Creation and environmental applications of 15-year daily inundation and vegetation maps for Siberia by integrating satellite and meteorological datasets

Abstract

As a result of climate change, the pan-Arctic region has seen greater temperature increases than other geographical regions on the Earth’s surface. This has led to substantial changes in terrestrial ecosystems and the hydrological cycle, which have affected the distribution of vegetation and the patterns of water flow and accumulation. Various remote sensing techniques, including optical and microwave satellite observations, are useful for monitoring these terrestrial water and vegetation dynamics. In the present study, satellite and reanalysis datasets were used to produce water and vegetation maps with a high temporal resolution (daily) and moderate spatial resolution (500 m) at a continental scale over Siberia in the period 2003–2017. The multiple data sources were integrated by pixel-based machine learning (random forest), which generated a normalized difference water index (NDWI), normalized difference vegetation index (NDVI), and water fraction without any gaps, even for areas where optical data were missing (e.g., cloud cover). For the convenience of users handling the data, an aggregated product is provided, formatted using a 0.1° grid in latitude/longitude projection. When validated using the original optical images, the NDWI and NDVI images showed small systematic biases, with a root mean squared error of approximately 0.1 over the study area. The product was used for both time-series trend analysis of the indices from 2003 to 2017 and phenological feature extraction based on seasonal NDVI patterns. The former analysis was used to identify areas where the NDVI is decreasing and the NDWI is increasing, and hotspots where the NDWI at lakesides and coastal regions is decreasing. The latter analysis, which employed double-sigmoid fitting to assess changes in five phenological parameters (i.e., start and end of spring and fall, and peak NDVI values) at two larch forest sites, highlighted a tendency for recent lengthening of the growing period. Further applications, including model integration and contribution to land cover mapping, will be developed in the future.

1 Introduction

The pan-Arctic region, which includes Siberia, is known for its extremely cold climate. However, it has recently been experiencing a higher rate of temperature increase than other regions of the Earth’s surface, mainly as a result of global climate change (Hantemirov et al. 2022, Arctic Monitoring and Assessment Programme (AMAP) 2021). The higher temperatures increase the risk of climatic extremes and disasters, such as heatwave-induced wildfires (Witze 2020), thawing permafrost that promotes the release of greenhouse emissions from these areas (Yokohata et al. 2020), and river ice jams and consequent flooding (Madaeni et al. 2020; Sakai et al. 2015).

In order to evaluate the impact of climate change on the pan-Arctic region, detailed monitoring of specific environmental parameters is essential. This has been accomplished using remote sensing techniques. Water-related parameters (Suzuki and Matsuo 2019) such as total water storage, snow water equivalent (SWE), soil moisture, and surface water coverage (Velicogna et al. 2012; Suzuki et al. 2020; Yang et al. 2007; Bartsch et al. 2009; Watts et al. 2012; Mizuochi et al. 2021a) have been used to investigate local and global hydrological cycles. Another focus is vegetation parameters (Nagai et al. 2019) such as plant functional type (i.e., land cover) map, aboveground biomass, leaf area index, and growing season duration (Leroy et al. 2006; Myneni et al. 2001; Kushida et al. 2007; Buitenwerf et al. 2015), which are important for ecological monitoring and carbon cycle research. Water and vegetation parameters are closely interlinked via energy budgets and meteorological and ecological schemes in relation to evapotranspiration, albedo, vegetation growth, canopy interception of precipitation, and surface aerodynamic properties (e.g., Mizuochi et al. 2021b).

Optical remote sensing has been traditionally used in such environmental studies to monitor changes in surface water and vegetation. Band indices, such as the (modified) normalized difference water index (NDWI: Xu 2006) and the normalized difference vegetation index (NDVI: Hatfield et al. 1984; Perry and Lautenschlager 1984), are widely used to characterize spectral reflectance features and enhance the appearance of water and vegetation bodies in satellite imagery, respectively.

Another promising approach is microwave remote sensing using either active radar or passive radiometers, which has the advantage of being less sensitive to cloud cover than optical sensors. Backscatter signals observed by synthetic aperture radar (SAR) can be used to extract open water and forest areas and to estimate surface soil moisture content (Twele et al. 2016; Zakharov et al. 2020; Reiche et al. 2018). Microwave radiometers also can estimate the surface water fraction, soil moisture content and vegetation optical depth (Fily et al. 2003; Owe et al. 2008; Moesinger et al. 2020).

When using satellite data to retrieve physical parameters, the divergent characteristics of the data and the technical difficulty in data handling often result in trade-offs between the spatial and temporal resolutions of the resulting product. For example, previous broad-scale water maps (e.g., open water and wetland distribution, water fraction within pixels) have resolution ranging from tens to hundreds of meters (Lehner and Döll 2004; Fluet-Chouinard et al. et al. 2015; Yamazaki et al. 2015; Pekel et al. 2016). While they offer detailed spatial information, they often provide no temporal information, or if such information is provided, the time intervals are long. Conversely, maps with high temporal resolutions (i.e., daily to monthly) typically provide relatively coarse spatial information (e.g., several tens of kilometers) (Papa et al. 2010; Schroeder et al. 2015).

To overcome such trade-offs, researchers have used multiple datasets in a synergistic manner (i.e., data fusion; Belgiu and Stein 2019). There has been particular interest in combining image data obtained by optical and microwave sensors (Suzuki and Matsuo 2019; Beamish et al. 2020). This is because information that is missing from optical sensor observations due to cloud cover or insufficient solar illumination can be compensated for using data obtained by microwave sensors. Many applications of such data fusion approaches have been reported for wetland mapping (Chasmer et al. 2020; Zhang et al. 2021; Mizuochi et al. 2021a), snow and thawing detection (Armstrong and Brodzik 2001; Kim et al. 2012) and vegetation and ecosystem monitoring (Kimball et al. 2009; Mavrovic et al. 2023). These studies involved intercomparison of different data sources and/or complementary inputs for process-based models or machine learning (ML). However, there have been few studies on simultaneous mapping of areas of water and vegetation that are interlinked. In addition, the creation and handling of broad-scale datasets that combine both high temporal resolution (e.g., daily) and moderate spatial resolution (e.g., sub-kilometer scale) are challenging, although there has been one report on the creation of daily maps of soil moisture content with a spatial resolution of 1 km (Zhu et al. 2023).

The goal of the present study was to produce simultaneous continental-scale inundation and vegetation maps for Siberia. By integrating optical and microwave satellite data with other ancillary data, maps were obtained with a daily temporal resolution and a moderate spatial resolution of 500 m. In this paper, the map generation process and the characteristics and utility of the created products are described in detail. It is expected that this data will provide fundamental information that can be used to analyze the impacts of climate change on the risk of disasters, and also to clarify changes in water and carbon cycles in the pan-Arctic region. As example applications, we performed a time-series analysis of the NDWI and NDVI on a continental scale and extracted phenological features at two experimental sites.

2 Methods

2.1 Target area and product specifications

The target is a wide area of Siberia, approximately corresponding to 50–70° N and 31–180° E at maximum (Fig. 1). Considering potential applications (i.e., large-scale monitoring and model integration), computational resources and the spatial coverage of permafrost over Siberia (Shestakova et al. 2021), the specifications of our product are shown in Table 1.

Fig. 1

MODIS tiles used in this study. The tiles (h20–h25, v02–v03) were superimposed on a sinusoidal map projection of the study area

Table 1 Specifications of developed product

To meet these requirements, we relied mainly on the Moderate Resolution Imaging Spectroradiometer (MODIS) to identify vegetation and water distributions at optimal spatiotemporal resolutions. In this way, the spatiotemporal resolution of our target product is consistent with those of MODIS: the original spatial resolution was 500 m, and the temporal resolution was daily. The product duration (2003–2017) was determined after considering the temporal characteristics of all of the data sources (see Sect. 2.2.1).

In polar regions in particular, commonly used geographic projections (i.e., latitude/longitude) exaggerate distances along the longitude direction, so distorting the shape of objects on the map and increasing pixel-based computational costs. Consequently, in the present study, data fusion processing was conducted using a MODIS sinusoidal projection (Fig. 1), which is originally provided at a resolution of 500 m. For convenience, an aggregated product using a latitude/longitude projection with a 0.1° grid resolution is also provided as a postprocessing output (see Sect. 2.2.5).

2.2 Algorithm and data sources

The basic algorithm used for map generation was previously developed by us for use in a limited study area (Mizuochi et al. 2021a). In the ML-based algorithm used in this study, a pixel-based random forest (RF; Breiman 2001) approach is used to fill observation gaps in MODIS data (explained variable) by referring to other matching data sources (explanatory variable). Subsequently, bias correction is applied using a conditional generative adversarial network (pix2pix; Mirza and Osindero 2014; Isola et al. 2017). This study expands the previous study area to a pan-Arctic scale and includes an update to the NDVI map, the details of which are summarized in Table 1. To conserve computational resources, no pix2pix bias correction was used, as the procedure only improved the RMSE accuracy by ~ 1% (Mizuochi et al. 2021a).

The overall procedure involves (1) data download, (2) preprocessing including extraction of feature variables, map translation and coregistration, (3) ML training and prediction, and (4) postprocessing including validation, visualization, creation of the water fraction maps, and product format conversion.

Freely available optical and microwave satellite data, together with reanalysis data, were downloaded from the relevant websites (see the ‘Availability of data and materials’ section). Preprocessing and postprocessing treatments that include geospatial analysis were implemented mainly using GRASS GIS (ver. 7.8.2) and QGIS (3.10.4). ML processing was implemented using Python 3.8.10 with PyCaret 2.3.10 for preliminary model selection and hyperparameter tuning, scikit-learn 1.2.1 was used for ML training and prediction, and ray 2.2.0 was used for multiprocessing. Detailed explanations of each step are provided in the following subsections.

2.2.1 Data preparation

The data sources and physical parameters were similar to those used by Mizuochi et al. (2021a); however, vegetation optical depth (VOD) data (Moesinger et al. 2020) were newly added since they may be useful for NDVI prediction. The data sources and retrieved physical parameters are summarized in Table 2.

Table 2 Data source and physical parameters for data fusion

Surface reflectance data obtained by the MODIS instrument on the Aqua satellite (MYD09GA, collection 6.1) were used to calculate the NDWI and NDVI as follows:

$$\begin{array}{*{20}c} {{\text{NDWI}} = \frac{{{\text{G}} - {\text{SWIR}}}}{{{\text{G}} + {\text{SWIR}}}}} \\ \end{array}$$

(1)

$$\begin{array}{*{20}c} {{\text{NDVI}} = \frac{{{\text{NIR}} - {\text{R}}}}{{{\text{NIR}} + {\text{R}}}}} \\ \end{array}$$

(2)

where G, R, NIR, SWIR are the green, red, near infrared, and short-wave infrared surface reflectance, which correspond to MODIS band 4 (545–565 nm), band 1 (620–670 nm), band 2 (841–876 nm), and band 7 (2105–2155 nm), respectively.

Cloud, cloud shadow, snow, and cloud-adjacent pixels were screened out (i.e., treated as missing data to be filled by ML) by referring to accompanying quality assurance data (“state_1km” subset with bitmask that indicates cloud and snow states). Also, pixels recorded during shadow time in the high-latitude winter were also treated as missing data to be filled using ML. The quality controlled NDWI and NDVI were the explained variables to be predicted from the following low-resolution explanatory variables.

Level-3 brightness temperature data, obtained by the Advanced Microwave Scanning Radiometer-EOS (AMSR-E) on the Aqua satellite and AMSR-2 on the Global Change Observation Mission-Water (GCOM-W) satellite, were used to calculate the polarization index (e.g., Moradizadeh and Srivastava 2021; Sawada et al. 2017) and surface water fraction. These indices are sensitive to soil moisture and/or abundance of surface water. The calculation method employed for the indices was the same as that in Mizuochi et al. (2021a). In addition, the SWE product retrieved from the AMSR series was used to determine snow masks. To maintain consistency with the AMSR series (on the Aqua satellite or a satellite with a similar overpass time), the other MODIS product on the Terra satellite platform (MOD09GA) was not used.

Sixteen meteorological parameters provided by ERA5-land (ECMWF 2023) along with the VOD product (Moesinger et al. 2020) estimated from multiple microwave satellite data sources (i.e., SSM/I, TMI, AMSR-E, AMSR2 and WindSat), were also used. These were selected because they were likely to be sensitive to the water and vegetation conditions on the surface. In addition, day of year (DOY) information was used to represent the season. To ensure continuity between the end of one year and the beginning of the next, DOY was calculated in cosine and sine forms as follows:

$$\begin{array}{*{20}c} {{\text{DOY}}_{\cos } = \cos \left( {2\pi \times \frac{{{\text{DOY}}}}{365.25}} \right) \;{\text{and}}\;{\text{ DOY}}_{\sin } = \sin \left( {2\pi \times \frac{{{\text{DOY}}}}{365.25}} \right)} \\ \end{array}$$

(3)

In total, 23 low-resolution maps (derived from AMSR series, ERA5-Land, VOD and DOY) were assumed to be the common explanatory variables to predict the two explained variables (MODIS NDWI and NDVI), since the NDWI (water) and NDVI (vegetation) are likely to be closely interlinked.

2.2.2 Preprocessing

Preprocessing consisted of calculating the abovementioned variables, converting map projections, coregistering maps, and performing climatology calculations. All of the processes were managed on individual MODIS sinusoidal tiles, which encompassed 12 tiles in total (vertical tile numbers 02–03 and horizontal tile numbers 20–25). First, the MODIS HDF files were converted into raw binary images containing 2400 × 2400 pixels using a sinusoidal projection with the HDF-EOS-to-GeoTIFF-conversion tool (HDF-EOS Tools and Information Center 2023). Using GIS software, data from other data sources (i.e., explanatory variables) were processed to align them with the MODIS sinusoidal image. This involved resampling, cropping (i.e., cutting out the region of interest), and converting different data formats and on different projections into raw binary images with matching locations and dimensions (i.e., coregistration). Importantly, the process employed nearest-neighbor resampling, which involves oversampling (i.e., dividing a coarse pixel into multiple pixels while retaining the same pixel value). This step was necessary as the other data sources have lower resolution than MODIS.

After coregistration, daily climatology maps were prepared for the explanatory variables by averaging the maps for the same DOY over the period 2003–2017. These daily climatology maps were assumed to represent typical values of the variables on a given DOY and were used to fill any missing pixels in the explanatory maps for ML prediction (see Sect. 2.2.3).

2.2.3 ML processing

In our data fusion scheme, pixel-based ML was used to train matching pairs of explanatory and explained variables, allowing the prediction of explained maps without any gaps (Mizuochi et al. 2018; 2021a). Before initiating ML processing, a preliminary study was conducted to select the ML model and to tune the hyperparameters using a limited number of sampling pixels (100 pixels) from several tiles. In total, 25 ML algorithms (different types of regression models) were compared based on tenfold cross validation using PyCaret. RF was chosen as a robust regressor, with the following hyperparameters: maximum tree depth 5, minimum impurity decrease 0.005, minimum samples at leaf node 2, minimum samples for internal node splitting 7 and number of trees 60.

Strictly speaking, the rank of the models fluctuated depending on the tile, the presence of absence of snow, the NDWI or NDVI (detailed below), and the referred accuracy criteria. However, ensemble models of decision trees (RF and extremely randomized trees) usually showed top scores. For simplicity, we decided to use RF for all the tiles, since it is widely used in remote sensing (Zhao et al. 2022) and is also consistent with our previous research (Mizuochi et al. 2021a).

The spectral features of snow are characterized by a continuous decrease in reflectance from the visible to the infrared region (Petty 2006), a pattern that resembles that for water bodies in NDWI and NDVI maps. To consider the sensitivity of the snow pixels to the NDWI and NDVI, RF models with and without snow were created separately by checking if the SWE retrieved by AMSR-E/AMSR2 was positive or zero. Four original images taken in May, August, November, and February of a randomly picked year during the study period were selected for validation of each season. Each variable was then normalized by subtracting the spatiotemporal mean from the original values and dividing by the spatiotemporal standard deviation for the RF training.

For each pixel, historical match-up pairs were searched during the period 2003–2017, where all the explained and explanatory variables existed altogether. In other words, one matching pair containing a full set of 23 explanatory and 2 explained variables was used as one training sample for each the pixel.

SWE and ERA5-Land included permanently missing pixels for water and ocean masks. To avoid wasting information for the other variables and to make a wall-to-wall product, ML was performed on these pixels by filling zero for SWE and the spatially averaged values of climatology maps for ERA5-Land.

After RF training, RF prediction was performed for each pixel: explanatory maps were again input to the trained RF model to predict the explained maps (i.e., MODIS NDWI and NDVI) without gaps. To this end, it was ensured that the explanatory maps at this stage had no missing values. Temporary gaps (much fewer than in the explained maps but still non-negligible) were filled by daily climatology maps. It should be noted that this treatment was applied only during the prediction process, not during the training process, to avoid overfitting to the average map. Ancillary flag maps indicating the absence/presence of this treatment (0: original explanatory variables, 1: explanatory variables filled by climatology data) and maps showing the presence of snow were also produced.

Table 3 summarizes the tile-averaged rate of missing pixels filled by climatology data in each explanatory map, and the number of available explanatory variables. AMSR-related parameters (no. 3, 4) were approximately 60% available, and ERA-related parameters (no. 5 to 21) were 100% available, both of which were insensitive to the tile location. VOD (no. 22) was sensitive to the tile location, fluctuating in the range 66–90%. In all cases, 21–22 of a total of 23 parameters were available in the spatiotemporal average for all tiles. 18 parameters were obtained even in the worst case, which is understandable because ERA5-Land and DOY are at least available at any time. Therefore, if some variables were missed and filled by climatology data, the remaining variables are expected to offer realistic input data to some extent.

Table 3 Rate of missing pixels in each explanatory variable (parameter numbers 3–22) and number of available explanatory variables

2.2.4 Postprocessing: creating water maps by NDWI thresholding

Inundation pixels were extracted as physically intuitive information by thresholding the NDWI created by ML. Through visual interpretation of NDWI maps and high-resolution satellite imagery from Google Earth, ground references were established for three typical land cover categories (barren land, vegetation and water). Three hundred samples for each category were then used to produce category-dependent NDWI histograms (Fig. 2). A Gaussian distribution was fitted to each histogram, and the intersection between the Gaussian curve representing water and those representing the other categories was extracted as the tentative NDWI threshold for the data (i.e., the maximum likelihood approach).

Fig. 2

Scheme of NDWI thresholding. (Top) Visual interpretation of three land cover categories (red: barren ground, green: vegetation, and blue: water); (bottom left) Summary table of extracted thresholds for different seasons; (bottom right) Example of maximum likelihood approach using data obtained in August 2017

The tentative thresholds between vegetation and water, derived from 12 (snow-free) monthly averaged maps for a particular tile (h24v02), showed a limited fluctuation range (Table 4). Consequently, the average value (NDWI =  − 0.043) with a standard error of 0.003 (N = 12) was assumed as the fixed threshold for all seasons. Utilizing this threshold, binary maps were generated to show 0 (non-water) and 1 (water).

Table 4 Summary of estimated NDWI thresholds in different seasons

To assess the uncertainty in the water map, a sensitivity analysis of the threshold setting was conducted. In the average map derived from the 12 monthly NDWI scenes, the threshold was slightly adjusted from the average value (NDWI =  − 0.043) within one standard deviation (0.012) using Gaussian noise. This process was repeated 12 times to calculate variations in the water fraction across the entire tile (h24v02). The mean water fraction and the standard error (N = 12) were 0.032 ± 0.001, corresponding to 3% uncertainty in water fraction estimation.

This study also considered fluctuating wetlands, i.e., areas where the extent of inundation changes seasonally, rather than permanent open water bodies such as lakes, ponds, rivers and reservoirs, which often attract special interest in estimating methane emissions, a key process affecting climate change (Zhang et al. 2021). Thus, the mode value for the summer season was subtracted from the original water maps to create what we refer to as “floating water maps”, which likely represent areas that are intermittently inundated, and which exclude permanent open water.

2.2.5 Postprocessing: creation of analysis-ready data (ARD)

In addition to the raw binary maps based on MODIS tiles, which cover the wall-to-wall NDVI, NDWI, water, and quality flags at the original spatial resolution of 500 m, a merged product was also created using a commonly used latitude/longitude projection. The product was resampled using the nearest neighbor method and stored in GeoTiff format. To reduce the cost of data handling on a continental scale, the original resolution was aggregated into a 0.1° pixel spacing. Within these pixels, the average NDVI and NDWI values were stored, along with the fractional coverage of both permanent and intermittent wet areas.

2.3 Validation

The ML result was validated for each tile over four seasons (Table 5), by comparing the generated MODIS NDVI and NDWI images with the original images that were not used in the ML training process (i.e., by employing a hold-out approach). Spatial patterns and statistical criteria were evaluated, with the latter including the mean error, root mean squared error (RMSE), correlation coefficient ®, slope and offset of the regression line between the original and created images. These values were summarized for each MODIS tile and image acquisition season.

Table 5 Acquisition date of original MODIS images used for validation in each tile

Spatiotemporal patterns were also checked for the 0.1° ARD product. In addition to visualizing the NDVI and NDWI, spatially averaged time series and temporally averaged spatial patterns of the original and floating water maps were compared with a previous wetland map (WAD2M; Zhang et al. 2021). WAD2M primarily draws upon water fraction maps derived from multiple microwave satellite data sources (SWAMPS; Jensen and McDonald 2019), supplemented with ancillary satellite data and static maps for snow and permanent water masking, plus bias correction. The recent product, which emerged from the integration of multiple data sources through different schemes and at varying spatiotemporal resolutions (0.25°, monthly), seems useful for assessing the advantages and potential disadvantages of our product.

The terrestrial water storage (TWS) anomaly is also useful for time-series comparisons with surface water (Suzuki et al. 2018). Gravity Recovery and Climate Experiment (GRACE) monthly water mass datasets were prepared (Swenson 2012; Landerer and Swenson 2012), with contributions from the Center for Space Research (CSR), GeoForschungsZentrum Potsdam (GFZ), and NASA Jet Propulsion Laboratory (JPL), culminating in the calculation of the ensemble mean of the three datasets (Sakumura et al. 2014). In addition, as in Suzuki et al. (2018), reanalysis data from Global Land Data Assimilation System (GLDAS) version 2 (Li et al. 2019) were also compared for the other TWS data source. The TWS includes snowpack, river runoff and groundwater and is not directly comparable to the NDWI and NDVI. However, it is helpful for interpreting time-series NDWI (and derived surface water area) data in relation to the water budget.

2.4 Application: trend analysis and phenological feature extraction

The first application of the product developed in this study involved trend analysis at a continental scale. Time-series of snow-masked NDVI and NDWI maps were used to calculate the 15-year trend (i.e., Theil Sen’s regression slope) for each 0.1° pixel. The statistical significance of the trends was calculated using the Mann–Kendall test. The results were compared with previous reports for 15-year trends in vegetation and water in polar regions.

The second application involved the extraction of phenological features at two study sites covered by larch forest (Spasskaya Pad: 62°15′ 17’’N, 129°37′ 10’’E; Elgeeii: 60°01′ 01’’ N, 133°49′ 53’’ E), utilizing the high observation frequency of our product (Fig. 3); detailed descriptions of each study site can be found in Nagano et al. (2022). As a useful approach for phenological monitoring in terrestrial ecosystems, a moving average (window size = 5 days) of snow-masked NDVI time series data was fitted using a double sigmoid function (e.g., Ide and Oguma 2013; Myers et al. 2019; Yan et al. 2019) as follows:

$$\begin{array}{*{20}c} {{\text{NDVI}}\left( t \right) = {\text{NDVI}}_{{\text{b}}} + \frac{1}{2}{\text{NDVI}}_{{\text{a}}} \left[ {\tanh \left( {p\left( {t - D_{i} } \right)} \right) - \tanh \left( {q\left( {t - D_{d} } \right)} \right)} \right] } \\ \end{array} ,$$

(4)

where NDVI_b is the baseline value, NDVI_a is the amplitude, D_i is the day of year when the NDVI increases most rapidly, D_d is the day of year when the NDVI decreases most rapidly, and p and q determine the rate of increase and decrease of the NDVI, respectively. The initial parameter values were set to NDVI_b = 0.4, NDVI_a = 0.2, D_i = 150, D_d = 250, and p = q = 0.001, based on a preliminary investigation of interannual phenology patterns over the sites. These parameters were then precisely fitted to the actual NDVI time-series data during the growing season (assumed to be from DOY = 100 to DOY = 280), employing the curve_fit function of the Python SciPy library.

Fig. 3

Examples of double sigmoid fitting for seasonal NDVI. The NDVI with a 5-day moving average, sigmoid fitting and second derivative plots for sites at Spasskaya Pad (left) and Elgeeii (right) during 2017 are shown

Then, local maximum and minimum values of the second derivative in the fitted function were identified, allowing for the determination of four commonly used phenological features (i.e., D₁, D₂, D₃, and D₄), which correspond to the start of spring (SOS), end of spring (EOS), start of fall (SOF), and end of fall (EOF), respectively. The date of the peak NDVI (D₅) was also extracted, and the 15-year trends in these five parameters were investigated. A summary of the individual definitions of these phenological features is given in Table 6.

Table 6 Definition of five phenological parameters

3 Results and discussion

3.1 Created data and validation

Data fusion using RF produced wall-to-wall MODIS-like NDWI and NDVI maps for the period 2003–2017 with enhanced the spatiotemporal extent compared to the existing data fusion product (Mizuochi et al. 2021a).

Examples of comparisons between the original and newly created maps in a tile are shown in Figs. 4 and 5. The examples show that the cloud gaps present in the original images were completely filled in the images created by RF regression, closely resembling the original spatial patterns and values of the NDWI and NDVI. The scatterplots show that images in the fall season deviated from the 1:1 (i.e., unity) lines, and that spatially discontinuous patterns were observed in the southwest area.

Fig. 4

Comparison between original and created MODIS NDVI maps for tile h23v03. Spring image: 2007/05/03, summer image: 2003/08/01, fall image: 2015/11/21, winter image: 2010/02/26. Color scale of the scatterplot was configured by kernel density estimation (KDE)

Fig. 5

Same as Fig. 4, but for MODIS NDWI

The discrepancy in the scatterplot analysis was quantified by Theil-Sen’s regression slope for each tile and season (Table 7). While the regression slope was close to unity for both the NDWI and NDVI in summer, large deviations representing uncertainties in the values were observed in the seasons affected by snow or soil freezing. Since the RF models were trained separately for snow and no-snow pixels, the mixture of these pixels within a scene resulted in discontinuous spatial patterns and poor regression among multiple clusters in the scatterplots (Fig. 6, h22v03). The other reason for the poor regression may have been that the limited variable range of the NDWI and NDVI in the snow-covered winter season creates a featureless spectral pattern over the entire scene (Fig. 6, h23v02). Different tiles likely had different snow or soil freezing event distributions, resulting in varying accuracies among tiles (Table 7).

Table 7 Accuracy of RF prediction averaged for each tile, for each season with standard deviation

Fig. 6

Snow-affected images. Images showing regressions adversely affected by a mixture of snow and no-snow areas (h22v03), and areas with fewer spectral features due to snow cover (h23v02)

Table 7 also shows the small mean error for all tiles (-0.01 ± 0.04 for NDWI, and 0.02 ± 0.02 for NDVI), indicating minimal biases in the reproduced images. The RMSE for all tiles was approximately 0.1 for both the NDWI and NDVI.

Seasonal changes were clearly observed in the snow-masked NDWI and NDVI time-series data, averaged over the entire study region (Fig. 7A). The NDVI oscillated between being high in the growing season and low in the winter season, while the NDWI was high in the winter season and low in the growing season. A high local NDWI peak was observed in the spring season, possibly due to spring soil thawing and/or snow melting which rapidly supplied water to rivers and wetlands. Given the high ratio of null pixels in the winter season (~ 1.0), the low NDVI and the high NDWI observed at this time were attributed to the limited number of pixels affected by snow, even after the application of snow masking, which produced spectral patterns that differed from the original land cover. In addition, the low solar incidence angle in the winter season could reduce the signal-to-noise ratio for the measured radiance as well as the reliability of atmospheric correction for the optical data, potentially resulting in poor quality of the original MODIS products. As a result, studies on vegetation and surface water using our ARD product should focus on spring through fall, with special consideration given to snow masking.

Fig. 7

Time series of 0.1° ARD product spatially averaged over the entire Siberian region. A Snow-masked NDVI and NDWI time-series with ratio of null cells averaged over entire study region. B Comparison between original (blue line) and floating (red line) water fraction extracted from our product, WAD2M wetland maps (Zhang et al. 2021), and time-series anomaly of terrestrial water storage (TWS) volumes (mm) from GRACE monthly water mass dataset (Swenson 2012; Landerer and Swenson 2012) and Global Land Data Assimilation System (GLDAS) version 2 (Li et al. 2019). For comparison with our floating water fraction, the minimum value of the WAD2M time-series data was subtracted from the original WAD2M to adjust the baseline to zero

Gap-filling by the climatology maps (i.e., an average value for the same date over 15 years) can potentially lead to bias in the prediction. This is particularly the case when predicting recent climatic extremes away from the 15-year average. In fact, slight differences in patterns around the beginning of 2012 were likely caused by a gap in the microwave data (AMSR-E terminated on 2011/10/03 and AMSR2 started on 2012/07/02), which was filled by the daily climatology maps for ML prediction.

Time-series data for the original and floating water fraction (Fig. 7B) showed consistent seasonal variability, with high values recorded in the summer season (~ 0.07) and the spring local peak, and low values recorded in the winter season. The baseline for the original water fraction was adjusted to zero in the floating water fraction. Comparison with a previous wetland map (i.e., WAD2M; Zhang et al. 2021) showed similar seasonal changes during the spring to fall seasons, although large discrepancies were observed in the winter season. The water fraction obtained in the present study for the winter season was lower than that for WAD2M, in both the original and floating cases. This difference was likely due to differences in the definitions used for snow masking. Specifically, our product strictly masked pixels where the microwave SWE was positive. This screened almost all of the pixels in the winter season and resulted in the water fraction being nearly zero. Conversely, the water fraction data from SWAMPS version 3.2 (Jensen and McDonald 2019), mainly used in WAD2M, applied a more relaxed approach to snow masking, which resulted in a substantial water fraction even in the winter season.

In the spring to fall seasons, the original WAD2M values were close to the lower limit of our original water fraction. In contrast, those for the offset-corrected WAD2M (adjusted to make the minimum values zero) were close to the upper limit of our floating water fraction. The WAD2M did not sense the strong peaks in the water fraction associated with soil thawing or snow melting in spring, probably due to its lower observation frequency (monthly) and coarser spatial resolution (0.25°) compared to our product (daily, 0.1° in the ARD product and 500 m in the original product). Furthermore, differences in the variable range of the water fraction between the ARD product and WAD2M also seemed to be linked to differences in their data sources. Unlike WAD2M, our optical sensor-based output observes the tree canopy in densely vegetated areas, but cannot explicitly delineate inundation under the vegetation. However, our product does implicitly utilize sources such as microwave data (i.e., AMSR series) and other data sources (i.e., ERA5-Land, VOD, DOY) that may be sensitive to inundation under the vegetation to some extent.

The interpretations above are also supported by TWS from GRACE and GLDAS, both of which account for snowpacks and show large values in winter. Snowmelt from snowpacks increases river runoff in the spring (Suzuki et al. 2018), which is consistent with the peak observed in our water fraction data. The lowest TWS during the summer–fall period also likely reflected our water fraction, as seen in the local minimum in the summer season.

Figure 8 shows temporally averaged NDWI and NDVI maps, which are useful for visualizing the overall distribution of surface water and vegetation (Fig. 8A, B). The original water fraction map largely contained permanent water bodies, such as rivers, lakes, and coastal areas (Fig. 8C), which were mostly excluded in the floating water fraction map (Fig. 8D). The difference between the original and floating water maps exposes the intermittent wetlands. Comparisons with WAD2M showed a substantial underestimation of the water fraction on the west Siberian plain (Fig. 9E, F). High-resolution satellite images (Google Earth) confirmed the existence of abundant small-scale wetlands, which may not be detected by our moderate-resolution product (i.e., 500 m). Different effects of topography and vegetation cover on optical and microwave data may also cause discrepancies between WAD2M and our product. However, apart from this area, the exclusion of permanent water bodies from the original water map improved the consistency of the water fraction data between our product and WAD2M (Fig. 8F).

Fig. 8

Temporally averaged maps during period 2003–2017. A NDWI, B NDVI, C original water fraction, D floating water fraction, and difference between wetland map (WAD2M) provided by Zhang et al. (2021) and E original water fraction and F floating water fraction in this study at monthly intervals. The maps visualized on the Arctic Polar Stereographic projection (EPSG: 3995) are convenient for comparison with previous results for long-term terrestrial trends in the polar region

Fig. 9

Maps showing time-series regression. Regression slope maps for snow-masked NDVI, NDWI, and water fraction provided by Zhang et al. (WAD2M 2021) are shown. Only statistically significant (p < 0.05) pixels are shown in color for the NDVI and NDWI slope maps. The map projection is similar to that in Fig. 8

3.2 Application: trend analysis and phenological feature extraction

Our wall-to-wall products enable us to perform time-series trend analyses for vegetation and surface water (Fig. 9). Specifically, the NDVI slope map showed an overall decreasing trend especially in the southwest area, with sporadic hotspots in western to central Siberian wetlands and water bodies such as Lake Baikal. Several studies have reported an overall greening trend (i.e., increasing NDVI) in the Arctic region in recent decades, which they have attributed to increased temperatures and a longer growing season (May et al. 2020). However, there have been also reports of browning (i.e., decreasing NDVI) (Myers-Smith et al. 2020), with the distribution varying depending on the data sources and methodologies used. The browning patterns (i.e., negative NDVI values) shown in Fig. 9 are closely aligned with the spatial distribution of the annual maximum NDVI trend (Fig. 1 in Myers-Smith et al. 2020) based on the GIMMS3g AVHRR dataset (2000–2015) (Tucker et al. 2005; Pinzon and Tucker 2014), which supports the validity of our NDVI product. Extensive greening is expected over the northernmost Arctic Tundra region, but this could not be investigated as this region is not covered by our product. Consequently, expanding the map coverage area will be undertaken in future research. Expanding the study duration with similar data sources with a longer period is also important future work to monitor recent climate extremes after 2017 and to track the historical record before 2003.

The NDWI slope map showed an overall increasing trend across the study area. Given that the NDWI and NDVI are based on the inverse differences between visible and infrared reflectance, their overall contrasting spatial trends are expected. Exceptions were found in small hotspots around Lake Baikal, the lake around the Zeya Nature Reserve, and coastal areas facing the Sea of Okhotsk, where decreasing trends were observed even in NDWI slope maps. These hotspots were also observed in the water fraction map for WAD2M, although WAD2M showed an overall decreasing trend across the study area. Rather, our results were more consistent with previous findings of significant increasing trends in the water fraction from 2003 to 2010 (Watts et al. 2012).

The dense time series (i.e., daily) nature of our product is well suited to extracting phenological parameters in the spring and the fall seasons through double-sigmoid fitting (Tables 8 and 9, Fig. 10). In both Spasskaya Pad and Elgeeii, D₁ showed a negative trend (i.e., SOS occurring earlier), whereas D₃ and D₄ showed positive trends (i.e., SOF and EOF occurring later), although only the D₃ in Spasskaya Pad was statistically significant. Therefore, the duration of the growing season at the two larch forest sites increased from 2003 to 2017.

Table 8 Phenological parameters estimated by double sigmoid fitting of NDVI over Spasskaya Pad

Table 9 Phenological parameters estimated by double sigmoid fitting of NDVI over Elgeeii

Fig. 10

Anomalous time series of phenological parameters (D₁ to D₅) at the Spasskaya Pad and Elgeeii experimental sites. The parameters are described in Table 6. The black dashed line indicates the zero reference line

Nagai et al. (2020) determined the start of the growing season (SGS) and the end of the growing season (EGS) for the Spasskaya Pad and Elgeeii sites, based on in-situ observations and a degree-day model. Although the parameters that they used differed from those used in this study, the SGS was significantly correlated with our D₁ (SOS) (r = 0.55 with p < 0.05 for Spasskaya Pad; r = 0.93 with p < 0.01 for Elgeeii). These findings show that our SOS extraction is potentially well suited for use as a proxy for SGS. Conversely, unlike SGS, EGS showed a weaker correlation with our D₄ (EOF) (r = 0.34 with p = 0.21 for Spasskaya Pad; r = 0.54 with p = 0.17 for Elgeeii).

The D₁ time series data for Spasskaya Pad seem to have changed at around 2005–2007, with the SOS occurring earlier thereafter (Fig. 10). This change was likely associated with waterlogging events and changes in the composition of the understory/overstory vegetation from 2005 to 2008 at the site (Kotani et al. 2014; 2019; Ohta et al. 2014; Nagano et al. 2022). The DOY of the Peak NDVI (D₅) showed larger interannual variability than the other parameters, and the temporal changes in D₅ were similar between Spasskaya Pad and Elgeeii. Similarly, D₁ and D₂ showed time-series similarity between Spasskaya Pad and Elgeeii, suggesting that these parameters were sensitive to large-scale meteorological factors such as temperature. Nagano et al. (2022) also observed a strong correlation between temperature anomalies and precipitation between Spasskaya Pad and Elgeeii, supporting the importance of temperature as a large-scale phenological driver.

Nagai et al. (2020) also stressed the complexity of the sensitivity of leaf senescence to air temperature, noting that various internal (hormones and timing of spring budburst) and external (precipitation, photoperiod, drought, and heat stress) factors need to be considered in order to better understand leaf senescence, as doing so will improve predictions of the EOF and/or EGS.

In this study, our original and ARD products are considered to be valuable for monitoring water and vegetation distributions, and for performing trend analysis and phenological research. Further potential applications may include contributions to land surface modeling (Guimberteau et al. 2018) as input and/or reference data. In addition, enhancing time-series land cover maps (Brown et al. 2022) to explicitly quantify temporal changes in vegetation and water distributions could also be undertaken in the future.

4 Conclusions

This research provided 15-years water and vegetation maps with daily frequency and a 500-m spatial resolution over the Siberia region on a continental scale, integrating optical and microwave satellite data, and meteorological data. A systematic treatment based on pixel-based RF techniques and further postprocessing produced maps of the NDWI, NDVI, and water fraction (original and floating), without any gaps for the period from 2003 to 2017. In addition to the original sinusoidal maps with 500-m resolution, an analysis-ready product that merges all of the sinusoidal maps on a latitude/longitude projection was also rescaled to a resolution of 0.1° pixel spacing. The treated NDWI and NDVI images showed no overall systematic biases, with a RMSE of approximately 0.1 for all tiles. Based on the products, trend analysis showed the distribution of NDVI browning and increasing NDWI hotspots for the period from 2003 to 2017. Phenological features were extracted for each year at two larch forest sites (Spasskaya Pad and Elgeeii) based on double-sigmoid fitting of NDVI time-series data. The findings revealed a recent lengthening tendency in the growing period at the sites and confirmed that waterlogging caused an earlier start to spring at the Spasskaya Pad site. Our product can potentially be applied to spatiotemporal monitoring of water and vegetation, including trend analysis, phenological research, model integration and contribution to land cover mapping.