1. Introduction
Heterogeneous growing conditions within a field plot can benefit from modern precision agriculture in terms of spatially optimized management [
1] and illustrate the high importance of acknowledging the interaction between the soil conditions, topography and cereal production [
2,
3,
4,
5,
6,
7]. For an efficient and non-destructive assessment of the heterogeneity of soil conditions of arable land and plant growth characteristics, the use of remote sensing data is becoming increasingly established [
8,
9]. Plant growth across landscapes or within a single field can show temporally consistent spatial patterns, such as low- and high-yielding areas. Depending on the spatial resolution, vegetation indices (VI) allow for assessing such patterns and their change over time. Several studies have successfully used vegetation indices based on remote sensing data to monitor and predict plant growth at the field scale [
4,
5,
10,
11,
12]. However, the direct link between within-field changes in soil properties, topography and plant growth and their change over time has seldom been studied in more detail. It therefore often remains unclear if spatial growth differences strongly vary over a longer time period and are governed by external factors such as meteorological conditions, or are mainly influenced by intrinsic factors, such as soil parameters, which would lead to temporally consistent spatial growth patterns.
This study aims at providing an example and a workflow for using high-resolution satellite time series (Rapid Eye, Planet Scope) to assess spatially consistent vegetation patterns and their relation to soil characteristics at the field scale. The test site is an arable field of about six ha located in the state of Brandenburg, Germany. Using satellite data from the years 2015 to 2020 the following research hypotheses were tested: (1) the spatial distribution of Normalized Difference Vegetation Index (NDVI) mainly depends on soil heterogeneity and is consistent over time, (2) annual weather conditions can slightly modify the spatial pattern of NDVI. Years with dry conditions lead to more pronounced spatial patterns.
2. Materials and Methods
The study area is located in Großmutz, Brandenburg, Germany, between continental (low precipitation, low humidity) and maritime (high precipitation, high humidity) climates [
13] and is classified as a moderately dry lowland climate. The long-term averages between 1991 and 2020 (WHO climate normal) of the closest climate station of the German Weather Service (DWD) in Neuruppin-Alt Rupin (Station ID: 96) are 623 mm of annual precipitation and 9.6 °C mean air temperature [
14,
15]. The field has an area of six ha and is characterized by spatial heterogeneity in terms of soil properties, topography and crop growth [
16]. The altitude decreases from the southeastern corner (58 m a.s.l.) to the northern corner (51 m a.s.l.) (
Figure 1). A soil reconnaissance map by [
17] classifies the investigated area as an anhydromorphic substrate type “ASL w” (sandy ground moraine with undulated relief). The dominant soil types can be classified as Arenic Cambisol or Haplic Luvisol [
18]. The agricultural field is conventionally managed and not irrigated. The cereal crops cultivated during the observation period are listed in
Table 1. Summer oat (harvested in 2018) was excluded from this study to facilitate interannual comparisons between winter cereals.
The meteorological data used in this study were based on station data provided by the DWD (Germany’s national meteorological service (Deutscher Wetterdienst, DWD)) [
18]. The climatic water balance (CWB in mm) was calculated for each day by subtracting the potential evaporation from the precipitation. Daily water balances were summed up to monthly values.
Soil samples were taken up to one m depth with a Pürckhauer soil auger along nine transect lines as described in [
16]. The transect orientation followed the slope from the southwestern side to the northern corner with a spatial distance of 50 m from each other. The outer transects were 25 m apart from the edge of the field. Within the transects, samples were taken in 15 m intervals and each coordinate of the 87 sampling points was recorded with a Trimble Geo 7× Handheld GNSS System (Sunnyvale, CA, USA). Following the Munsell soil color charts [
19] and the FAO (Food and Agriculture Organization of the United Nations) guidelines for soil description [
20], the physical properties of the soil profiles were estimated in moist conditions by manually wetting soil probes. In general, all soil profiles down to 1 m depth were composed of a layer with sandy soil material (predominantly Loamy Sand or Sand) with variable thickness (35 to 100 cm) and a subsequent loam layer (Sandy loam to Loam) with variable thickness (0–65 cm). Based on the spatial variability of the vertical position (depth) of the loamy layer, three soil classes were derived as an indicator for differences in the available water capacity (AWC) of the soils [
16]. Soil class 1 (S1) was characterized by a shallow upper boundary of the loamy subsoil layer (<60 cm depth), class 2 (S2), a deeper upper boundary of the loamy layer at 60–100 cm depth and class 3 (S3) was sandy throughout 100 cm soil depth (no loamy layer). Assuming that sandy soils have a low water holding capacity [
21] that is improved by the presence of an underlying loamy layer, soil class 1 is assumed to show the highest and class 3 the lowest AWC. An interpolated soil map was created by kriging the soil sample points. To take the different altitude levels into account, which describe the position of the drilling points along the slopes (catena) at our site, a freely available digital terrain model (DTM) provided by the Amt für Landesvermessung und Geobasisinformation Brandenburg (LGB,
https://geobasis-bb.de/lgb/de/geodaten/3d-produkte/gelaendemodell/, last visited: 17 October 2022) was used to further distinguish between three altitude classes: A1: <53.5 m, A2: 53.5–55.37 m, A3: >55.37 m a.s.l. Both classifications are shown in
Figure 1.
The main type of satellite imagery used in this paper originates from the RapidEye satellite constellation. RapidEye was owned by Planet (Planet Labs, Inc., San Francisco, CA, USA) and the data were provided in cooperation with ESA’s Third-Party Missions Programme. The RapidEye satellite program consisted of 5 satellites collecting data in five spectral bands: red: 630–685 nm, green: 520–590 nm, blue: 440–510 nm, red edge: 690–730 nm and near-infrared (NIR): 760–850 nm. The resolution was 5 × 5 m per grid cell on orthorectified imagery resulting from 6.5 m GSD (Ground Sampling Distance) at the nadir [
22]. Only images captured within the main growing season after winter were selected, which was between April and June of each year. Cloud coverage was visually checked. Since RapidEye’s operation was discontinued in spring 2020, PlanetScope imagery was applied as the satellite data source for the 2020 data. PlanetScope was chosen because the images were delivered by the same provider (Planet Labs, Inc., San Francisco, CA, USA), had a similar high temporal and spatial resolution compared to RapidEye and were available free of charge, including an atmospheric correction. The first 12 satellites were launched on 22 June 2016; therefore, PlanetScope was not able to cover the whole investigation period either. The images used were taken with either sensor type PS2 or PS2.SD [
22]. The sensor PS2 consists of a four-band frame imager (red: 590–670 nm, green: 500–590 nm, blue: 455–515 nm, NIR: 780–860 nm) with a split-frame for the visible RGB (red/green/blue) and the NIR. PS2.SD consisted of a four-band frame imager (red: 650–682 nm, green: 547–585 nm, blue: 464–517 nm, NIR: 846–888 nm) with butcher-block filter providing separate stripes for blue, green, red and NIR. The GSD of 3.9 m leads to a higher resolution (3 × 3 m). Both RapidEye and PlanetScope provided atmospherically corrected imagery.
To calculate the Normalized Difference Vegetation Index (NDVI) [
23,
24], the Red (R) and Near Infrared (NIR) bands were used:
Subsequently, data were limited to the scene with the highest NDVI values in each month (maximum value approach) assuming that these images were least affected by cloud and haze. Image selection was undertaken with a preference for scenes captured in the middle of the month and around the same day each month to allow a constant observation frequency and to increase the comparability across years. However, this was not always feasible due to the limited availability of cloud-free satellite images. The 14 dates used are listed in
Table 2.
Due to the described inherent different spatial and spectral resolutions between RapidEye and PlanetScope imagery, and to reduce sources of bias in the NDVI data, PlanetScope data were only used for statistical data analysis when NDVI values were normalized and divided into quartiles. By using normalized NDVI values and a quantile-based approach, we evaluated general trends and patterns and accounted for the uncertainty in NDVI values due to interannual variations in crop phenological development and crop-specific optical properties.
For all statistical analyses and graphs, the statistical software R (version 3.6.3, Windows 10 [
25]) was used. Masks were applied to the satellite scenes to select corresponding grid cells for each altitude class (
Figure 1) and to subsequently extract information on soil classes (
Figure 1) within the altitude classes (package: “raster” [
26]). NDVI values were not normally distributed, which was tested with the Shapiro–Wilk test. Therefore, the non-parametric Kruskal–Wallis test was used to test for significant differences in the NDVI values between classes for each scene. Here, an analysis of variance was used to test whether independent samples belong to the same population [
27]. As a post-hoc test, the non-parametric Wilcoxon signed-rank test was applied whenever significant differences were found to specify which groups significantly differed by testing two paired samples for their central tendencies [
28]. To avoid type 1 errors in the Wilcoxon test, the
p-value was adjusted with the Bonferroni method [
29].
To compare NDVI values across years, the NDVI values were normalized by dividing all the NDVI values of each grid cell with the mean of all grid cells. Subsequently, the relative NDVI values of each scene could be classified into four different classes. To do so, the values were split into quartiles, and each quartile was assigned its own value (1 to 4). This processing resulted in images showing the classified relative NDVI values, ranging from low (class Q1) to high (class Q4) NDVI values.
First, the non-parametric Mann–Kendall trend test [
30] was applied to test for the presence of a monotonic trend in the time series over time (R-package: “Kendall” [
31]). Tests were applied for each month separately. This test is commonly used in environmental research [
32,
33]. Subsequently, a correlation matrix was created by using the “corrplot” package [
34]. In the first step, 10% of the pixels per scene were subsampled and pairwise correlations between those pixels were calculated across the years, using the non-parametric Spearman rank correlation using the global “cor” function implemented in R. To quantify the changes in the spatial distribution and size of NDVI clusters between years, cross-tabulations (crosstabs) for all possible combinations of months across all years were computed.