Open and Free Sentinel-2 Mowing Event Data for Austria

Abstract

The accurate detection of mowing events is important in many applications, including in agricultural contexts such as yield and fodder production, as well as biodiversity assessments, habitat modeling, and protected area monitoring. This work presents the first free and open dataset of mowing events covering the entire Austrian territory for the year 2023 at a spatial resolution of 10 × 10 m. We use the Sentinel-2 time series of the Normalized Difference Vegetation Index (NDVI) to detect mowing events, and additionally, we use the mean of the two ShortWave InfraRed (SWIR) bands to exclude misclassification caused by remaining cloud artifacts and shadows. The validation procedure builds on a visual interpretation of the Panomax webcam archive complemented by a selection of field observations. The final validation dataset consists of 211 mowing events recorded in 85 different locations across Austria. In total, 77.73% of these mowing events were detected with a mean time delay of 4 days. The detection delay in summer was smaller than the values recorded in spring and fall. The pixel-based approach exhibited superior efficacy, especially for meadows with three or more mowing events, compared to the polygon-based approach. The results of our study are consistent with those of previous works demonstrating the capacity to produce high-quality mowing event data for various grassland areas in a fully automated manner, independent from training datasets. The results could be used in research on biodiversity or in practical applications such as agricultural policy support and control, fodder supply evaluation, or impact assessment in nature restoration efforts.

1. Introduction

Grasslands play a significant role in land use in Austria, comprising 1.2 million hectares of land in 2020, which accounts for approximately half of Austria’s agricultural lands [1]. These grasslands serve multiple essential functions, including erosion prevention, carbon storage, the maintenance of soil quality and water regulation, the provision of fodder for livestock, and supporting diverse habitats for various species. Therefore, understanding mowing events and intensities is crucial in applications such as fodder yield and quality assessment, biodiversity evaluation, habitat modeling, and protected area monitoring.

Several studies in recent years have demonstrated the potential of remote sensing in detecting mowing events [2,3,4,5,6]. A comprehensive review by Reinermann et al. [7] revealed that around 40% of grassland-related studies focused on mowing and grazing management options. However, previous studies on mowing event detection were either limited to small areas [3,4,5,7] or conducted on a polygon level rather than on an individual pixel level [4,6]. In most cases, these polygons were derived from the European Union’s common agricultural policy (CAP) integrated administration and control system (IACS). Some studies relied on extensive training data, which can be challenging and costly to obtain [3,6]. In terms of spectral features, 46% of the studies investigated mowing events using remote sensing data by employing the Normalized Difference Vegetation Index (NDVI) [7]. Furthermore, studies that provided pixel-level results for larger areas using automated methods were unable to offer independent validation data and relied on the time series itself to interpret the mowing events [2]. Such a validation approach makes it impossible to calculate the detection delay for the mowing events. Further, not all datasets are open and freely available.

The research questions of this study are as follows:

(1)

What detection rate for mowing events can be expected from a fully automated, wall-to-wall, nationally applied approach independent from training data?

(2)

Is a pixel-based or a polygon-based application more suitable?

(3)

What is the detection delay overall and for the individual cut dates?

To address these research questions, we produced and evaluated the first dataset of mowing events for the entire Austrian territory in the year 2023, at a 10 × 10 m resolution, which is freely and openly accessible. Our methodology builds upon the approach introduced by Kolecka et al. [5], with some adaptations and extensions. We also compare the results derived per pixel and per IACS polygon for a smaller test site. To validate the national product, we use the completely independent Panomax webcam archive to digitize reference meadows and record the corresponding mowing events. Additionally, selected field observations are incorporated into the validation dataset. One main step forward is to show which accuracy is achievable without training data for a whole country and using independent validation data. The scientific sub-objectives are threefold: first, to show the benefit of including the SWIRMean into the analysis, second, to compare pixel-based and polygon-based approaches, and third, to analyze the detection delay for different seasons. The practical significance of our study output is the national wall-to-wall coverage, the provision of the results freely and openly, and the easy repeatability for other years, as no costly training data are needed.

2. Materials

2.1. Study Area and Input Data

This study is conducted in Austria, employing a wall-to-wall approach covering the entire territory, which spans approximately 84,000 square kilometers, encompassing flat lowlands in the east to the Central Alps in the western provinces. The diverse topography and different climatic conditions result in a wide range of grassland types and utilization intensities across Austria. The climatic conditions range from the Continental-Pannonian climate in the east and Illyrian climate in the south-east to the harsher Alpine climate in the main parts of central Austria. The Central European transitional climate dominates the north and we find some areas with Atlantic influences in the west. Grasslands are found in about 90% of the Austrian farms, but the main focus of dairy farming is in the Alpine and northern regions [8]. In this topographically, pedologically, and climatically diverse setting, favorable production sites and disadvantaged sites are often closely intertwined [6]. In combination with small ownership structures, where two thirds of the farmed units of grasslands are smaller than 0.5 ha [8], a large variety of grassland use intensity can be expected. Unlike other studies that focused solely on cultivated grasslands with at least one cut per year [6], our approach processes the entire country in a wall-to-wall manner, thus allowing a broader range of applications to benefit from our results. This includes biodiversity assessments in areas such as Natura 2000 sites or alpine grasslands, which are seldom cut. For this study, we utilized all Sentinel-2 datasets from the year 2023 available through the Copernicus Data Space Ecosystem (CDSE), accessible at https://dataspace.copernicus.eu/ (accessed on 5 August 2024).

2.2. Validation Data

Validation data were derived from the freely accessible Panomax webcam archive (www.panomax.com, accessed on 18 February 2025), which currently comprises 830 active cameras worldwide, with 721 of them being accessible to the public (as of 6 December 2024). In Austria, there are approximately 200 cameras; however, not all of them were operational in 2023 or do not display archived imagery from that year. Furthermore, many of the webcams are situated in urban areas or high-altitude regions, limiting the number of usable cameras for mowing event detection.

Using the available webcam data, the precise date and even time of the cutting event can be determined through visual interpretation, as most Panomax webcams store images in a 15 or 30 min interval. The meadows visible in the webcams were digitized within a Geographic Information System (GIS) environment, and the corresponding mowing events were recorded as related attributes. Additionally, field observations were included in the dataset. In total, 211 mowing events were documented across 85 mapped areas throughout Austria, with the majority located in Carinthia (29%), followed by Salzburg (24%) and Styria (17%). For three provinces (Burgenland, Vorarlberg, and Lower Austria), only one area per province is currently available (see Figure 1).

2.3. IACS Data for Comparison

The IACS dataset typically includes the LPIS (Land Parcel Identification System) and GSAA (GeoSpatial Aid Application) data, both of which constitute spatial information that supports IACS as a spatial reference for aid applications under the Common Agricultural Policy (CAP). The LPIS data comprises georeferenced blocks of agricultural areas that have been identified and digitized primarily from ortho-imagery. These areas are potentially eligible for EU aid application. On the other hand, GSAA consists of georeferenced parcels that are in agricultural use and for which aid has been applied. Therefore, GSAA is usually a component of LPIS. In the context of Austria and this study, we exclusively consider LPIS as the most comprehensive dataset within the IACS 2023 dataset used. The IACS categories used in this study as well as their definitions are given in

Table 1. IACS categories and their definitions [9].

IACS Category	Definition
pasture (Hutweide)	The pasture is a low-yielding, grazed permanent grassland (usually without maintenance cutting) on which mechanical fodder production or maintenance is not possible or is not carried out due to the nature of the soil. These areas must be fully grazed at least once per year.
mountain meadow (Bergmähder)	Mountain meadows are extensive mowing areas above the local permanent settlement boundary, whereby these areas must be above the altitude of the home farm and generally not directly adjacent to home farm areas of the same farm. The majority of the area must be above 1200 m.a.s.l. Mountain meadows must be fully mown at least once every two years and the mown material must be removed.
single-mown meadow (Mähwiese 1)	Single-mown meadows are areas on which the entire surface is mown once a crop year and the mown material is removed from the area.
mown meadow/pasture with two uses (Mähwiese 2)	Mown meadows/pastures with two uses are areas on which full-surface mowing is carried out either twice (including the removal of the mown material) or only once but combined with a full-surface grazing in the same year. A selective maintenance cut does not count.
mown meadow/pasture with three or more uses (Mähwiese 3)	Mown meadows/pastures with three or more uses are areas on which full-surface mowing is carried out either three times (including the removal of the mown material) or a combination of one or two mowing events (including the removal of the mown material) with full-surface grazing in the same year to reach at least three full-surface uses.

. Please note the ambiguity between the IACS category definition and the actual number of mowing/use events.

3. Methods

3.1. Mowing Event Detection Method

Similar to prior studies [5,10], the approach relies on time series data from Sentinel-2 to identify declines in the time series curves, with stringent pre-processing being crucial for successful mowing detection. This pre-processing includes atmospheric correction, cloud masking, and topographic correction. For atmospheric correction, we used Sen2Cor [11,12]. The Sen2Cor processor has enhanced the ability to convert Level-1C data (top-of-atmosphere reflectance) to Level-2A (bottom-of-atmosphere reflectance), including maps of aerosol optical thickness, water vapor, and scene classification for clouds and snow [13]. Sentinel-2 data undergo rigorous geometric correction using a Global Reference Image (GRI) database and automatic ground control point matching, improving geolocation accuracy to sub-pixel levels [14]. Cloud masking was done using Function of Mask (FMASK) [15], an automated algorithm that detects and masks clouds, cloud shadows, and snow in optical satellite imagery. Finally, topographic normalization is used to correct terrain-induced brightness variations in satellite imagery. The approach applies iterative classification to spectrally distinct clusters and derives specific Minnaert constants for each. This allows fully automatic, reference-free correction of topographic effects, which is essential for large-scale processing in mountainous areas [16,17]. In addition to previous methods [5], the ShortWave InfraRed (SWIR) bands are utilized alongside the Normalized Difference Vegetation Index (NDVI). The NDVI is calculated as the ratio between the difference and sum of the NIR and the red band (NDVI = (Band 8/NIR − Band 4/red)/(Band 8/NIR + Band 4/red). The two SWIR bands at a spatial resolution of 20 × 20 m are band 11 with a central wavelength at 1610 nm and band 12 with a central wavelength at 2190 nm. The SWIRMean is defined as the mean pixel value of the two SWIR bands.

The method initially examines the NDVI time series to identify significant drops, with a drop greater than 0.2 in NDVI flagged as a potential mowing event [5]. To account for noise or minor vegetation changes, two consecutive smaller NDVI drops, each less than 0.04, are also considered as potential mowing events. However, NDVI drops do not only occur due to mowing events, but also due to unmasked shadows, haze, or thin cirrus clouds, which reduce visible and near-infrared reflectance, leading to a spurious decline in NDVI. When NDVI drops due to thin cirrus clouds, cloud shadows, or haze in the data, the SWIRMean typically remains unchanged or drops.

In contrast, a genuine mowing event results in a drop in NDVI and a simultaneous increase in the SWIRMean due to changes in vegetation structure and moisture content. This complementary behavior is used to differentiate between atmospheric artifacts and actual mowing events, enhancing the robustness of the detection approach. The behavior differences in a ‘look-alike’ and a real mowing event are shown in Figure 2. For the NDVI drop on 30 May (marked with a black vertical line), there is no corresponding SWIRMean increase, leading to the filtering out of this drop as a ‘look-alike’ mowing event. In contrast, the detected mowing event on 19 June (marked with a green vertical line in Figure 2) shows the typical behavior of both indices and was also identified in the Panomax reference images on 11 June (red line). In addition to this visual check, we also tracked the amount of ‘look-alike’ mowing events for the small test site to see how much the SWIRMean test actually sorted out. We found 226.896 valid mowing events and 236.274 ‘look-alike’ events. Based on these numbers, the importance of the SWIRMean as an additional criterion becomes evident. While the SWIRMean effectively helps mitigate misclassifications caused by atmospheric artifacts such as haze, shadows, and thin cirrus clouds, its performance may be affected by complex terrain. In mountainous areas, topographic effects—such as varying slopes and aspects—can influence reflectance even after correction. Although the applied Minnaert correction reduces terrain-induced illumination variability, residual effects may remain.

Upon detecting a mowing event, a regeneration phase of two weeks is assumed [5], during which no further mowing events are identified. This temporal buffer allows for the natural recovery of vegetation and prevents the misclassification of subsequent vegetation changes within this timeframe as separate mowing events. The outputs of the mowing detection process include the total count of mowing events, the dates of (a maximum of six) detected mowing events, and the number of valid observations available in the time series for each detection, providing a comprehensive overview of mowing activity during the analyzed period.

The method is designed to accommodate both vector and raster data. When applied to vector data, mowing events can be detected within predefined field boundaries, such as those provided by the IACS system. For pixel-level analysis, the method identifies mowing events on a per-pixel basis. Nonetheless, additional post-processing is recommended to smooth the results and minimize noise, ensuring spatial consistency and accuracy in the detection outcomes. We applied a simple 3 × 3 majority filter. The dual capability allows the method to be flexibly applied across different spatial scales while maintaining robustness in mowing event detection. Although the method is capable of working both retrospectively and in near-real time, it was not processed in near-real time for this study; thus, differences in accuracy from real-time processing cannot be assessed.

3.2. Validation Method

Ensuring the accuracy of the analysis, the validation of mowing detection results is crucial. This process involved leveraging completely independent and high-resolution geotagged imagery from Panomax webcams (www.panomax.com, accessed on 18 February 2025) to identify and delineate mowing areas. Figure 2 illustrates the validation workflow, demonstrating the integration of webcam observations with geographic data. As depicted in Figure 3A, the georeferenced location of the Panomax webcam and its field of view were analyzed to outline specific mowing areas. These visible areas in the imagery were digitized in QGIS as polygons to represent the observed extent of mowing activities. The alignment of the polygons with the actual mowing fields was facilitated using Panomax’s satellite map and live-view imagery (Figure 3B).

To enhance the precision of the analysis, a 5 m inverse buffer was applied to each polygon. This adjustment aimed to minimize the influence of potential inaccuracies from digitization and the effect of mixed pixels by focusing on the core regions of the validation polygons.

To evaluate the detection rate and delay, the most frequently detected mowing date in each mapped meadow was compared to the reference mowing date. The reference mowing event was confirmed when the grassland in the Panomax image visibly changed from tall, uniform vegetation to a cut appearance—often accompanied by visible swath rows, drying hay, or mowing machinery. The meadow areas visible in the webcams were digitized as polygons based on perspective-corrected aerial views from the webcams static map layer. We then linked each observed mowing date to the corresponding polygon. For the 24 field datasets and the three EU grassland watch observations, a window of −3 days was allowed to still be considered correct, as during the field survey, it was not always possible to ascertain if the mowing occurred exactly on the day of observation or a few days earlier. For logistical reasons, no field surveys were carried out over the weekends, which means mowing events from, e.g., a Friday evening, may only have been surveyed and reported on the following Monday. This procedure to allow a −3 day window during the analysis aligns with the methodology employed in previous works [6,18]. Any detections that occurred more than three days before the v date were classified as ‘not detected’. Moreover, any detections with a delay exceeding one of the specified thresholds (30 days, 15 days, and 7 days) were also deemed ‘not detected’ in the respective statistics. For the remaining detections, the delay [days] in comparison to the validation date was calculated.

To assess commission errors, a full screening of the entire Panomax archive inspecting each individual day is necessary. Due to the time-consuming nature of this full screening, we randomly selected 20 out of the 85 areas for the full screening. For these 20 areas, we screened each individual day and evaluated grazing events alongside the mowing events.

3.3. Method for Comparison with IACS

We utilized five categories from the IACS dataset [9], and the definitions for these categories are provided in Section 2.3. However, these definitions do not allow a clear assessment of the number of mowing events per category in the year 2023. Therefore, we had to make assumptions regarding the typical number of mowing events for the purpose of comparison:

-

No mowing events were assumed for pastures (Hutweide).

-

One annual mowing event was assumed for mountain meadows (Bergmähder).

-

It was assumed that there was one mowing event per year for single-mown meadows (Mähwiese 1).

-

Two annual mowing events were assumed for mown meadows/pastures with two uses (Mähwiese 2).

-

It was assumed that mown meadows/pastures with three or more uses (Mähwiese 3) experienced three or more mowing events per year.

It is important to note that this estimation of mowing events per category may not be entirely reliable, as there are cases where mountain meadows, for example, are mown twice a year, while others are mown only every second year. Therefore, we consider this analysis as a plausibility check rather than an accuracy assessment.

4. Results

4.1. Mapping Result

We applied the pixel-based approach to process Austria wall-to-wall, and the resulting map clipped to IACS categories (as listed in Table 1) is presented in Figure 4. The map effectively illustrates the more intensively used grassland areas in the northern foreland, predominantly in Upper Austria, and in the inner-Alpine valleys compared to the less intensively used areas in Carinthia, for example.

4.2. Accuracy Assessment Based on Validation Dataset

We conducted a comparison between the results of our pixel-based mowing event detection for Austria and all the validation data obtained using the methods described earlier. An example for the successful detection is shown in Figure 5 with all reference events detected with a maximum delay, in this case, of 10 days.

The findings revealed a detection rate of above 80% (80.57% for a maximum delay of 15 days and 81.99% for a maximum delay of 30 days). The minimal difference between the 15- and 30-day delays indicates that most mowing events are typically detected within 15 days. Regarding detection delay, the average delay (mean absolute error, MAE) for all detections was approximately 4 days (4.07 for 15 days and 4.30 for 30 days, as shown in

Table 2. Mean absolute error (detection delays) for individual cuts and all cuts [days] and overall detection rate.

Maximum Allowed Delay	Delay in Cut Detection (Mean Absolute Error-MAE) [Days]						Detection Rate [%]
	1	2	3	4	5	All Cuts
30 days	5.3	3.3	4.1	3.9	5.7	4.30	78.67
15 days	5.3	3.3	3.5	3.2	5.7	4.07	77.73
7 days	3.28	2.6	2.5	2.9	3	2.84	63.98

). Within 7 days, 63.98% of the mowing events were detected with an MAE of 2.8 days. This result is noteworthy considering the global 5-day revisit time of Sentinel-2, although it is important to note that the revisit time is not globally consistent. Due to overlapping orbits at higher latitudes, certain areas in Austria receive a 3-day revisit time [19].

Further, we assessed the delay for individual cuts separately. Table 2 shows a slightly higher delay for the first and the last cut compared to the summer cuts (for both 15- and 30-day maximum allowed delay). For the 7-day maximum allowed delay, the third cut shows the smallest delay. The timeliness of mowing event detection strongly depends on the available number and temporal density of cloud-free observations. Areas and/or periods of high cloud cover show longer time lags between the mowing event and the next available interpretable optical satellite image. Previous studies have shown significant spatial (e.g., from single or double Sentinel-2 orbit coverage) and temporal (inter-annual and seasonal) differences in the availability of cloud-free Sentinel-2 data. A recent study on mowing event mapping with Sentinel-2 and Landsat-8 at national scale in Germany reported differences in mowing detection accuracies related to the inhomogeneous availability of cloud-free datasets [18].

For the randomly selected 20 areas of full screening, we evaluated grazing events alongside the mowing events. We found an average commission error of 0.6 events when considering grazing as a land use event and an average commission error of 0.9 events when grazing is not considered an event (see

Table A1. Full evaluation of 20 Panomax areas also assessing grazing activities and commission error.

Location	Panomax Mowing Events	Panomax Grazing Events	Detected Events	Commission (Only Mowing)	Commission (Incl. Grazing)
Griffen	4	0	4	0	0
Pyhrn	5	0	5	0	0
Pertisau	3	0	5	2	2
Liebenau	4	0	3	0	0
Nassereith	2	1	3	1	0
Ramsau im Zillertal	3	1	4	1	0
Pertisau	3	0	4	1	0
Scharnitz	2	1	4	2	1
Donnersbachwald	2	1	3	1	1
Altenmarkt 1	4	0	5	1	0
Altenmarkt 2	4	1	5	1	1
Mondsee	5	0	4	0	0
Bad Mitterndorf	4	0	3	0	0
Großarl 1	4	1	5	1	0
Großarl 2	4	0	6	2	2
Großarl 3	4	0	6	2	2
Großarl 4	3	0	5	2	2
Westendorf 1	4	1	4	0	0
Westendorf 2	4	0	4	0	0
Westendorf 3	3	1	3	0	0
Average				0.9	0.6

in Appendix A).

4.3. Plausibility Check in Comparison with IACS Dataset

Due to the unclear number of mowing events per category (see Section 2.3), we can consider this comparison as a plausibility check rather than an accuracy assessment.

(a)

Pixel-based versus polygon-based for the test site only

To evaluate the suitability of a polygon-based versus a pixel-based approach, we conducted a comparative analysis using a test area in Carinthia (as indicated by the purple outline in Figure 1). Within this area, we separately processed the region using the same settings and methods for polygons and individual pixels. Due to the small number of Panomax webcams and field data, a full-fledged accuracy assessment was not representative. Instead, we compared the results only to the IACS data. Figure 6 displays the percentage deviation of the summed mowing event areas from the IACS (referred to as the reference values) for both polygon-based and raster-based results. In this visual representation, each bar represents the deviation from the IACS, with positive values indicating overestimation (areas with more mowing events than assumed using the IACS categories) and negative values reflecting underestimation (areas with fewer mowing events than assumed using the IACS categories).

The results revealed substantial overestimation in the category “mown meadows/pastures with three or more uses” by the polygon method, reaching nearly 600 percent. This indicates that the polygon method classifies approximately six times more area in this category than the IACS. Such a significant overestimation serves as an indicator of substantial commission errors in the polygon-based approach. In contrast, other categories exhibited smaller deviations, with both methods often underestimating areas, particularly in cases like pastures and mountain meadows, where we assumed “no mowing or other use event” in accordance with the IACS.

Across all categories, the pixel-based method demonstrated smaller deviations compared to the polygon-based approach. These findings align with the results of the accuracy assessment, further supporting the conclusion that the pixel-based approach is superior to the polygon-based method.

(b)

Pixel-based comparison with IACS for whole of Austria

Building upon the earlier findings that favored the pixel-based approach for its superior performance, we expanded the comparison conducted in the small test site to encompass the entirety of Austria using the pixel-based approach. The deviations per category are displayed in Figure 7, following the same format as Figure 6.

Notably, similar deviations are observed for category “mountain meadows” (underestimation) and for category “mown meadows with two uses” (overestimation) compared to the small test site. However, distinctive behaviors are observed for other categories. In the small test site, pastures were underestimated, yet for the full Austrian coverage, there is a slight overestimation in area. This same pattern is also observed for the categories “single-mown meadows” and “mown meadows with three or more uses”. Notably, the latter two categories exhibit excellent results with minimal deviations evident in both the test site and the full coverage.

5. Discussion

This study has explored the feasibility of achieving nationwide and fully automated mowing event detection using Sentinel-2 time series data. Our findings exhibit similar detection delay and overall accuracy to other studies, both with (f-score of 0.79 for [6], 60% for [10]) and without training data (85% for [3], 77% for [5]). Since the overall detection rates are very similar, it seems that our relatively simple, physically determined approach is able to generate the same quality as high-end deep learning tools. Notably, previous studies [4,10] highlighted challenges related to data availability and cloud interference in mowing event detection. We found a moderate relationship between mean monthly cloud cover and detection success. Monthly cloud fraction data from the Sentinel-5P satellite (TROPOMI instrument) were obtained as Level-3 gridded product in NetCDF-format from the Copernicus S5P-PAL portal (https://maps.s5p-pal.com/cloud-fraction/ accessed 28 April 2025). The original grid was interpolated to a regular 0.05° × 0.05° (~5 km) resolution and average cloud cover values were calculated for all reference data locations. The two months with the highest detection rates (August and September) show low (September) and moderate (August) mean cloud cover values (see

Table 3. Comparison of mean monthly cloud cover over reference sites and detection success per month.

Month	No. of Reference Mowing Events	Detected Mowing Events [%]	Monthly Mean Cloud Cover over Reference Sites [%]
April	1	0	80
May	40	50	65
June	46	85	59
July	41	73	53
August	34	97	50
September	30	100	30
October	19	68	48

). It must be noted that such an analysis has some problems: First, in the case that the reference event occurs towards the end of one month, the satellite data acquisition may already fall into the subsequent month. In that case, the mean cloud cover of reference month would actually not affect detection. Second, in a mountainous country like Austria, cloud cover can vary significantly according to region. Mean monthly cloud cover is not able to depict these regional differences. Figure 8 depicts an example to better understand the effects of missing data. The first mowing event on 25 May (Figure 8, first red vertical line) was preceded by a long period of clouds and, therefore, missing Sentinel-2 data (Figure 8). The event could not be detected by the subsequent observation, as the NDVI drop was too small and there was no second observation to compare to. Similar situations were demonstrated in prior research [5].

Other studies integrated SAR data to mitigate the cloud cover problem [20,21,22,23]. One study showed slight improvements (MAE 4.1 from 4.6 and f-score 0.81 from 0.77) based on SAR and weather data integration for meadows larger than 2 ha [21]. Other authors found improvements of 10% by including SAR data [20] with a deep learning system for selected test sites in Germany. Our main reason for relying solely on optical data is the wall-to-wall requirement and the Alpine topography of Austria, which poses issues in SAR signals [22]. In our observations using the polygon-based approach, we also encountered excessive detections, as depicted in Figure 6. However, our meticulous preprocessing in conjunction with the pixel-based assessment notably mitigated these false detections (see Figure 6, pixel-based approach). Similar findings were previously noted by Kolecka et al. (2018, [5]). Nevertheless, we recognize that data availability varies annually due to weather patterns, which may still impact detection accuracy, as shown in Figure 8.

The resulting map from our study (see Figure 4) effectively illustrates the distribution of intensively and extensively used grasslands across Austria, with the most heavily used areas situated in the northern provinces of Upper and Lower Austria, followed by the south-eastern foreland of Styria. This distribution closely aligns not only with previous studies [6], but also with the official statistics from Statistics Austria [24], which indicate that these three provinces have the highest shares of intensively used grasslands (see Figure 9).

In our accuracy assessment, we deliberately employed much stricter criteria compared to previous authors, who allowed a three- [18] or even a six-day [10] window before an event to be considered correct, whereas we only allowed three days in case of field visits and none for the Panomax data. Our results robustly reinforce their assertion of the superiority of the pixel-based approach over the polygon-based approach (Figure 6 as well as Figure 10), aligning with the findings of other studies [5]. This preference may stem from factors such as (i) the presence of mixed and non-grass pixels within the polygons, including trees, hedges, or artificial surfaces [5] or (ii) different mowing patterns within a single polygon.

Regarding detection delay (MAE in days), Reinermann et al. (2023, [7]) demonstrated the capability to detect 60.3% of all mowing events within 7 days, with an average delay of 2.5 days, using solely Sentinel-2 data. Their reference data were also independent webcam data but building on a small set of reference areas for parametrization. Our findings closely mirror this, with 68.2% of events detected within 7 days and an average delay of 2.6 days. Moreover, with a permissible delay of 15 days, the detection rate exceeds 80%. In a related study, Kolecka et al. (2018, [5]) reported an overall accuracy of slightly less than 80% based on visual interpretation of the utilized Sentinel-2 data. Our results fall within the same range, even when utilizing a wholly independent validation dataset from the Panomax network.

Our study is limited in several ways. First, we recognize that the available Panomax data are not distributed evenly and there might be a regional bias to the results, as the provinces of Lower Austria, Burgenland, and Vorarlberg with significant grassland extent are not well covered. In the case of Vorarlberg, we do not expect significant impacts on the overall validity, as we have good coverage of Tyrol, a similar region. However, for the relatively dry eastern zones of Austria (Lower Austria and Burgenland), there might be more impact on the results once including more reference data from this area. The second study limitation concerns the transferability to other countries and regions. Although covering the whole of Austria means to cover a large variety of grassland types and conditions, our study results are not automatically transferable to other areas. A limitation might be found in the thresholds used for both NDVI drop and SWIRMean increase for much drier or wetter regions. Third, the processing of only one year is a further limitation to the general validity of our findings. This is specifically true for years with extreme weather conditions such as, for example, 2024, when heavy rainfall caused flooding, which could potentially jeopardize the accuracy of mowing event detection. This is, of course, not only true for our approach, but also for other methods.

Therefore, there is scope for future work and improvements. One future work direction would be to expand the reference data to gain a better understanding of the remaining errors. It is important to analyze how accuracy varies with elevation and/or regions. A second future expansion would be to address challenges such as cloud cover and related missing measurements over time by combining Sentinel-2 data with other sensors, such as Planet Scope data, although output quality is expected to be much lower due to the missing SWIR bands. The results for the integration of SAR data, as shown in previous works, are encouraging and should also be further explored. Furthermore, carrying out the processing for additional years would improve the assessment of the method’s robustness and its sensitivity to changing weather patterns and evolving management practices. Lastly, conducting a comprehensive comparison of methods (in line with the concept of the mowing detection intercomparison exercise, MODCiX [25]), for a consistent year and the same region would be an important step forward. For this comparison, it is essential to employ independent data for accuracy assessment (which were not utilized for training), as this would represent a significant advancement in effectively comparing and evaluating the applicability of various methods.

6. Conclusions

The main aim of this study was to show, based on independent validation data, how well a fully automated approach for mowing event detection from Sentinel-2 time series works in the diverse Austrian landscape. The method is based on a combination of NDVI drops and SWIRMean increase and works without any kind of training data. In conclusion, this study is an important step forward in mowing event detection, as it proves that detection accuracies comparable to those in the reviewed literature without costly and difficult-to-come-by training data are possible. The detection rate for mowing events reached almost 78% with 15 days of maximum allowable delay. Thanks to rigorous preprocessing schemes and the inclusion of the SWIRMean in the processing, the commission error is less than one (0.9 on average) event. A comparison confirmed previous statements that a pixel-based approach performs better than a polygon-based application. Additionally, the pixel-based approach enables the assessment of areas without polygon boundaries (e.g., outside IACS, see Figure 10) and can differentiate between mowing events on different days within a single IACS polygon (Figure 10). The detection delay throughout the year averages approximately 4 days, with slightly lower delays for cut dates in summer than for those occurring early or late in the season. This fact could mainly be attributed to cloud cover and related missing observations in spring and fall. Our study is limited by the application to one year only, the focus on Austria only, and the available amount of validation data, which could bias the result. Further work should, therefore, focus on processing multiple years, larger areas, and assessing the accuracy based on more validation data for individual regions related to varying weather and growing conditions.

Our results are shared freely and openly to allow for numerous follow-up applications. They can be used for habitat mapping, biodiversity assessments, evaluating usage intensity, or control adherence to specific regulations, such as late mowing (e.g., after 15 June to support biodiversity). We hope this comprehensive study will contribute to advancing the understanding of grassland dynamics and aid in informed decision making for sustainable land-management practices.