Assessment of Temporal Aggregation of Sentinel-2 Images on Seasonal Land Cover Mapping and Its Impact on Landscape Metrics

doi:10.21203/rs.3.rs-4723895/v1

Download PDF

Research Article

Assessment of Temporal Aggregation of Sentinel-2 Images on Seasonal Land Cover Mapping and Its Impact on Landscape Metrics

https://doi.org/10.21203/rs.3.rs-4723895/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Landscape metrics (LM) play a crucial role in fields such as urban planning, ecology, and environmental research, providing insights into the ecological and functional dynamics of ecosystems. However, in dynamic systems, generating thematic maps for LM analysis poses challenges due to the substantial data volume required and issues such as cloud cover interruptions. The aim of this study was to compare the accuracy of land cover maps produced by three temporal aggregation methods: median reflectance, maximum normalised difference vegetation index (NDVI) and a two-date image stack using Sentinel-2 (S2), and then to analyse their implications for LM calculation. The Google Earth Engine platform facilitated data filtering, image selection, and aggregation while mitigating cloud cover effects. A random forest algorithm was employed to classify five land cover classes across ten sites, with classification accuracy assessed using global measurements and Kappa index. LM were then quantified. The analysis revealed that S2 data provided a high-quality, cloud-free dataset suitable for analysis, ensuring a minimum of 25 cloud-free pixels over the study period. The two-date and median methods exhibited superior land cover classification accuracy compared to the max NDVI method. In particular, the two-date method resulted in lower fragmentation-heterogeneity and complexity metrics in resulting maps compared to the median and max NDVI methods. Nevertheless, median method holds promise for integration into operational land cover mapping programs, particularly for larger study areas exceeding the width of S2 swath coverage. These results highlight the importance of appropriate temporal aggregation techniques when using satellite data for landscape analysis and monitoring.

Sentinel 2

Google Earth Engine

Data summary

Random Forest

Pixel compositing

Landscape metrics (LM) serve as important tools in disciplines such as urban planning, ecology, and environmental research (Nowosad and Stepinski 2019; Altieri, Cocchi, and Roli 2018). They provide insights into the structure and spatial arrangement of land cover, which are crucial for understanding the ecological and functional dynamics of ecosystems. This understanding is essential for valuing ecosystem benefits and different land cover configurations (Abdollahi et al. 2023).

To calculate landscape metrics, various algorithms are used to measure specific spatial characteristics of the landscape mosaic, its patches, classes and surroundings (McGarigal 2015). These metrics include patch density, the ratio of patch perimeter to area, and the average distance to the centre of each patch. Combined, these metrics describe landscape characteristics such as heterogeneity (Awuah et al. 2018; Tran, Julian, and De Beurs 2014), degree of fragmentation (Silva et al. 2018), and structural change (Statuto, Cilis, and Picuno 2018). Accurate thematic maps are crucial for improving the reliability of these metrics (Awuah et al. 2018; Shao and Wu 2008).

The production of accurate thematic land cover maps faces two major challenges. Firstly, dealing with the large volumes of data that need to be acquired and processed (Amani et al. 2019; Azzari and Lobell, 2017; Okoro et al. 2016). Second, addressing information gaps due to cloud cover interference (Griffiths, Nendel, and Hoster 2019; Orynbaikyzy, Gessner, and Conrad 2019). Temporal aggregation of satellite data has emerged as an approach to overcome these difficulties (Phan, Kuch, and Lehnert 2020).

Temporal aggregation, a pixel-level image fusion method (White et al. 2014), combines individual pixel values from each input dataset into new composite pixel values. The median is often chosen as the summary statistic for temporal aggregation due to its robustness to extreme values such as those caused by clouds and shadows (Bey et al. 2020). In this method, the output is calculated pixel by pixel, with each pixel containing the median of all images in the collection at that location.

Differences in phenology between land cover categories can lead to high spatial variance (Nolasco, Bocco, and Scavuzzo 2018). To minimise this variance, pixel values in the composite can be set to approximately the same phenological stage, such as the time of maximum plant greenness as indicated by the maximum NDVI value (Fontana et al. 2009). Alternatively, land cover mapping can be performed using a subset of images or a two-date stack, which provides good accuracy for land cover classification in dynamic systems (Rujoiu-Mare et al. 2017).

In this study, we aim to evaluate the accuracy and spatial configuration of land cover maps generated using three temporal aggregation methods applied to Sentinel-2 (S2) satellite data. Specifically, we evaluate the accuracy of land cover classifications using three easy-to-use temporal aggregation methods: median reflectance values, maximum NDVI, and a two-date image stack. In addition, we compare the spatial configuration of the resulting land cover maps.

2.1. Study area and sites

The study was conducted in 10 sites surrounding the metropolitan area of the Córdoba city, Argentina (31°20′S, 64°15′W) (Fig. 1). Each site was located in the peri-urban area and had a 1000-m radius buffer area. The study area is situated in the Pampas Loess Plains. The soils have good physical and chemical conditions for agricultural use. Annual mean temperature is 17° C and thermal amplitude is 14° C. Annual rainfall is 750 mm, increasing to the east, and is concentrated in spring and summer (Aliaga, Ferrelli, and Piccolo 2017) which is the growing season; therefore, it is often cloudy during the growing season.

2.2. Data collection, pre-processing, and temporal aggregation

The dataset was created using images from Sentinel-2 satellite constellation (Drusch et al. 2012). S2 carries a MultiSpectral Instrument optical sensor and is the data source most commonly used for medium–high resolution land cover mapping. In this study S2 level-1C product, available in the Google Earth Engine (GEE) web platform (Gorelick et al. 2017), was used. The level-1C product is composed of 100km² tiles of orthorectified and radio corrected images. Per pixel radiometric measurements are provided in Top of Atmosphere (TOA) reflectance. Only visible bands (2, 3, 4), red edge bands (5, 6, 7, 8A), near-infrared band (8), and SWIR bands (11, 12) of Sentinel-2 data were used with original spatial resolutions (10 or 20 m). S2 constellation covers the study area completely by two orbit paths. Images with zero or low (< 20%) cloud coverage, taken between December 2017 and March 2018 (growing season), were selected to construct the image collection. To evaluate cloud-free features of each image, the cloud information on the QA60 band was retrieved for each pixel in the study area during the analysed period. QA60 is a bitmask band with cloud mask information; cloud masks allow identification of dense clouds, also called opaque clouds, and Cirrus clouds, which are thin, transparent or semi-transparent clouds formed at high altitudes, approximately 6–7 km above the Earth surface.

Regarding calculation of the median of the reflectance value, GEE was instructed to pick the median value in each band in the stack. The function median() implemented in GEE was applied to the image collection. The latter aggregation method was based on the maximum NDVI value of the period and the qualityMosaic() function was applied to the image collection. This function works as follows: at the pixel level, the entire time series of images is examined and the pixel with the maximum value in the NDVI is set as the composite value. Thus, a single 10-band image is obtained for each aggregation function. NDVI was calculated using the red and near-infrared (NIR) bands:

$$\:NDVI=(\rho\:NIR\:-\:\rho\:RED)/(\rho\:NIR\:+\:\rho\:RED)$$

where ρ represents spectral reflectance and characterizes the “greenness” of the surface. NDVI band was added to every image in S2 data collection. Finally, a two-date dataset was created using single-date images. The four-month period was split in two, for the first and second two-month period, the image with the smallest cloud cover and covering the whole area was selected. Then the bands from both images were stacked, resulting in one image with 20 bands.

The analysis consists of 7 steps classified into 4 work boxes (Fig. 2). Box 1: pre-processing of input data to create TOA reflectance (Level 1-C) image collection and digitalization of reference field data; Box 2: data processing for temporal aggregation, which involves the creation of three composite datasets, and the generation of training and validation samples; Box 3: data classification, using Random Forest algorithm to map the target classes; Box 4: evaluation, first confusion matrices for assessing the classification accuracies are computed; then, landscape metrics are calculated.

Steps 1 to 5 were conducted on GEE platform, steps 6 and 7 were performed using “landscapemetrics” package in R (R Development Core Team 2023). Step 3 was carried out using the three different temporal aggregation strategies (median, maximum NDVI based-methods, and two-date stack). For the last, the image with less percentage cloud coverage in the first and in second two-month subperiod were selected.

2.3. Field reference data

The ground reference data were collected on two occasions: in the first two-month period and in the last two-month period of the growing season. Five thematic land cover categories were visually identified in the field: forest, crops, pasture, bare soil, and built area. These five categories were used as thematic classes in the classification. The field reference dataset was complemented with visual sampling through RGB color composite of Spot & Sentinel 2 data of the season, in conjunction with high resolution Google maps images.

Polygons were drawn over each reference site for each thematic land cover class, representing at least 200 random selected pixels (Table 1). These reference samples were split into two sets of disjoint training (80%) and validation samples (20%).

Table 1

Reference field data size (pixels) per thematic class at each studied site.
Class	Site 1	Site 2	Site 3	Site 4	Site 5	Site 6	Site 7	Site 8	Site 9	Site 10
Forest	764	2119	2703	635	388	748	420	704	1111	4788
Bare soil	592	825	550	345	320	281	290	437	629	651
Builded	248	300	278	442	441	518	333	344	248	289
Crop	272	4567	4987	1489	685	1081	1772	2901	1348	3445
Pasture	1700	1112	1112	1209	887	1587	695	717	896	869

2.5. Land cover classification and assessment

Land cover classifications were performed for each dataset by applying Random Forest classifier (Breiman 2001). The Random forest classifier was selected since it is less sensitive than other classifiers to the features of training samples and to overfitting (Belgiu and Drăgu 2016). The Random Forest nTree parameter was set to 100 (Rodriguez-Galiano et al. 2012; Wang, Azzari, and Lobell 2019) and Mtry to the square root of the number of input variables (Izquierdo-Verdiguier and Zurita-Milla 2020; Rufin et al. 2019).

The classifications were validated using k-fold cross validation (80:20 split ratio). The k-fold cross validation method (Lyons et al. 2018) involves partitioning the data once into k-folds; here, a 5-fold cross validation divides the data into five partitions. Then, in an iterative manner, each fold was used once as the test set while the remaining k − 1 folds formed the training sets. Differences between the validation data and the classifications were calculated using classification error matrices (Stehman and Foody 2019). The overall accuracy (OA) and the kappa coefficient (κ) were used to compare the classification performance obtained with datasets derived from each temporal aggregation method.

Finally, temporal aggregation methods were compared using a series of landscape metrics (McGarigal 2015). We select a suite of minimally redundant metrics (Fan and Myint 2014, Hargis, Bissonette and David 1998, Awuah et al. 2018, Luo et al. 2019, Zhou et al, 2020). All metrics (Table 2) were calculated using R 4.0.2, and the “landscapemetrics” (Hesselbarth et al. 2019) package. Thus, the spatial metric calculation was based on the land cover maps generated with each temporal aggregation method, in each site. The parameterization of the metrics consisted in defining the number of directions in which the patches should be connected equal to 8. For the calculation of conditional entropy, pairs were considered as ordered, and bits was used as measurement unit.

Table 2

Description of landscape metrics used in the study.
Landscape metric (abbreviation, unit)	Abbreviation & unit	Class	Level	Description
Patch density	PD (n°/ha)	Aggregation metric	Class	$\:PD=\frac{{n}_{i}}{A}10000100$ where n_i is the number of patches and A is the total landscape area in square meters.	(2)
Mean perimeter-area ratio	PARA (m/m²)	Shape	Class	$\:PARA=mean\left(PARA\right[{patch}_{ij}\left]\right)$ where $\:PARA\left[{patch}_{ij}\right]$ is the perimeter-area ratio of each patch.	(3)
Mean radius of gyration	GM (m)	Area – Edge	Class	$\:GM=mean\left(GYRATE\right[{patch}_{ij}\left]\right)$ where $\:GYRATE\left[{patch}_{ij}\right]$ is the radius of gyration of each patch.	(4)
Shannon’s diversity index	SHDI	Diversity	Landscape	$\:SHDI={\sum\:}_{i=1}^{m}\left({P}_{i}*ln{P}_{i}\right)$ where $\:{P}_{i}$ is the proportion of class $\:i$. where a_ij is the area of each patch and A is the total landscape area.	(5)
Contagion	CONTAG (%)	Aggregation	Landscape	$\:CONTAG=1+\frac{{\sum\:}_{q=1}^{{n}_{a}}{P}_{q}ln\left({P}_{q}\right)}{2ln\left(t\right)}$ where $\:{P}_{q}$ is the adjacency table for all classes divided by the sum of that table, and $\:t$ is the number of classes in the landscape	(6)
Conditional Entropy	CONDENT	configurational complexity	Landscape	$\:CONDENT=-{\sum\:}_{i=1}^{K}{\sum\:}_{j=1}^{K}p\left(x={c}_{i},y={c}_{j}\right)$ $\:p(y={c}_{i}\|x={c}_{j})\:\:$ The value of CONDENT is the number of bits needed on average to specify a class of an adjacent cell if the class of the focus cell is $\:{c}_{i}$. $\:p$ is the probability distribution.	(7)

3.1. Temporal aggregation

Cloud-free images were available for most of the dates. There was at least one completely cloud-free image every four consecutive dates (Fig. 3), covering a period of 7–13 days. Regarding the images with cloud evidence, there was an image with more than 90% of cloud-free pixels at least once every three consecutive dates, covering a period of 2–10 days (Fig. 3). All pixels had at least 25 cloud-free images in the four-month period under analysis (Fig. 4), with 28 being the most common value.

Table 3 shows the summary statistics for the information of spectral bands of each temporal aggregation method and studied site. Similar values of visible spectral bands were obtained for the two-date images and the median method. Maximum NDVI (hereafter "max NDVI") exhibited the lowest values in visible bands. The values of NIR and red edge spectral bands of date 1 images were similar to those of median method images, and those of date 2 images were similar to those of max NDVI method images. The highest and lowest values of SWIR bands corresponded to date 1 and max NDVI, respectively. NIR, red edge 3 and red edge 4 reflectance bands had the highest reflectance values and, at the same time, the highest variation among temporal aggregation methods. Finally, the comparison of the results in each spectral band reveals a low variation between sites (coefficient of variation).

Table 3

Summary statistics for spectral band information of each temporal aggregation method for 10 studied sites. First, mean values were calculated per site; then, mean values and coefficient of variation (between “()”) of the 10 sites were calculated.
Band	Temporal aggregation method
	Two-date		Median	Max NDVI
	Date 1	Date 2	Median	Max NDVI
Blue	0.11 (0.10)	0.11 (0.11)	0.11 (0.11)	0.09 (0.12)
Green	0.11 (0.11)	0.11 (0.13)	0.11 (0.12)	0.09 (0.14)
Red	0.11 (0.17)	0.10 (0.22)	0.10 (0.19)	0.07 (0.25)
NIR	0.25 (0.06)	0.30 (0.06)	0.26 (0.05)	0.30 (0.08)
Red Edge 1	0.14 (0.11)	0.13 (0.14)	0.13 (0.13)	0.11 (0.17)
Red Edge 2	0.22 (0.07)	0.26 (0.04)	0.22 (0.04)	0.25 (0.05)
Red Edge 3	0.26 (0.07)	0.31 (0.06)	0.26 (0.04)	0.31 (0.08)
Red Edge 4	0.29 (0.06)	0.34 (0.06)	0.29 (0.04)	0.34 (0.08)
Swir 1	0.26 (0.08)	0.23 (0.08)	0.23 (0.10)	0.20 (0.11)
Swir 2	0.18 (0.15)	0.15 (0.18)	0.15 (0.18)	0.11 (0.23)

3.2. Classification results after applying simple temporal aggregation methods

Accuracy assessment indexes are shown in Fig. 5. In nine of the 10 sites, the highest overall accuracy (OA) was achieved by the stacked pair of images (two-date method). Of the other two methods that aggregate several date images, median composite had the highest OA value (Fig. 5), whereas max NDVI composite had the lowest value for all sites. In general, kappa index values were closely related to the OA, i.e. datasets with a high OA also obtained high kappa values.

Although the percentage of land cover types differed between sites (Table 4), OA rates above 95% were achieved in eight of 10 sites, ranging from 96.2–98.2%. Surprisingly, for site 6, the median method achieved 94.4% OA, an almost similar value to that achieved by the two-date method (94.2%). The lowest OA was achieved in Site 1, with 87.4%, 84.6%, and 81.7% for the two-date, median and max NDVI methods, respectively. For the 10 sites, OA had on average 3% of error at 95% confidence level.

Table 4

Mean area per patch and land cover percentage per thematic class for the three methods of temporal aggregation applied at 10 studied sites.
	Site 1		Site 2		Site 3		Site 4		Site 5		Site 6		Site 7		Site 8		Site 9		Site 10
Landcover	MA	LP	AM	LP	AM	LP	AM	LP	AM	LP	AM	LP	AM	LP	AM	LP	AM	LP	AM	LP
Method = Two-date
Forest	0.51	20	0.45	11	0.35	10	0.28	14	0.07	5	0.33	15	0.19	8	0.21	12	0.23	7	1.03	31
Bare soil	0.16	23	0.60	14	0.08	0	0.08	10	0.51	35	0.21	12	0.27	15	0.33	20	0.34	7	0.08	7
Builded	0.06	11	0.12	12	0.06	7	0.06	6	0.03	4	0.04	1	0.04	2	0.05	1	0.13	5	0.19	2
Crop	0.03	5	0.82	50	2.35	69	0.11	16	0.05	3	0.32	29	0.93	44	0.66	44	0.8 1	43	0.62	40
Pasture	0.45	42	0.29	13	0.22	14	1.01	54	0.97	53	0.49	43	0.40	31	0.42	23	2.18	39	0.24	20
Method = Median
Forest	0,40	19	0,20	6	0,28	10	0,34	18	0,10	6	0,33	12	0,28	9	0,10	4	0,36	7	0,90	30
Bare soil	0,13	21	0,27	5	0,47	4	0,09	11	0,34	30	0,18	13	0,20	13	0,24	15	0,23	8	0,09	9
Builded	0,04	7	0,10	11	0,09	9	0,06	6	0,04	7	0,05	1	0,04	2	0,05	1	0,09	5	0,13	2
Crop	0,05	15	1,24	65	1,36	66	0,20	25	0,04	6	0,19	23	0,81	57	2,01	60	2,50	60	0,60	42
Pasture	0,31	38	0,15	14	0,10	10	0,37	40	0,90	52	0,68	52	0,18	19	0,52	20	0,23	20	0,19	18
Method = Max NDVI
Forest	0,20	19	0,11	4	0,16	10	0,23	17	0,05	4	0,19	13	0,13	14	0,12	10	0,17	9	0,46	30
Bare soil	0,13	23	0,16	7	0,04	2	0,09	11	0,75	41	0,09	8	0,14	12	0,54	30	0,18	4	0,06	7
Builded	0,03	10	0,07	13	0,04	5	0,04	6	0,02	2	0,03	1	0,04	2	0,04	1	0,09	8	0,13	2
Crop	0,03	7	0,66	60	1,73	72	0,05	15	0,04	3	0,19	32	0,41	49	0,68	42	0,66	48	0,42	40
Pasture	0,26	41	0,12	16	0,08	11	0,88	52	0,60	50	0,37	45	0,16	23	0,12	17	0,42	31	0,17	21
MA: Mean area per patch (hectares); LP: Landcover percentage (%).

Regarding producers’ and users’ accuracies (PA & UA) of land cover maps obtained from the different temporal aggregation methods in the 10 sites (Table 5), in most cases the two-date method yielded the best results, followed by median and max NDVI methods.

Regarding results per class, the forest class achieved the highest accuracy values (mean = 98%), followed by the crop class (mean = 96.5%). For pasture, bare soil and built area classes, accuracy values were above 90%, except for site 1, where the highest values for these classes ranged between 78% and 88%. Bare soil and built areas achieved the lowest mean values (93% and 94%, respectively). The low values for these classes are mainly due to substantial misclassification of bare soil as built area class. The greatest difference among methods corresponded to bare soil and crop classes, with the two-date method performing much better than max NDVI method (about 5.7% and 6.2% difference for bare soil and crop classes, respectively, obtained for PA, and 7.6% and 5.8% difference, respectively, for UA) (Table 5). The two-date method classified the forest class with the highest accuracy; this class shows the lowest difference from the results obtained with max NDVI method, of 2.7% and 2% for PA and UA, respectively.

Table 5

Producer’s and user's accuracy for 10 studied sites and aggregation method. The highest values are underlined.
Temporal Aggregation Method†	Producer's accuracy					User's accuracy					Site
Temporal Aggregation Method†	Forest	Bare Soil	Built area	Crop	Pasture	Forest	Bare Soil	Built area	Crop	Pasture	Site
Two-date	0.97	0.78	0.81	0.95	0.87	0.97	0.79	0.88	0.93	0.80	1
Median	0.96	0.74	0.79	0.91	0.84	0.97	0.81	0.89	0.78	0.80
Max_NDVI	0.95	0.70	0.77	0.89	0.79	0.95	0.72	0.82	0.84	0.75
Two-date	0.99	0.98	0.93	0.96	0.97	0.99	0.97	0.96	0.95	0.96	2
Median	0.98	0.96	0.93	0.96	0.92	0.98	0.96	0.95	0.89	0.95
Max_NDVI	0.97	0.92	0.76	0.90	0.95	0.98	0.84	0.88	0.87	0.92
Two-date	0.99	0.99	0.93	0.97	0.96	0.99	0.98	0.95	0.96	0.96	3
Median	0.98	0.98	0.91	0.95	0.93	0.99	0.96	0.95	0.90	0.94
Max_NDVI	0.96	0.95	0.82	0.90	0.93	0.97	0.90	0.92	0.86	0.91
Two-date	0.99	0.97	0.92	0.96	0.98	0.99	0.96	0.99	0.96	0.91	4
Median	0.99	0.98	0.94	0.94	0.93	0.99	0.97	1.00	0.92	0.89
Max_NDVI	0.97	0.96	0.92	0.86	0.95	0.99	0.92	0.99	0.89	0.87
Two-date	0.97	0.99	0.98	0.99	0.99	1.00	0.98	0.99	0.96	0.97	5
Median	0.97	0.98	0.98	0.97	0.97	0.99	0.98	0.98	0.96	0.94
Max_NDVI	0.95	0.98	0.97	0.95	0.97	0.99	0.97	0.99	0.97	0.91
Two-date	0.99	0.91	0.93	0.95	0.95	0.97	0.91	0.98	0.90	0.95	6
Median	1.00	0.92	0.91	0.99	0.92	0.97	0.90	0.98	0.95	0.93
Max_NDVI	0.94	0.90	0.89	0.94	0.92	0.97	0.88	0.97	0.86	0.91
Two-date	0.97	0.96	0.96	0.98	0.95	0.97	0.93	0.99	0.97	0.96	7
Median	0.98	0.95	0.95	0.99	0.95	0.97	0.95	0.98	0.95	0.96
Max_NDVI	0.95	0.89	0.90	0.89	0.94	0.94	0.85	0.92	0.92	0.93
Two-date	0.99	0.98	0.95	0.99	0.97	0.99	0.93	0.99	0.97	0.98	8
Median	0.98	0.92	0.95	0.99	0.96	0.99	0.92	0.97	0.93	0.97
Max_NDVI	0.98	0.91	0.91	0.95	0.95	0.98	0.86	0.97	0.96	0.93
Two-date	1.00	0.98	0.97	0.99	0.98	1.00	0.97	0.98	1.00	0.98	9
Median	0.99	0.98	0.92	1.00	0.97	0.99	0.93	0.98	0.98	0.97
Max_NDVI	0.99	0.93	0.93	0.98	0.98	0.99	0.94	0.95	0.97	0.96
Two-date	1.00	0.95	0.89	1.00	1.00	1.00	0.89	0.96	1.00	0.98	10
Median	0.99	0.91	0.87	0.98	0.97	0.99	0.87	0.97	0.95	0.94
Max_NDVI	0.96	0.86	0.83	0.93	0.92	0.97	0.80	0.92	0.94	0.88

3.3. Landscape metrics

The land cover percentage of each mapped class differed between sites, but was similar across temporal aggregation methods (Table 4). Sites 2, 3, and 7–10 were dominated by agricultural land class; sites 1, and 4–6 were dominated by grassland class. In general, the median composite dataset overestimated crop areas and underestimated grass and bare soil areas. Interestingly, the max NDVI and two-date methods reached similar land cover proportion values for all classes, although these datasets achieved the lowest and highest accuracy, respectively. Patch density (PD) was also diverse between sites. This metric (Table 2) is a simple measure of the extent of fragmentation of the classified cover, which illustrates the diffuse urban sprawling and the fragmentation of peri-urban areas. Sites 1 and 9 presented the highest and lowest PD values, respectively. Max NDVI was the aggregation method with highest PD value (except for site 5). As number of patches, PD often has limited interpretative value per se, because it conveys no information about the size, shape, or spatial distribution of patches. Therefore, it is important to analyse PD in relation to other metrics.

Patch shape complexity can be evaluated using the perimeter-area ratio (PARA). The results obtained from each aggregation method showed that, considering all classes together, max NDVI had the highest PARA. Thus, according to PD and PARA values, thematic maps based on the max NDVI method had the most complex shape.

Patch mean radius of gyration (GM) values were calculated as a measure of patch extent. GM is a measure of the distance from each cell to the patch centroid and is based on cell centre-to-cell centre distances. This metric is affected by both patch size and patch compaction. Its value is zero when the patch consists of a single cell and increases as the patch increases in extent (McGarigal 2015). GM evidenced that max NDVI temporal aggregation method-based maps are composed of many compact patches in comparison to two-date and median methods.

The values of the landscape level metrics are reported in Fig. 7. The SHDI is equal to 0 when only one patch is present and increases, without limit, as the number of classes increases while the proportions are equally distributed. There was no evidence of a relationship between SHDI values and the three temporal aggregation methods. CONTAG is a metric that approaches zero when the classes are disaggregated to the maximum (i.e., each cell belongs to a different class) and interspersed. CONTAG is equal to 100 when all the classes are aggregated at most, i.e., when the landscape consists of individual patches. In general, the thematic maps based on the max NDVI composited dataset had the lowest CONTAG values (Fig. 7).

Finally, an entropy metric was evaluated. Entropy is employed to measure the heterogeneity of observations (Altieri, Cocchi, and Roli 2018). Conditional entropy (CONDENT) is a metric of the configurational complexity of a pattern. A small value means that cells of a particular class are predominantly adjacent to only one cell class, whereas a high value of CONDENT indicates that cells (of any class) are adjacent to cells of several different classes (Nowosad and Stepinski 2019). An increase in CONDENT values was observed from the two-date method to the max NDVI temporal aggregation method (Fig. 7).

Figure 8 shows for site 10 a detail of the land cover map created using Max NDVI method together with that created using Two-date method. It can be seen from the maps in Figure 8 the differences in landscape heterogeneity, as well as its discrepancies with the validation data.

Pixel-based temporal aggregation is a paradigm that applies a suite of user-defined rules to reduce redundant information and generate cloud-free, radiometrically and phenologically consistent, spatially contiguous image composites (White et al. 2014). From the S2 archive, three images were generated, one that corresponds to the median spectral response, one that captures the peak value of NDVI in the seasonal cycle of vegetation, and a two-date composite.

A crucial point of land cover mapping applications using multitemporal satellite images is the radiometric consistency of the acquired images (Syariz et al. 2019). The degree of influence of clouds, sun angle, water-vapour, aerosols and directional surface reflectance is minimized in maximum NDVI composite imagery with increasing length of the quantizing interval (Holben 1986). Specifically, in order to apply pixel-based composite techniques, the availability of cloud-free images in the study period is mandatory (Griffiths et al. 2013). Accuracy in thematic map production is a function of the quantity and quality of imagery available; data deficiency during the peak of the growing season is a severe limitation (Johnson 2019). In this study, the cloud coverage values per date and the number of cloud-free images per pixel over the growing season were analysed.

For the data assessed in this study and the four-month period considered, at least one cloud-free image is ensured every 13 days. This value is high, considering that the study area belongs to a monsoon regime with a rainy season concentrated in the growing season (Aliaga, Ferrelli, and Piccolo 2017). In areas of Australia characterised by a low precipitation regime, the period between two cloud-free S2 acquisitions was on average less than two weeks (Sudmanns et al. 2020). S2 data have higher frequency than Landsat 8, the alternative satellite for use in this type of studies; this is because the study area is located in the centre of the Landsat 8 orbit, so scenes do not overlap and the maximum revisit frequency is 16 days. In practice, this frequency cannot be met because of cloud coverage (Frantz et al. 2016), which is high especially when considering that the analyzed growing season coincides with the rainy season.

The inconsistency in the number of cloud-free pixels available over a period is one of the major challenges of using image temporal aggregation techniques. In this study, we found that for a four-month period that includes the summer crop season, at least 25 cloud-free images are ensured. The values observed are far better than those observed by Carrasco et al. (2019), who found a minimum of seven cloud-free images per pixel for a one-year interval of S2 data in the United Kingdom. In our study, such high number of free-cloud images per pixel allowed us to generate composites. Composites are helpful for characterizing vegetation phenology, and have shown to improve classification accuracies (Azzari and Lobell 2017; Bey et al. 2020; Johnson 2019). How to generate a cloud-free input dataset for analysing land use/cover change in GEE is an important and crucial step, and the effects of the selected method on the classification results is generally underestimated (Phan, Kuch, and Lehnert 2020).

Here, the two-date method yielded the highest accuracy indexes, followed by the median method, with the max NDVI method having the poorest results. These findings are consistent with those of Carrasco et al. (2019) and Phan, Kuch, and Lehnert (2020), who also found that traditional multi date datasets obtained the highest classification accuracy compared with median and max NDVI methods applied to Sentinel 2 data.

Here, patch size and homogeneity of forest and crop classes were detected as possible characteristics that may increase classification accuracy. In central Córdoba, forests are characterized by a net decrease of species richness due to grazing pressures and fire recurrence (Aguilar et al. 2018), whereas crop fields are characterized by large-scale commercial mechanized farms. The simplicity of the forest and the machinery used for crop sowing, planting and harvesting increases homogeneity of these types of land cover patches. This homogeneity and the large mean area (MA) size of forest patches and field plots (Table 4) contribute to a clearer differentiation of these classes using these methods (Bey et al. 2020). Using multi-date images in classification enhances the spectral separability of the land cover classes and increases the overall accuracy of the land cover maps (Rujoiu-Mare et al. 2017). On the other hand, difficulties in distinguishing the built-up area from the bare land patches may be explained by their similar spectral response.

The use of the median method was found to not only significantly reduce data volume but also remove noisy, very dark, and very bright pixels that may occur due to shadow and haze (Amani et al. 2019; Azzari and Lobell, 2017; Okoro et al. 2016; Phan et al. 2020). These factors may explain the relatively good performance of land cover maps based on the median method with respect to max NDVI. Even though the max NDVI method is useful to indicate the highest level of vegetation abundance (Huang et al. 2017), it did not perform well as a temporal aggregation method. Depending on the study context, max NDVI may be useful for identifying specific land use changes, such as conversion of cropped land to conservation, non-irrigated to irrigated land, or forest land variation (Maxwell and Sylvester 2012).

Landscapes with high values of PD and PARA, and low GM can generally be considered spatially heterogeneous (Awuah et al. 2018; Silva et al. 2018; Statuto, Cilis, and Picuno 2018; Tran, Julian, and De Beurs 2014). Here, the land cover maps with the highest landscape heterogeneity were obtained using the max NDVI method. Bey et al. (2020) also found that Seasonal Greenest Pixel composites included more hazy and shaded pixels than the Median composite. These authors suggested that the inclusion of hazy pixels may be indicative of a more systematic bias, since hazy pixels often had higher reflectance values than clear pixels in the bands required for NDVI calculation. Ruefenacht (2016) and Xie et al. (2019) also found that the max NDVI composite has Landsat anomalies, clouds, shadows, and abnormal pixelation. Therefore, it is likely that max NDVI is not a convenient method for constructing a cloud-free composite from time series.

Landscape metrics integrate all patch classes over the full land extent and represent the spatial complexity of the pattern of the entire land cover map (McGarigal 2015). Land cover map with either high or low CONDENT values can generally be considered a land cover map with a complex pattern (Altieri, Cocchi, and Roli 2018; Awuah et al. 2018; Nowosad and Stepinski 2019). Again, our results showed that the max NDVI method was the one revealing the most complex pattern. The comparison among the 10 sites showed that site 1 had the land cover map with the highest complexity, regardless of the temporal aggregation method used to generate the composite database. This result was also observed in PD values, with site 1 showing the highest fragmentation. Site 1 is characterized by an urbanized natural area, with a very low proportion of impervious surfaces compared with urban environments. Awuah et al. (2018) also found the highest values of spatial heterogeneity in the transition zone between urban and rural areas. In those cases, the authors suggest the use of finer resolution images that can retain more spatial and thematic details in land cover.

In the current study, landscape metrics values were found to be significantly related to accuracy indexes, which impacts map quality. Maps exhibiting high levels of landscape fragmentation-heterogeneity and complexity were closely correlated with low accuracy assessment values. This observation is supported by the hypothesis that highly heterogeneous land cover classes generally have lower accuracy with respect to simple and homogeneous land cover patches, not only because of the possibility of geolocation error, but also because these subsets include all edge pixels that are generally more difficult to classify correctly than centre patch pixels (Glick et al. 2016; Stehman and Foody 2019). Awuah et al. (2018) also found that high fragmentation can result in an increase in the number of mixed pixels, affecting classification accuracy. This fact can be observed in our results, since the max NDVI method was related to high heterogeneity and low overall accuracy.

It is known that uncertainty of landscape metrics is negatively related to classification accuracy. In most landscape metrics, the values tend to stabilize as the overall accuracy of image classification approaches 90%. This means that a high degree of classification accuracy is required for ensuring the consistency and reliability of landscape metrics (Shao and Wu 2008). The two-date method showed the best performance in accuracy indexes and, at the same time, it exhibited the lowest value of fragmentation and complexity in land cover maps. For small areas, its implementation is also computationally and technically less demanding than that of the median and max NDVI methods, since GEE has a built-in command that facilitates the procedure.

An interesting future research avenue is the use of this methodology in a wider area and including several crop growing seasons; we will focus on the precise discrimination of principal crops to determine the predominant crop sequences in the area, because proper sequencing of crops is an important component of successful dynamic cropping systems (Ashworth et al. 2017; Krupinsky et al. 2006). Moreover, the current open access to historical satellite data sets and free GIS tools give us the opportunity to identify relationships between the main dynamics in land use, land use policies and socioeconomic contexts through techniques and methodologies that can be applied around the globe.

LM plays a crucial role in environmental research, providing insights into the ecological and functional dynamics of ecosystems. However, in dynamic systems, generating thematic maps for LM analysis poses challenges due to the substantial data volume required and issues such as cloud cover interruptions. In this study, we applied three temporal aggregation methods over Sentinel-2 image time series to create composites that served as data input for land cover mapping. The analysis was applied in 10 sites with different land use proportions, agricultural crops, and landscape relief.

Our findings confirm that temporal resolution provided by S2 data offered a very useful cloud-free collection, i.e. a combination of high coverage and low average cloud cover. A minimum of 25 cloud-free pixels were ensured for the four-month period analysed. Two-date and median methods outperformed maximum NDVI method in all 10 sites, with overall accuracies ranging between 87.4% and 98.3%. LM were related to the classification performance, as the classification accuracy is sensitive to the spatial heterogeneity and complexity. Thematic maps based on max NDVI method showed the highest values of fragmentation-heterogeneity and complexity. These findings demonstrate that high levels of land cover map fragmentation-heterogeneity and complexity are closely correlated with low overall accuracy. Although the two-date method showed the best performance in accuracy indexes and, at the same time, exhibited a low value of landscape fragmentation and complexity, creating a two-date composite may have practical limitations in large areas, since the orbital swath width is 290 km in S2 images.

Given the good classification accuracy obtained without post-classification processing and the simplicity of the temporal aggregation method, the median method could be directly integrated into operational programs used for land mapping and monitoring based on satellite image time series.

Acknowlegdements

The authors thank Phd. Juia Asteringo y Martin Videla for their assistance in data recording and scientific advice, and the Facultad de Ciencias Agropecuarias (UNC) for providing research facilities.

Author contributions Miguel Nolasco: Conceptualization, Investigation, Validation/Visualization, Methodology, Writing–original draft. Mónica Balzarini: Investigation, Formal analysis, Review and editing.

Funding

This research was partially supported by CONICET through the doctoral scholarship assigned to the first author, and by Agencia Nacional de Promoción de la Investigación, el Desarrollo Tecnológico y la Innovación (FONCYT PICT 2017 1071).

Data availability

The datasets used and/or analyzed during the cur-rent study are available from the corresponding author on reasonable request.

Competing interests

The authors declare that they have no competing interests.

Ethical approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Abdollahi S, Ildoromi A, Salmanmahini A, Fakheran S, Kulczyk S (2023). Quantifying the relationship between landscape patterns and ecosystem services along the Urban–rural gradient. Landsc Ecol Eng 19(3), 531-547. https://doi.org/10.1007/s11355-023-00561-x
Aguilar R, Calviño A, Ashworth L, Aguirre-Acosta N, Carbone LM, Albrieu-Llinás G, Nolasco M, Ghilardi A, Cagnolo L (2018) Unprecedented Plant Species Loss After a Decade in Fragmented Subtropical Chaco Serrano Forests. PloS one 13(11), e0206738. https://doi.org/10.1371/journal.pone.0206738
Aliaga VS, Ferrelli F, Piccolo MC (2017) Regionalization of Climate Over the Argentine Pampas. Int J Climatol 37, 1237–1247. https://doi.org/10.1002/joc.5079
Altieri L, Cocchi D, Roli G (2018) A New Approach to Spatial Entropy Measures. Environ Ecol Stat 25(1), 95-110. https://doi.org/10.1007/s10651-017-0383-1
Amani M, Mahdavi S, Afshar M, Brisco B, Huang W, Mirzadeh SMJ, White L, Banks S, Montgomery J, Hopkinson C (2019) Canadian Wetland Inventory Using Google Earth Engine: The First Map and Preliminary Results. Remote Sens 11(7), 1–20. https://doi.org/10.3390/RS11070842
Ashworth AJ, Allen FL, Saxton AM, Tyler DD (2017) Impact of Crop Rotations and Soil Amendments on Long‐Term No‐Tilled Soybean Yield. Agron J 109(3), 938-946. https://doi.org/10.2134/agronj2016.04.0224
Awuah KT, Nölke N, Freudenberg M, Diwakara BN, Tewari VP, Kleinn C (2018) Spatial Resolution and Landscape Structure Along an Urban-Rural Gradient: Do They Relate to Remote Sensing Classification Accuracy? – A Case Study in the Megacity of Bengaluru, India. Remote Sens Appl: Soc Environ 12, 89–98. https://doi.org/10.1016/j.rsase.2018.10.003
Azzari G, Lobell DB (2017) Landsat-Based Classification in the Cloud: An Opportunity for a Paradigm Shift in Land Cover Monitoring. Remote Sens Environ 202, 64–74. https://doi.org/10.1016/j.rse.2017.05.025
Belgiu M, Drăguţ L (2016) Random Forest in Remote Sensing: A Review of Applications and Future Directions. ISPRS J Photogramm Remote Sens 114, 24-31. https://doi.org/10.1016/j.isprsjprs.2016.01.011
Bey A, Jetimane J, Lisboa SN, Ribeiro N, Sitoe A, Meyfroidt P (2020) Mapping Smallholder and Large-scale Cropland Dynamics with a Flexible Classification System and Pixel-based Composites in an Emerging Frontier of Mozambique. Remote Sens Environ 239, 111611. https://doi.org/10.1016/j.rse.2019.111611
Breiman L (2001) Random Forests. Mach Learn 45(1), 5–32. https://doi.org/10.1023/A:1010933404324
Carrasco L, O’Neil AW, Daniel Morton R, Rowland CS (2019) Evaluating Combinations of Temporally Aggregated Sentinel-1, Sentinel-2 and Landsat 8 for Land Cover Mapping with Google Earth Engine. Remote Sens 11(3), 288. https://doi.org/10.3390/rs11030288
Drusch M, Del Bello U, Carlier S, Colin O, Fernandez V, Gascon F, Hoersch B, et al. (2012) Sentinel-2: ESA's Optical High-Resolution Mission for GMES Operational Services. Remote Sens Environ 120, 25-36. https://doi.org/10.1016/j.rse.2011.11.026
Fan C, Myint SW (2014) A comparison of spatial autocorrelation indices and landscape metrics in measuring urban landscape fragmentation. Landscape Urban Plan 121, 117-128. https://doi.org/10.1016/j.landurbplan.2013.10.002
Fontana FMA, Trishchenko AP, Khlopenkov KV, Luo Y, Wunderle S (2009) Impact of Orthorectification and Spatial Sampling on Maximum NDVI Composite Data in Mountain Regions. Remote Sens Environ 113(12), 2701–2712. https://doi.org/10.1016/j.rse.2009.08.008
Frantz D, Röder A, Stellmes M, Hill J (2016) An Operational Radiometric Landsat Preprocessing Framework for Large-Area Time Series Applications. IEEE Trans Geosci Remote Sens 54(7), 3928-3943. https://doi.org/10.1109/TGRS.2016.2530856
Glick HB, Routh D, Bettigole C, Seegmiller L, Kuhn C, Oliver CD (2016) Modeling the Effects of Horizontal Positional Error on Classification Accuracy Statistics. Photogramm Eng Remote Sens 82(10), 789-802. https://doi.org/10.14358/PERS.82.10.789
Gorelick N, Hancher M, Dixon M, Ilyushchenko S, Thau D, Moore R (2017) Google Earth Engine: Planetary-Scale Geospatial Analysis for Everyone. Remote Sens Environ 202, 18–27. https://doi.org/10.1016/j.rse.2017.06.031
Griffiths P, Nendel C, Hostert P (2019) Intra-Annual Reflectance Composites from Sentinel-2 and Landsat for National-Scale Crop and Land Cover Mapping. Remote Sens Environ 220, 135–151. https://doi.org/10.1016/j.rse.2018.10.031
Griffiths P, van der Linden S, Kuemmerle T, Hostert P (2013) A Pixel-Based Landsat Compositing Algorithm for Large Area Land Cover Mapping. IEEE J Sel Top Appl Earth Obs Remote Sens 6(5), 2088-2101. https://doi.org/10.1109/JSTARS.2012.2228167
Hargis CD, Bissonette JA, David JL (1998) The behavior of landscape metrics commonly used in the study of habitat fragmentation. Landscape Ecol 13(3), 167-186. https://doi.org/10.1023/A:1007965018633
Hesselbarth M.H., Sciaini M., With K.A., Wiegand K., Nowosad J. (2019) Landscapemetrics: An Open‐Source R tool to Calculate Landscape Metrics. Ecography 42(10), 1648-1657. https://doi.org/10.1111/ecog.04617
Holben BN (1986) Characteristics of Maximum-Value Composite Images from Temporal AVHRR Data. Int J Remote Sens 7(11), 1417–1434. https://doi.org/10.1080/01431168608948945
Huang H, Chen Y, Clinton N, Wang J, Wang X, Liu C, Gong P, Yang J, Bai Y, Zheng Y, Zhu Z (2017) Mapping Major Land Cover Dynamics in Beijing Using All Landsat Images in Google Earth Engine. Remote Sens Environ 202, 166–176. https://doi.org/10.1016/j.rse.2017.02.021
Izquierdo-Verdiguier E, Zurita-Milla R (2020) An Evaluation of Guided Regularized Random Forest for Classification and Regression Tasks in Remote Sensing. Int J Appl Earth Obs Geoinformation 88, 102051. https://doi.org/10.1016/j.jag.2020.102051
Johnson DM (2019) Using the Landsat Archive to Map Crop Cover History Across the United States. Remote Sens Environ 232, 111286. https://doi.org/10.1016/j.rse.2019.111286
Krupinsky JM, Tanaka DL, Merrill SD, Liebig MA, Hanson JD (2006) Crop Sequence Effects of 10 Crops in the Northern Great Plains. Agric Syst 88(2-3), 227-254. https://doi.org/10.1016/j.agsy.2005.03.011
Luo X, Tong X, Qian Z, Pan H, Liu S (2019) Detecting Urban Ecological Land-cover Structure Using Remotely Sensed Imagery: A Multi-area Study Focusing on Metropolitan Inner Cities. Int J Appl Earth Obs Geoinformation 75, 106–117. https://doi.org/10.1016/j.jag.2018.10.014
Lyons MB, Keith DA, Phinn SR, Mason TJ, Elith J (2018) A Comparison of Resampling Methods for Remote Sensing Classification and Accuracy Assessment. Remote Sens Environ 208, 145–153. https://doi.org/10.1016/j.rse.2018.02.026
Maxwell SK, Sylvester KM (2012) Identification of “Ever-Cropped” Land (1984-2010) Using Landsat Annual Maximum NDVI Image Composites: Southwestern Kansas Case Study. Remote Sens Environ 121, 186–195. https://doi.org/10.1016/j.rse.2012.01.022
McGarigal K (2015) FRAGSTATS help. University of Massachusetts: Amherst. Accessed 20 September 2020. https://www.umass.edu/landeco/research/fragstats/documents/fragstats.help.4.2.pdf
Nolasco M, Bocco M, Scavuzzo CM (2018) Mapeo de Superficies Agrícolas en Base a Datos Espectrales y de Referencia a Campo no Contemporáneos. 2018 IEEE Biennial Congress of Argentina (ARGENCON) 1-7. https://doi.org/10.1109/ARGENCON.2018.8646305
Nowosad J, Stepinski TF (2019) Information Theory as a Consistent Framework for Quantification and Classification of Landscape Patterns. Landscape Ecol 34(9), 2091–2101. https://doi.org/10.1007/s10980-019-00830-x
Okoro SU, Schickhoff U, Böhner J, Schneider UA (2016) A Novel Approach in Monitoring Land-Cover Change in the Tropics: Oil Palm Cultivation in the Niger Delta, Nigeria. J Geogr Soc Berlin 147(1), 40–52. https://www.die-erde.org/index.php/die-erde/article/view/254
Orynbaikyzy A, Gessner U, Conrad C (2019) Crop Type Classification Using a Combination of Optical and Radar Remote Sensing Data: a Review. Int J Remote Sens 40(17), 6553-6595. https://doi.org/10.1080/01431161.2019.1569791
Phan TN, Kuch V, Lehnert LW (2020) Land Cover Classification Using Google Earth Engine and Random Forest Classifier—The Role of Image Composition. Remote Sens 12(15), 2411. https://doi.org/10.3390/rs12152411
R Development Core Team (2023) R: A language and environment for statistical computing.
Rodriguez-Galiano VF, Ghimire B, Rogan J, Chica-Olmo M, Rigol-Sanchez JP (2012) An Assessment of the Effectiveness of a Random Forest Classifier for Land-Cover Classification. ISPRS J Photogramm Remote Sens 67, 93-104. https://doi.org/10.1016/j.isprsjprs.2011.11.002
Ruefenacht B (2016) Comparison of Three Landsat TM Compositing Methods: A Case Study Using Modeled Tree Canopy Cover. Photogramm Eng Remote Sens 82(3), 199–211. https://doi.org/10.14358/PERS.82.3.199
Rufin P, Frantz D, Ernst S, Rabe A, Griffiths P, özdoğan M, Hostert P (2019) Mapping Cropping Practices on a National Scale Using Intra-Annual Landsat Time Series Binning. Remote Sens 11(3), 232. https://doi.org/10.3390/rs11030232
Rujoiu-Mare MR, Olariu B, Mihai BA, Nistor C, Săvulescu I (2017) Land Cover Classification in Romanian Carpathians and Subcarpathians Using Multi-Date Sentinel-2 Remote Sensing Imagery. Eur J Remote Sens 50(1), 496–508. https://doi.org/10.1080/22797254.2017.1365570
Shao G, Wu J (2008) On the Accuracy of Landscape Pattern Analysis Using Remote Sensing Data. Landscape Ecol 23(5), 505-511. https://doi.org/10.1007/s10980-008-9215-x
Silva CHL, Aragão LEOC, Fonseca MG, Almeida CT, Vedovato LB, Anderson LO (2018) Deforestation-Induced Fragmentation Increases Forest Fire Occurrence in Central Brazilian Amazonia. Forests 9(6), 305. https://doi.org/10.3390/f9060305
Statuto D, Cillis G, Picuno P (2018) GIS-based Analysis of Temporal Evolution of Rural Landscape: A Case Study in Southern Italy. Nat Resour Res 28, 61–75. https://doi.org/10.1007/s11053-018-9402-7
Stehman SV, Foody GM (2019) Key Issues in Rigorous Accuracy Assessment of Land Cover Products. Remote Sens Environ 231, 111199. https://doi.org/10.1016/j.rse.2019.05.018
Sudmanns M, Tiede D, Augustin H, Lang S (2020) Assessing Global Sentinel-2 Coverage Dynamics and Data Availability for Operational Earth Observation (EO) Applications Using the EO-Compass. Int J Digital Earth 13(7), 768-784. https://doi.org/10.1080/17538947.2019.1572799
Syariz MA, Lin BY, Denaro LG, Jaelani LM, Van Nguyen M, Lin CH (2019) Spectral-consistent relative radiometric normalization for multitemporal Landsat 8 imagery. ISPRS J Photogramm Remote Sens 147, 56-64. https://doi.org/10.1016/j.isprsjprs.2018.11.007
Tran TV, Julian JP, De Beurs KM (2014) Land Cover Heterogeneity Effects on Sub-pixel and Per-pixel Classifications. ISPRS Int J Geo-Information 3(2), 540–553. https://doi.org/10.3390/ijgi3020540
Wang S, Azzari G, Lobell DB (2019) Crop Type Mapping Without Field-Level Labels: Random Forest Transfer and Unsupervised Clustering Techniques. Remote Sens Environ 222, 303–317. https://doi.org/10.1016/j.rse.2018.12.026
White JC, Wulder MA, Hobart GW, Luther JE, Hermosilla T, Griffiths P, Coops NC, et al. (2014) Pixel-Based Image Compositing for Large-Area Dense Time Series Applications and Science. Can J Remote Sens 40(3), 192-212. https://doi.org/10.1080/07038992.2014.945827
Xie S, Liu L, Zhang X, Yang J, Chen X, Gao Y (2019) Automatic Land-Cover Mapping Using Landsat Time-Series Data Based on Google Earth Engine. Remote Sens 11(24), 3023. https://doi.org/10.3390/rs11243023
Zhou G, Wang H, Chen W, Zhang G, Luo Q, Jia B (2020) Impacts of Urban land surface temperature on tract landscape pattern, physical and social variables. Int J Remote Sens 41(2), 683-703. https://doi.org/10.1080/01431161.2019.1646939

No competing interests reported.

Download PDF

Reviewers agreed at journal
31 Jul, 2024
Reviewers invited by journal
31 Jul, 2024
Editor assigned by journal
23 Jul, 2024
Submission checks completed at journal
23 Jul, 2024
First submitted to journal
11 Jul, 2024

You are reading this latest preprint version

Assessment of Temporal Aggregation of Sentinel-2 Images on Seasonal Land Cover Mapping and Its Impact on Landscape Metrics

Status:

Version 1

Abstract

Figures

1. Introduction

2. Material and Methods

2.1. Study area and sites

2.2. Data collection, pre-processing, and temporal aggregation

2.3. Field reference data

2.5. Land cover classification and assessment

3. Results

3.1. Temporal aggregation

3.2. Classification results after applying simple temporal aggregation methods

3.3. Landscape metrics

4. Discussion

5. Conclusions

Declarations

References

Additional Declarations

Status:

Version 1