Standard preprocessing steps were applied, including cloud masking using the ‘QA60’ band and the Scene Classification Layer (SCL), and the use of Level-2A products ensured atmospheric correction via the Sen2Cor processor. However, GEE may still provide pixel values affected by residual cloud contamination or saturation. To address this, we implemented a proprietary software solution (Spectralgeo) to further refine the dataset by correcting errors undetected by GEE. These errors were identified by analyzing the temporal evolution of specific spectral indices, assuming that near-continuous trends are expected in vegetation dynamics. Meteorological data from OpenMeteo (https://open-meteo.com) were incorporated. The monthly averages of temperature (minimum, mean, and maximum), monthly precipitation, and the monthly mean relative humidity were obtained. These data were downloaded for the years of the study, ensuring temporal consistency with the spectral indices. A total of 12 Sentinel-2 satellite bands (‘B1’, ‘B2’, ‘B3’, ‘B4’, ‘B5’, ‘B6’, ‘B7’, ‘B8’, ‘B8A’, ‘B9’, ‘B11’, ‘B12’) were used to develop the models. These included the native bands provided by the satellite, along with derived vegetation and water indices. Using this corrected dataset, the temporal evolution of these bands was analyzed to detect trends in habitat dynamics.

Predictive spatial modeling of A. scherman habitat

A machine learning model was developed to identify potential habitat plots, for which 82,905 pixels (10 × 10 m each) were labelled based on A. scherman presence data collected from field surveys in the municipality of Cervantes. For each pixel, cloud-free data from the first day of each month was selected as a reference. The dataset used to train the model included the temporal evolution of 12 aforementioned Sentinel-2 satellite bands and 14 vegetation indices (‘NDVI’, ‘NDWI’, ‘NDRE’, ‘SAVI’, ‘GCI’, ‘NDMI’, ‘MSI’, ‘SIPI’, ‘EVI’, ‘TCARI’, ‘GRDIFF’, ‘RVI’, ‘AOT’, ‘WVP’) from the preceding year, as well as altitude, slope, and geospatial coordinates. A Random Forest algorithm was chosen as the classification model. Additionally, a Standard Scaler was applied to standardize the features, which ensures that features with small scales can influence the model. Missing values were imputed using the mean. The Synthetic Minority Oversampling Technique (SMOTE) was applied to balance the classes, ensuring more accurate predictions and was used only on the training set to avoid overfitting and maintain the integrity of the validation and test data. The labelled pixels were initially categorized into two classes: habitat (label 1) and non- habitat (label 0). To enhance inclusivity in the classification, the Random Forest model’s decision threshold was adjusted to 0.4, such that any pixel with a probability of 40% or higher of being classified as habitat was labelled as such. To evaluate model performance, a confusion matrix was generated. This adjustment significantly improved prediction accuracy. Ten representative small zones were selected across the study area in Lugo. Two of these, located within the municipality of Cervantes, were designated as the test set, while the remaining eight were used for training and validation. To ensure spatial independence and avoid spatial autocorrelation, the test zones were located at least 5 km away from the training areas. The target labels spanned multiple years, thereby capturing temporal variability in the data. Furthermore, the distribution of the target variable was carefully preserved across both the training and test sets to maintain consistency in class representation. The final model was applied to a total of 15,697,648 pixels across the eight municipalities.

Quantitative model for estimating A. scherman-induced vegetation damage

To mitigate potential misclassification risks arising from aerial or high-resolution optical imagery artifacts that may resemble vole activity signs, vole abundance was inferred indirectly through the severity of vegetation damage. Previous vegetation indices have been tested to discriminate general annual trends in the habitats. Another index, termed the Damage Index (DI), was generated by applying a machine learning model to custom combinations of Sentinel-2’s near-infrared (NIR) and shortwave infrared (SWIR) bands. A total of 16,768 Abundance Index (AI) data from sampled plots across the eight municipalities were considered to test feasibility of these indices. Annual climatic conditions were found to significantly influence these indices, as noted in previous studies^36,37. To address this, monthly precipitation, temperature and humidity were incorporated to optimize DI. Based on this, an Optimized Damage Index (ODI) was developed, combining DI bands adjusted for potential water stress in the vegetation. The DI and ODI are novel metrics developed specifically for this study and are protected under intellectual property rights by Spectralgeo. Both indices are designed to specifically reflect vegetation damage caused by vole activity, with lower values indicating higher levels of damage.

All analyses and visualizations in this study were conducted using QGIS 3.28 (https://qgis.org/) and Python 3.11.5 (https://www.python.org/). QGIS was primarily used for data labelling and the creation of spatial figures (Figs. 1, 4, 8, and 10). Python was employed for the remaining tasks, including data acquisition, preprocessing, statistical analysis, and model training. Pearson correlation analyses were carried out to assess the relationship between the ODI and the monthly field-recorded relative abundance (Abundance Index, AI). Data visualizations were generated using the Matplotlib (https://matplotlib.org/) (Figs. 2, 3, 5, 6, and 9) and Seaborn libraries (https://seaborn.pydata.org/) (Fig. 7).

Fig. 4.

An orthophoto of the prediction test area and the classification results are presented, where 10 × 10-meter pixels assigned to A. scherman habitat are highlighted in light yellow. Map generated with QGIS 3.28 (https://qgis.org/).

Fig. 8.

Area affected by A. scherman, where both the habitat identification model and the Optimised Damage Index (ODI) model were applied to infer species abundance in relation to vegetation damage across potential habitats. The colour of the contours represent the inferred Abundance Index (AI) from each plot for each year, as indicated by the gradient on the right. Three plots corresponding to the graphs in Fig. 9 (a, b, c) are referenced. Map generated with QGIS 3.28 (https://qgis.org/).

Fig. 10.

Area affected by A. scherman, where both the habitat identification model and the Optimised Damage Index (ODI) model were applied to infer species abundance in relation to vegetation damage across potential habitats. Plots shown in blue indicate locations with field-survey data on relative abundance. The colour of the contours represent the inferred Abundance Index (AI) from each plot for each year, as indicated by the gradient on the right. Maps generated with QGIS 3.28 (https://qgis.org/).

Fig. 2.

Abundance Index (AI) frequency of A. scherman by number of plots surveyed during field sampling. Graph generated using Python 3.11.5 (https://www.python.org/) and Matplotlib (https://matplotlib.org/).

Fig. 3.

Confusion matrix generated to assess the performance of the artificial intelligence-based model for identifying A. scherman habitats. Graph generated using Python 3.11.5 (https://www.python.org/) and Matplotlib (https://matplotlib.org/).

Fig. 5.

Annual NDVI evolution estimated across randomly selected meadows as habitats for A. scherman (10 plots) and non-meadows (10 plots) from 2021 to 2024. Graph generated using Python 3.11.5 (https://www.python.org/) and Matplotlib (https://matplotlib.org/).

Fig. 6.

Graphs showing the temporal dynamics of the Damage Index (DI) in relation to the Abundance Index (AI) of A. scherman for the plots with repeated surveys. Each line represents the DI trajectory from January 1 to December 31 for a given plot and year. The colour of each line corresponds to the AI estimated from field surveys for that specific plot and year, with warmer colours indicating higher vole abundance and cooler colours indicating lower abundance, as defined by the gradient scale on the right. This visualisation highlights how damage intensity (DI) varies seasonally and how it relates with independently estimated vole abundance (AI). Graphs generated using Python 3.11.5 (https://www.python.org/) and Matplotlib (https://matplotlib.org/).

Fig. 9.

Graphs showing the temporal dynamics of the Optimised Damage Index (ODI) in relation to the Abundance Index (AI) of A. scherman for the plots with repeated surveys indicated in Fig. 8. Each line represents the ODI trajectory from January 1 to December 31 for a given plot and year. The colour of each line corresponds to the AI estimated from field surveys for that specific plot and year, with warmer colours indicating higher vole abundance and cooler colours indicating lower abundance, as defined by the gradient scale on the right. This visualisation highlights how damage intensity (ODI) varies seasonally and how it relates with independently estimated vole abundance (AI). Graphs generated using Python 3.11.5 (https://www.python.org/) and Matplotlib (https://matplotlib.org/).

Fig. 7.

Annual and monthly correlation between Optimised Damage Index (ODI) averages and A. scherman Abundance Index (AI). Graph generated using Python 3.11.5 (https://www.python.org/) and Seaborn libraries (https://seaborn.pydata.org/).

Results

Between autumn 2021 and spring 2024, a total of 8,712 plots were surveyed across the study area (Fig. 1), encompassing 8,058 hectares affected by fossorial water voles. The temporal distribution of plot surveys was as follows: 99 in autumn 2021, 5,640 in spring 2022, 3,960 in autumn 2022, 2,374 in spring 2023, 2,360 in autumn 2023, and 2,335 in spring 2024. Thus, 16,768 abundance estimates were recorded over the entire monitoring period (Fig. 2), reflecting the intensity of the monitoring. The number of plots visited once, twice, three, four, five, and six times was 5,687, 735, 181, 1,484, 618, and 7, respectively, indicating the temporal sampling effort.

The model for identifying A. scherman habitats achieved an accuracy of 0.9701 using an 80 − 20 split for training and testing data. Through k = 10 cross-validation, the average accuracy was 0.9639, with a standard deviation of 0.0016. The confusion matrix generated (Fig. 3) classified 11,436 pixels identified as unfavorable, and 7,798 pixels identified as habitat, totaling 19,234 correctly predicted pixels. The application of the model performs acceptably in terms of habitat prediction (Fig. 4). When the prediction model is applied, 30% of the total pixels considered (15,697,648) were identified as suitable habitat for A. scherman in this geographic area.

The vegetation indices tested, such as NDVI (Fig. 5), were able to distinguish general annual trends in meadows from those in non-habitat patches for fossorial water voles. This reveals its effectiveness in detecting significant reductions in vegetation damage. Similarly, the newly developed, study-specific DI index followed a comparable seasonal pattern (Fig. 6). However, performance analysis revealed similar declines in index values across plots with varying vole abundances (Fig. 6a, b), despite the expected reduction in vegetation damage at higher vole densities. This suggests that drought conditions can lead to reductions in vegetation damage similar to those caused by vole activity, complicating the discrimination between climate-driven and vole density-related effects.

When applying the ODI, the AI is inferred based on the state of vegetation cover within the habitat considering climatological conditions. Using a total of 16,768 AI data points from field samplings, average ODI values were calculated for annual and each monthly periods (Table 1; Fig. 7). Annual, April, and August periods proved to be the most reliable for detecting reductions in vegetation damage linked to vole abundance. Significant negative correlations were found in all three cases (annual: r = − 0.26; April: r = − 0.39; August: r = − 0.42; p < 0.01), indicating that the relative abundance of fossorial water voles can be consistently inferred during these times. The range of the annual average ODI was 0.08, while April and August showed broader ranges of 0.13, reinforcing their suitability for discriminating vole impacts (Fig. 7). In contrast, months such as June, September, November, and December showed weak or inconsistent patterns between vegetation damage and vole density, with non-significant correlations (p > 0.01 in all cases), suggesting limited usefulness for damage detection during those periods.

Table 1.

Monthly and annual average variation of optimised damage index (ODI) according to abundance index (AI) of A. scherman.

Average ODI	AI = 0	AI = 1	AI = 2	AI = 3	AI = 4
January	0.76	0.74	0.74	0.73	0.70
February	0.77	0.77	0.73	0.74	0.71
March	0.81	0.80	0.77	0.74	0.74
April	0.85	0.82	0.77	0.76	0.72
May	0.84	0.81	0.79	0.77	0.70
June	0.79	0.73	0.74	0.69	0.71
July	0.72	0.72	0.69	0.68	0.63
August	0.74	0.70	0.69	0.63	0.61
September	0.72	0.71	0.72	0.67	0.70
October	0.79	0.76	0.74	0.71	0.69
November	0.81	0.74	0.77	0.74	0.74
December	0.81	0.77	0.78	0.73	0.75
Annual average	0.75	0.74	0.73	0.68	0.67

After applying the final model to 15,697,648 pixels across eight municipalities, two representative areas with available field data were selected for detailed analysis. The predicted 10 × 10-meter pixels were converted into polygons based on cadastral parcel boundaries. Within each area, habitats potentially occupied by A. scherman were identified. Figure 8 shows one of these areas affected by the species across four consecutive years, 2021, 2022, 2023, and 2024, highlighting temporal variations in vole distribution and density. Additionally, three plots with data available were selected to directly link the ODI model with field data (Fig. 9a–c). Overall, the years 2021, and especially 2022, exhibited lower average ODI values, indicating higher levels of vegetation damage and increased vole activity. In contrast, a decline in vole populations was observed in 2023 and 2024, coinciding with signs of vegetation recovery and a continued decrease in field-recorded vole densities.

At a finer spatial scale, Fig. 10 presents the identification of plots classified as habitats for A. scherman, along with the evolution of their occupancy and inferred population density from 2021 to 2024. Although field-based abundance estimates are unavailable for most plots in this area, the model infers population densities that follow consistent and biologically plausible patterns, comparable to those observed in plots with field-survey data on relative abundance (Fig. 10).

Discussion

This study underscores the potential of Sentinel-2 remote sensing for high-resolution, large-scale monitoring and prediction of A. scherman habitats. The findings are consistent with previous research highlighting the utility of remote sensing for assessing rodent distribution and forecasting the impact of harmful species in agricultural and natural ecosystems^27,32,33,39. The integration of a high-accuracy habitat identification model with an optimized spectral index for damage assessment represents a promising step forward in the management of fossorial water vole populations. This approach has the potential to improve resource allocation and illustrates the feasibility of using satellite imagery as a cost-effective alternative to labor-intensive ground surveys²¹, enabling more precise surveillance and real-time assessment of crop damage. While further validation and adaptation to other contexts will be necessary, this methodology could contribute to more sustainable and economically viable strategies for controlling vole populations, with the potential to help mitigate agricultural losses and related public health risks^10,23.

The Damage Index (DI), based on Sentinel-2 red-edge bands, similar to indices such as NDVI, EVI, SAVI, ARVI, GCI, SIPI, and NDRE, showed interannual inconsistencies in tracking fluctuations in the Abundance Index. Although overall trends aligned with expectations, environmental factors influencing vegetation health likely affected model outputs. As noted in previous studies^36,37, annual climatic variations significantly impact vegetation indices. For example, drought conditions can reduce vegetation vigor in ways that mimic vole-induced damage, complicating the differentiation between biotic and abiotic stressors. This highlights the need to account for seasonal and climatic variability when interpreting the relationship between vegetation damage and vole abundance^33,38. In contrast, the Optimized Damage Index (ODI), which incorporates climatological variables, helped to reduce these discrepancies and provided a more consistent inference of A. scherman relative abundance. These results suggest that vegetation damage caused by voles, even when influenced by climate-related variability, may be monitored more reliably using carefully adjusted spectral indices⁶. Comparative analyses across years and seasons revealed that April and August were the most informative periods for ODI-based inference. These months correspond to extremes in vegetation health—peak and decline—which enhance the detectability of vole-related damage. This is partially consistent with spring field surveys, typically conducted during this period due to the stronger correlation between surface activity and vole density¹⁰. Fossorial water vole activity varies seasonally, with burrowing behavior peaking in spring and autumn, driven by foraging and reproduction needs^13,45, and supported by sufficient rainfall to reduce soil dryness^6,22. Given that vegetation health also tends to peak in spring, April appears particularly suitable for accurate ODI inference. Although these data are inherently noisy and shaped by multiple interacting ecological factors, the significant correlations observed in this study can be considered biologically meaningful, as they are supported by plausible ecological mechanisms. However, these findings should be interpreted as a partial, not definitive, solution to the challenge of disentangling vole impacts from climate-related effects. Further refinement of the methodology, such as incorporating higher-resolution climate data and validating remote-sensing inferences with longer-term field observations, could enhance the robustness of the results.

The habitat identification model achieved highly acceptable accuracy (97%), successfully detecting both occupied and potential habitats. The application of the final abundance model across a broad landscape made it possible to identify areas that are likely to be occupied by the species. Within these areas, spatial patterns of vole distribution and density were analyzed, and relationships between vegetation damage and vole activity were established. In zones with available field data, modelled abundance trends showed coherence with vegetation damage levels, confirming the model’s ability to capture biologically plausible population dynamics even in the absence of continuous field monitoring. Furthermore, the model provided inferences about population changes in years for which empirical data were unavailable, suggesting its potential usefulness as a complementary tool for monitoring in data-limited contexts. This capability may help support more targeted surveillance in complex landscapes, where prioritizing management actions remains challenging¹⁹. Dispersal and habitat colonization are central to fossorial water vole population dynamics^12,18,46. Juvenile and subadult voles disperse overland⁴⁷, initially over short distances that facilitate local colonization, with longer movements occurring during the population growth phase¹⁸. During this phase, a tightly coupled plant-herbivore dynamic emerges. The species exhibits marked habitat selectivity, often overexploiting high-quality areas, leading to local population surges followed by resource depletion²². Drone-based monitoring of dandelion–vole interactions has advanced our understanding of settlement patterns and management optimization²². However, traditional monitoring limitations remain. A spectral index can be incorporated as an explanatory variable to capture reflectance patterns indicative of dense dandelion presence, potentially serving as a proxy for environmental conditions favorable to A. scherman. Integrating the optimized habitat prediction model with a spectral index sensitive to dandelion aggregation, together with the ODI, may enhance habitat suitability mapping, improve damage assessment, and enable earlier identification of high-risk areas. This integration could enhance the ecological relevance of predictive outputs. To shift from reactive to predictive vole management, it is critical to develop a robust model that accurately simulates successful dispersal and subsequent settlement patterns across spatial and temporal scales. Enhancing damage prevention in heterogeneous landscapes requires future research focused on incorporating advanced machine learning techniques to refine land-use classification and assess its influence on fossorial water vole dispersal^19,29. Additionally, combining thermal sensor data with UAV imagery could improve the detection of burrows and damage in dense vegetation²⁸. These technological advancements will strengthen the predictive capabilities of vole monitoring strategies, supporting the implementation of long-term, large-scale control programs¹¹.

It is strongly recommended that both the habitat prediction model and the ODI be updated and recalibrated for each specific geographic area and ecological context. Nevertheless, their broader application to grassland voles inhabiting meadows and grasslands across Atlantic Europe^4,6,9 may be relatively straightforward, owing to the similarity in vegetation composition and damage patterns observed across both native and cultivated herbaceous species in these landscapes^41,42. The applicability of the ODI to other cropping systems will depend largely on the visibility of rodent-induced damage, which tends to be more apparent in annual crops⁴¹ and more cryptic in perennial systems such as fruit orchards, where damage accumulates gradually through root consumption¹³. For other pest rodent species occurring in distinct agroecosystems or climatic regions, habitat prediction models must be specifically developed^33,38,39, and the ODI adapted based on crop type and damage expression^34,35. Although the use of customized indices and proprietary processing pipelines may limit immediate replicability, both methodologies remain conceptually transferable to other contexts with appropriate adjustments. While remote sensing variables and presence data are generally accessible, and the ODI can be recalibrated for different pests and crops, the success of any machine learning–based monitoring framework ultimately hinges on the integration of high-quality field data that demonstrates strong correlations between pest abundance and quantifiable crop damage. To mitigate overfitting, future applications should prioritize incorporating diverse training datasets and independent validation across spatial and temporal gradients. These findings highlight the need for continued methodological refinement and field-based validation to enhance the robustness, reliability, and scalability of remote sensing–based rodent monitoring systems.

Conclusion

This study indicates that integrating high-resolution satellite imagery with optimized spectral indices and predictive habitat modelling holds considerable potential as a scalable tool for monitoring fossorial water vole populations. The proposed approach shows promise for inferring vole abundance across damage gradients, even where direct field observations are limited, and may offer biologically meaningful and temporally consistent insights into population dynamics. By addressing some key limitations of traditional vegetation indices through the development of the Optimized Damage Index (ODI), this methodology could improve the precision and reliability of remote sensing–based assessments under varying climatic conditions. The findings suggest several potential practical implications for agricultural and environmental management: (1) Early Detection and Targeted Control: by identifying high-risk habitats and periods of peak vole activity, land managers could implement control measures more efficiently and reduce the risk of widespread crop damage; (2) Cost-Effective Surveillance: the use of freely available Sentinel-2 data may reduce the need for extensive fieldwork, supporting large-scale monitoring with limited resources; (3) Integration into Decision-Support Systems: the models developed could potentially be incorporated into regional pest management platforms to inform real-time decisions based on spatial risk predictions; and (4) Adaptability to Other Species or Regions: with further refinement, the framework may be adaptable for monitoring other herbivorous pests or for application in different agroecosystems facing similar challenges.

Author contributions

Aitor Somoano conceptualized and designed the study, coordinated the fieldwork, compiled and organized the field data, and drafted the initial manuscript. Aitor Somoano and Ana del Cerro integrated feedback from all co-authors and secured funding for both the research and its publication. Luis Varona, Antonio Rubio, and Carlos Tarragona conducted the remote sensing analyses and developed the mathematical models. All authors contributed to the interpretation of the results and approved the final version of the manuscript.

Funding

This work was supported by Dirección Xeral de Gandaría, Agricultura e Industrias Agroalimentarias de la Consellería do Medio Rural de la Xunta de Galicia, Dirección General de Medio Natural y Planificación Rural de la Consejería de Medio Rural y Cohesión Territorial del Principado de Asturias, AGROALNEXT program (Ministerio de Ciencia, Innovación y Universidades, Spanish Government) with funding from European Union NextGenerationEU (PRTR-C17.I1), and by GRUPIN NySA IDE/2024/000764 (PCTI 2024–2026).

Data availability

Field data collected in this study are available from the corresponding author, free of charge, upon reasonable request for the purpose of reproducibility and validation. The environmental datasets used for the development of the mathematical models are openly accessible from public repositories. Satellite imagery data were obtained from the Sentinel-2 mission through the Google Earth Engine platform (https://developers.google.com/earth-engine/datasets/catalog/COPERNICUS_S2_SR_HARMONIZED?hl=es-419) which provides free access for research purposes. The same *Sentinel-2* data can also be directly accessed via the Copernicus Open Access Hub (https://browser.dataspace.copernicus.eu). Meteorological variables were retrieved from the OpenMeteo database (https://ddei5-0-ctp.trendmicro.com:443/wis/clicktime/v1/query?url=https%3a%2f%2fopen%2dmeteo.com%2f&umid=4D7DBFFD-40CB-7F06-A49A-1CD615BDC920&auth=78cfaeb67d6fcbc33fb1b0a50590653357551b54-1483a9f711408a49e1f4cd67b10d473cba6ed06f), which offers freely available global weather and climate data. The Damage Index (DI) and the Optimized Damage Index (ODI) that partially support the findings of this study are available from Spectralgeo but restrictions apply to their availability, which were used under license for the current study, and so are not publicly available. Mathematical models are however available from the authors upon reasonable request and with permission of Spectralgeo.

Declarations

Competing interests

L.V., A.R., and C.T., affiliated with Spectralgeo, hold the intellectual property rights to the Damage Index (DI) and the Optimized Damage Index (ODI) presented in this study and have interests related to their use. The remaining authors (A.S. and A.dC.) declare no conflicts of interest.