Shifting cultivation, which is still an important agricultural system in tropical regions, is being rapidly replaced by large-scale intensive farming. Such transitions have a complex impact on ecosystems and societies, which are not fully understood. Availability of free Landsat imagery and the development of novel methods to analyze dense satellite time-series opens the opportunity to understand the temporal and spatial patterns of shifting cultivation in biodiversity-rich tropical regions. We utilized 456 atmospherically and terrain-corrected Landsat images, which were condensed into 28 average normalized difference moisture index (NDMI) annual composites from 1984 to 2016 to assess the dynamics of shifting cultivation in the Yurimaguas region of Peru covered by WRS 2 path 08 row 64. We used Breaks For Additive Season and Trend (BFAST) to detect structural breaks on every object of the spatially segmented NDMI stack and derived information on the change of length and frequency of shifting cultivation cycles. We also classified the breaks and remaining pixels into Shifting Cultivation, Forest and Other classes with Random Forest classifier. We assessed both thematic accuracy and timing of classified shifting cultivation cycles. Finally, we performed Morphological Spatial Pattern Analysis (MSPA) of habitat fragmentation due to the changing frequency of shifting cultivation and related to change in the habitat of endemic tropical forest species. Our study revealed that 31% of the study area was under shifting cultivation by 2016. Shifting Cultivation Class was detected with User’s Accuracy of 71% and Producer’s Accuracy of 85%. Approximately 75% of shifting cultivation fields had one to two cultivation cycles and were in close proximity to settlements and rivers. However, we also observed an overall increase in cultivation cycles, which were correlated with the population increase in the study area (R² = 0.62). New shifting cultivation fields (opened in 2011-2016) were located far away from settlements and resulted in additional fragmentation of the forest frontier. The analysis of habitat fragmentation revealed a habitat loss for endemic tropical forest species, such as interior forest birds and butterflies, as a result of the increasing shifting cultivation intensity. In sum, the BFAST method adapted to track the temporal breaks for the spatial segments was robust to classify shifting cultivation fields and monitor the change of shifting cultivation cycles in tropical areas. It can be extended to elsewhere to track the structural breaks in ecosystem functioning and land use.

Area A - Board 28

Vegetation phenology has acquired special interest in recent decades. The behaviour of the biological cycles of plants is closely related to global change processes, being considered a key biological indicator of climate change. Shifts in the spatiotemporal patterns of phenology can influence the functioning of ecosystems (the water and carbon cycles or energy flows), and also have important impacts in agriculture, forestry, human health and even in the global economy. Land surface phenology (LSP), the study of the timing of occurrence of biological cycles of plants using time-series of vegetation indices (VI) derived from satellite data, can improve the monitoring of the response of vegetation to environmental changes. The development of satellite remote sensing technology and its ability to provide temporally continuous information about the response of ecosystem to climate change at different spatial scales could explain the increase of the LSP studies in recent decades.

A systematic literature review was carried out to evaluate and summarize the LSP studies at global level. This review aimed to answer three main questions: (i) where and when has been analysed LSP?; (ii) what was the purpose for studying LSP?; and, (iii) which methodologies have been used to estimate LSP?. The literature search was performed using two electronic scientific databases (Scopus and Web of Science). The term “Phenology” was used as a keyword in the search and only peer-reviewed articles published between 1980 and 2017 were considered. This search yielded 38975 papers. Non-duplicated articles were refined using different terms and keywords such as: “remote sensing”, “land surface phenology”, “climate change”, “seasonality”; and classified by geographical distribution. For each geographical area, only LSP studies based on estimates of phenological parameters (e.g. start, end and length of the growing season) were considered. Finally, a total of 438 papers were selected to be included in this review.

This review showed an unequal spatial distribution of LSP studies at the regional level. More than 75% of the studies evaluated LSP in mid-latitudes of the Northern Hemisphere. The 80% of the studies estimated LSP from two main satellite sensors at coarse and medium spatial resolutions: AVHRR and MODIS. Landsat, RapidEye, HJ-1 A/B or Sentinel data allowed to estimate LSP at a finer spatial resolution, but only the 6% of the studies used the information derived from these satellites. The most popular VI for LSP extraction were NDVI and EVI. Prior to LSP parameters extraction, different smoothing techniques were applied to remove atmospheric disturbances and noise, being logistic function based methods and Savitzky-Golay filters the most popular. Finally, the threshold and derivative based-methods were the most used to extract phenological parameters.

The main purpose of using LSP has changed from 2010s onwards. In 1990s and 2000s, studies were based especially on the characterization (spatial and/or temporal) of the behaviour of different biological cycles of vegetation. However, in the recent years, the majority of the studies are focused in the study of the relationship between vegetation phenology and environmental drivers (mainly climatic drivers), using medium or coarse spatial resolutions.

Area A - Board 29

The primary focus of this work was to investigate an ensemble of observation networks composed of flux tower measurements, near surface remote sensing and ground phenological observations to find global representativeness for validation of satellite derived land surface phenology products. The specific networks includes, Fluxnet, phenocams, USA National Phenological Network (NPN) and Pan European Phenological Network (PEP725),

A new global scale “phenoregion” analysis was carried out to stratify the phenology of the world based on a k-means clustering approach trained from yearly Land Surface Phenology (LSP) estimates, considering the median and standard deviation of the Start of Season (SOS), End of Season (EOS) and Growing Season Length (GSL) for two different periods: 1981-1998 and 1999-2016. The yearly LSP estimates were acquired from the VIP Lab (Vegetation Index and Phenology Laboratory), which are estimated from MODIS and AVHRR sensors at a spatial resolution of 0.05 degrees (https://vip.arizona.edu/viplab_data_explorer.php#). Numerous k-means models were built using the Hartigan-Wong algorithm and considering increasing number of clusters from 10 to 100. The internal variability of clusters converged around 30 phenoregions for both periods. Analysis of the representativeness of the ensemble sites entailed comparison of distances among phenoregions (k-means centroids) in a 3-dimensional feature space (SOS, EOS. and GSL). The ESA Climate Change Initiative (http://cci.esa.int/) global land cover was used to quantify the representativeness of networks for different land covers and phenoregions.
In general, southern hemisphere has greater number of phenoregion compared to northern hemisphere. The spatial distribution of phenoregions is very different to that of land cover classes, being a phenoregion composed of multiple land covers. In general, all except 3 phenoregions has some representataion through ground observation which could be used for validation of LSP products. Three phenoregions in South America, India and Australia lacks any observational network. Therefore, any future investment in monitoring network should target those area to have a complete global representation.

Area A - Board 30

Satellite observations of land surface phenology have gained increasing importance for studying responses of vegetation to environmental and anthropogenic changes over space and time. With the launch of Sentinel-2AB, dense medium resolution time series (between 2-3 days in mid-latitudes) have become available that have the potential to improve phenological remote sensing models. Combining Sentinel-2 with Landsat data will even further shorten revisit times. Considering the variability and complexity of land surface phenology at continental scales, approaches are needed that are able to deal with variations in vegetation type and environmental factors like snow cover and cloud conditions. The choice of the modeling approach is often influenced by the biogeographic region and vegetation type. For example, non-parametric methods have the potential to produce better time series fits when data availability is good but they may not be as robust under heavy cloud conditions.
The aim of this study is to develop robust and generalizable methods to estimate vegetation green-up, applicable over a wide range of ecosystems in Europe. We use state-of-the-art multi-sensor time series from Landsat ETM+, OLI, and Sentinel-2 MSI, atmospherically corrected using the Framework for Operational Radiometric Correction for Environmental monitoring. Based on test sites selected to cover a climate and land-use gradient ranging from tundra to Mediterranean vegetation in Europe, we evaluate the sensitivity of parametric and non-parametric algorithms to variably data conditions and their generality across sites. Furthermore, we assess how to best integrate prior geographic and climate knowledge into models, via ensemble modeling and hierarchical modeling, to better constrain remote sensing-based models and to provide estimates of uncertainty. We evaluate our models against mechanistic (meteorology-based) phenology models and ground-based phenological observations. The outcomes of this study contribute to an improved understanding on the utility of high-resolution sensor data (and specifically the Copernicus Land Monitoring Service) for mapping vegetation phenology and to an improved understanding of the spatial variability of vegetation green-up in Europe.

Area A - Board 31

Information on phenology is an important indicator for the characterization of the status and dynamics of land use/cover. Thereby, monitoring land cover dynamics requires data covering a preferably long time period. This information is provided by earth observation data with a high temporal repetition rate as acquired from sensors such as MODIS and NOAA AVHRR. For studies where spatially detailed information is required, for instance in areas with a heterogeneous land cover pattern, the spatial resolution of these dense time series (250 m x 250 m up to 4 km x 4 km) is often too coarse for providing information in sufficient detail. Earth observation data with an appropriate spatial resolution and covering long time periods such as provided by Landsat are somewhat limited due to the low temporal frequency (16 days orbital repeat). Therefore, these data often lack consistent phenological information because of cloud cover in many regions of the globe. In Southern Africa for example the repeat coverage is on average 48 to 64 days in the dry season and even 90 to 100 days in the wet season.

Reflectance fusion methods such as the spatial and temporal adaptive reflectance fusion model STARFM combine the advantages of both sensor types and thus, have the capability to generate data with both high spatial resolution and high temporal information. Unfortunately, using these methods to create dense time series that cover large areas over long periods is difficult for several reasons including storage capabilities and processing time. Therefore, we introduced the fusion method ImproPhe that directly predicts phenology metrics at medium resolution by fusing coarse resolution phenology metrics with medium resolution reflectance data (see figure).

In this study, we derive a 17-year time series of fused LSP metrics like seasonal amplitudes or integrals. We explore the capabilities and limits of these data to support land cover studies in a highly heterogeneous and dynamic landscape in Southern Angola and Northern Namibia. We thereby focus on the delineation of major land use systems and compare these to the original MODIS-derived land surface phenology as well as to Landsat-derived land cover information. For evaluating the capabilities of land cover change detection of the fused dataset, we use common time series tools to derive linear trend parameters but also segmentation algorithms to detect modifications of land cover and abrupt land cover changes.

The results demonstrate that this dataset is capable to monitor LSP with a high spatial detail and thus, is a valuable asset to support land cover mapping as well as for an improved monitoring of land cover changes by employing time series of phenology metrics with Landsat spatial resolution.

Area A - Board 32

The objective of the study is to describe the methodologies used to complete or fill gaps in the Sentinel 2 (A and B) MSI time series by using Landsat 8 OLI acquisitions, knowing that sensors acquiring data at the quasi same high spatial resolution could be used. The interest of this data merging is important and aims first at providing end users with consistent time series at high spatial resolution while keeping the spectral properties of Sentinel 2.

A new method for the production of dense Landsat 8 / Sentinel 2 time series has been already presented and has demonstrated its superiority among several other methods. This method is based on the use of the Dynamic Time Warping technics in the context of multi sensor data fusion.

To manage this merging, two steps were needed. First showing that the consolidation of the sentinel 2 time series could be achieved using the DTW method, and second by showing that the densification of Sentinel 2 data could be performed using Landsat 8 data.

To validate the first step, data acquired by S2A has been downloaded over six validation sites, representing a large diversity of landscapes (forest, agricultural sites, snowy sites) with low to high cloud coverage. The accuracy of the method among three has been statistically demonstrated.

The second step requested the development of different modules to simulate S2 acquisitions from Landsat 8/OLI acquisitions. The densification chained three modules: spectral band pass adjustment, geometric re-gridding, and BRDF corrections. Hyperspectral data have been used to the spectral adjustment, and existing BRDF databases have been used to normalise the time series in a same geometry.

Area A - Board 33

The short revisit time of Sentinel-2 provides new opportunities to study seasonal vegetation dynamics. Several studies demonstrated that phenological parameters can be accurately assessed with 10-30m resolution optical data for ecosystems with a single long growing season in the year. In semi-arid rangelands, vegetation green-up and senescence occur more rapidly because rainfall is concentrated over short periods. Obtaining sufficient cloud-free acquisitions to accurately describe the short vegetation cycles can thus be challenging. This study evaluates if the combined Sentinel-2A and -2B data allow for accurately mapping vegetation phenology in a semi-arid rangeland. The study is executed in a 150-km2 experimental farm situated in Machakos County, Kenya. The average annual rainfall is about 550mm, divided equally over two rainy seasons but with a large interannual variability. Both livestock and wildlife graze on its vegetation, consisting of grasses, shrubs, and trees. Using all available Sentinel-2 imagery between October 2017 and September 2018, we derived temporal NDVI profiles and compared these for three locations with greenness chromatic coordinate (GCC) series derived from digital repeat RGB-cameras. Double hyperbolic tangent models were fitted to the time series, separately for the two rainy seasons locally referred to as short and long rains. We derived start- and end-of-season (SOS/EOS) for camera and satellite series using a 50 percent threshold between minimum and maximum levels of the modelled time series. For the two seasons considered, the SOS retrievals were within 8 days of those derived from the camera series, while the EOS retrievals showed larger deviations of up to 20 days. This difference could be explained largely by the effect of non-photosynthetic elements that are more dominant in the oblique camera view. Subsequently, spatial variability of the phenological parameters was mapped by applying the approach to all 10-m resolution grid cells within the farm boundaries. Our analysis showed that for most of the area Sentinel-2 could provide sufficient temporal detail for accurately estimating the phenology of short vegetation cycles. This is important for better understanding climate and degradation impacts on rangeland vegetation, particularly for semi-arid rangeland systems with large interannual variability in phenology and productivity.

Area A - Board 34

To monitor land productivity trends and assess land degradation in woodlands in large areas with remote sensing, it is necessary to have a long time series of indicators of vegetation dynamics with a high temporal resolution. Contrary to typical forests that are characterised by a mostly continuous perennial tree cover, woodlands in the Mediterranean Basin are composed of relatively sparse trees, notably of cork oak (Quercus suber) and holm oak (Quercus ilex), associated with shrubs and grasses. Trees and shrubs have a perennial cover whereas the grasses are composed mostly of annual plants which are only visualised by remote sensing during the productive period. Thus, different components of Mediterranean vegetation have different dynamics over time, this means different phenology. The images provided by Sentinel 2 A/B, multispectral instruments with a 5 days temporal resolution developed under the Copernicus Program, have a 10-meter spatial resolution over the visible and near-infrared range, which is a major advantage for the study of vegetation phenology. However, the satellite is only operational since 2015 and thus cannot provide an archive with enough temporal depth to evaluate trends in vegetation phenology. The main goal of this study is to use MODIS (Moderate Resolution Imaging Spectroradiometer), an instrument that provides daily data since 2000 but with a coarser spatial resolution of ~250m, to re-create an un-mixed NDVI time series for each of the underlying vegetation cover types within the mixed pixels (in this case for trees and grasses). Based on these synthetic NDVI time series, it is then possible to follow the phenology of each vegetation cover type over time. This dataset can then serve as a baseline for future studies relying on Sentinel 2.
Using MODIS in such a complex and heterogeneous landscape requires particular attention to the instrument’s observational geometry and the resulting observational support. The spatial response of this sensor can be modelled with a point spread function (PSF) adapted to MODIS, which considers the view zenith angle, latitude and longitude of the study area. MODIS has a large instantaneous field of view and a wide swath, which implies that the scan angle increases considerably from the nadir to the edge; thus, the footprint of each observation encoded in a pixel increases towards the edge of an image, while observations from neighbouring pixels increasingly overlap. Consequently, there is a mismatch between the observation support and the grid cell, i.e., the signal assigned to a certain pixel also comes from the surrounding pixels.
To unmix signals of trees and grasses from MODIS observations, recent aerial photography is used to build a binary mask of tree crown cover to identify pixels of interest. A reasonable assumption is made here that most of these trees have remained throughout the period of interest (2000-2018). Then, the purity of each MODIS pixel at each image is determined by applying a PSF convolution over the binary images. The spatial variability of different purity conditions over the landscape is then exploited to establish a relationship between purity (of either trees or grasses) and MODIS NDVI using a generalised linear model. This is done for a series of dates between 2016 to 2018 when corresponding images of Sentinel 2 are also available. These are used to calibrate and validate the relationship, also evaluating the change in performance when alternating between dry and humid seasons. The calibrated models are used to estimate a pair of a retrospective NDVI unmixed time series for “trees” and “grasses” for each pixel during the period of 2000-2018. These new time series are then used to characterise phenology metrics of cork and holm oak across different Mediterranean regions. This novel methodology to extract un-mixed signals from heterogeneous pixels offers promising potential for various other ecosystems, such as tropical savannahs and boreal taigas, where major phenological changes are expected to happen as a consequence of climate change.

Keywords
Mediterranean woodlands, MODIS, Sentinel-2A/B, Spatial Response, Pixel purity, NDVI unmixed time series, phenology

Area A - Board 35

Land surface phenology (LSP) (the timings of biological events in the life cycle of plants) and its variation in space and time is a valuable tool for assessing the impact of climate variability and anthropogenic impact on the biosphere. Over the last decade there has been a rapid development of deriving LSP from satellite sensor data, primarily at a coarse spatial resolution (>300m). In many cases this is not good enough resolution to capture the heterogeneity in the landscape and in turn the phenological process. This results in uncertainty in quantifying the landscape processes, such as carbon cycling, which uses LSP as an input.
In addition, these coarse resolution products can be difficult to validate due to the lack of homogenous elementary sampling units at the pixel resolution and most of the ground measurements representing small areas in the pixel.
In parallel to the development of LSP products from satellite imagery a series of near surface remote sensing initiatives, such as the PhenoCam network (Accessible at: https://phenocam.sr.unh.edu/webcam/), have been recently developed. These initiatives aim to provide real time, high spatial and temporal resolution data for monitoring phenology across a range of land cover types. These in situ measurements are vital for the validation of satellite derived products and have create a permanent database that can be analysed both quantitatively and qualitatively. The introduction of cameras which are sensitive to both visible and near infrared wavelengths has allowed for time series of colour indices (green chromatic coordinate [gcc]) and vegetation indices (camera normalized difference vegetation index [NDVIc]) to be generated monitoring canopy structural and leaf biochemical cycles. However, these are limited in their spatial representation, mostly confined within their ecosystem. Satellite data at the appropriate spatial resolution could provide complementary information which could upscale these phenological observation to regional scale.
Until recently, there is a spatial mismatch between commonly used satellite derived land surface phenology and Phenocam data, therefore the complimentary information has not been fully exploited. Sentinel-2 can close this gap by providing data at 20 m resolution, matching the area captured by the Phenocam images and representing the dominant vegetation cover, with the potential of getting data once in every five days. However, suitability of Sentinel 2 data for land surface phenology estimation this has not been properly tested. Therefore, this study aims to look at key phenological metrics at 43 Phenocam sites (for a total of 60 years of data) covering a range of vegetation types (agricultural, deciduous broadleaf, evergreen broadleaf, evergreen needle leaf, grassland, shrub land and tundra) at a variety of spatial scales using Sentinel-2 data. Phenological matrixes derived from the Sentinel 2 are compared with those from the PhenoCam and MODIS data to assess the use of Sentinel-2 data for monitoring land surface phenology and bridging the scale gap between near surface and satellite remote sensing.

Area A - Board 36

Global warming which also referred as global climate change impacts on ecosystem functioning, and makes the monitoring of vegetation dynamics more difficult. One of the most sensitive indicators of the effect of global climate change is Phenology. However, monitoring the plant phenology and predicting the Start and End dates Of the growing Season (SOS,EOS) are challenging, because, drivers are not fully understood. Moreover, different studies have applied different methods to describe these phenological metrics. Greenness-based vegetation indices (GBVIs), such as Normalized Difference Vegetation Index (NDVI), are often used to parameterize land surface models, but in fact they are not ideal proxies for canopy photosynthesis and structure. These GBVIs fail in predicting the correct SOS and EOS, especially when greenness and photosynthesis become uncoupled (e.g. during drought, during winter in boreal needle-leaf trees). Furthermore, the validation of GBVIs-based SOS and EOS estimations with in-situ observations is still lacking for many ecosystems due to the broad spatial (1 km) and poor temporal resolution (bi-weekly or monthly) of commonly used sensors to derive GBVIs (e.g. MODIS). The advent of novel Very High Resolution Platforms (e.g. Sentinel-2) allows tracking phenology more accurately. In addition, spectral bands of Sentinel-2 offer the opportunity to calculate VIs related to pigment content such as the Modified Terrestrial Chlorophyll Index (MTCI).

In this study, we tested several structural and chlorophyll sensitive VIs derived from Sentinel-2 to improve the estimation of the SOS/EOS for three different plant types in Belgium.
Different phenological extraction methods (e.g. Savitzky-Golay, Polynomial function, and Harmonic Analysis of time series) with different ways of estimating the SOS and EOS (e.g. percentiles, first derivative function) are applied. The new pigment-based prediction of phenology metrics are validated against in-situ observations of Gross Primary Production, based on eddy covariance measurements, and biophysical parameters (e.g. Leaf Area Index) measured at the field site.
Keywords: Sentinel-2 product; Gross Primary Productivity; Greenness; Structural and chlorophyll indices; Smoothing function methods; Start and end of Season.

Area A - Board 37

Wetlands play a key role as biodiversity hotspots and provide a wide spectrum of ecological goods and services. Among the diverse types of wetlands, peatlands stand out for their extraordinary capacity to accumulate large amounts of carbon over thousands of years due to their slow decomposition rates of organic matter acting as carbon sinks. Although peatlands are mainly found in the boreal and subarctic regions of the Northern Hemisphere, a unique system of peat-accumulating wetlands occurs across the largest high-altitude semiarid region of South America represented by the Altiplano plateau in the Central Andes. These peatlands, locally known as “bofedales”, are sustained by groundwater flows and snow- and glacier-melt, and play a key role as a) water reservoirs and forage resources for the livestock of local indigenous communities; b) carbon sinks with higher accumulation rates than those of the Northern Hemisphere; and c) valuable paleoclimate archives as they store up to 10 m of peat accumulated over the last ~10.000 years. In the present study, we mapped all peatlands for the entire western Altiplano and assess their spatio-temporal trends in land surface phenology (LSP) using a non-parametric approach (the npphen R package) and dense Landsat NDVI time series. The npphen package implements basic and high-level functions for manipulating vector and raster data to obtain high resolution spatial and temporal LSP reconstructions. Advantages of ‘npphen’ are: its flexibility to describe any LSP pattern (suitable for arid ecosystem), it handles time series or raster stacks with and without gaps (suitable for Landsat data), and it provides confidence interval for the expected LSP at annual basis, useful to judge anomalies in LSP in a probabilistic way. Using this approach and taking advantage of a dense Landsat time series available for this relatively cloud-free semi-arid area in this study we: 1) developed the first peatland inventory for this vast region, 2) characterize the multi-decadal annual LSP of 5,665 peatlands, and 3) assessed the spatio-temporal patterns of the LSPannual cycle and the phenological stability of these ecosystems.

Area A - Board 38

Improved estimates of irrigated versus non-irrigated areas in Africa are essential to properly assess crop productivity, yield gaps, the consumption of water resources and drought impacts. Information on farming systems (defined as irrigated versus rainfed agriculture) is, however, not readily available in Africa. One reason may be the complexity of mapping croplands in Africa in general, given that African croplands are highly dynamic and fragmented agro-ecological systems, which are usually characterized by small-scale agriculture and subsistence farming. We hypothesized that optimized vegetation index (VI) harmonics (defined as the sines and cosines function of the periodicity of vegetation) from either Sentinel-2 or Landsat time series data can be used to accurately map farming systems in Africa. Zimbabwe and Free State, South Africa, were used as study sites. The harmonics fitting curves for Sentinel-2 and Landsat time series were optimized according to the amount of noise and gap lengths in the respective series, changes in the harmonic degree (one to twelve), length of the harmonics period (one single year versus five years) and the inclusion or exclusion of the regression trend parameters. The optimization settings were gauged against the mapping accuracies for classifying rainfed versus irrigated croplands. A non-parametric machine learning (random forest) classification algorithm in Google Earth Engine was used since the harmonic parameters are essentially non-linear. Random forest can deal well with non-linear data variables and various data properties while accounting for data overfitting and outliers. Training points representing irrigated and non-irrigated fields were collected using high resolution imagery provided in Google Earth (in total 9.900 reference points for each class in each study site). The results showed that the 7th harmonics degree rendered the most accurate result for mapping farming systems in both data sets and study areas. Moreover, using long NDVI time series as the input for the harmonic analysis, rendered more accurate results than using only one harmonic period or single year. In highly dynamic landscapes, however, and in areas where the data gap and number of noisy pixels within a period/year is large, the periodicity of vegetation (or crop seasonality) may be too unstable to produce accurate mapping results.
Applying the most optimal regression settings on Sentinel-2 and Landsat time series data, respectively, farm systems mapping accuracies of >96% could be attained for both data sets and study areas (the Free State in South Africa and Zimbabwe). The accuracies for Sentinel-2 data were only 1 percent higher than for Landsat data. This small difference could be due to a shorter time series of the Sentinel-2 (2015-2018) data compared to Landsat-8 data (2013-2018). The results confirmed the potential of best-fit harmonics to correct for noise artefacts and data gaps for mapping spatially variable land use in Africa (assuming stable land use classes over the observation period). The testing of optimized time series analysis tools provides important insights for the development of Analysis Ready Data for the Earth System Data Cube that aims to provide users with ready to interpret satellite variables.

Area A - Board 39

1. Abstract
The present study evaluates the viability of monitoring extensive asparagus fields in Peru using the capabilities of the Sentinel 1 satellite. Ground data acquired during two campaigns per year for 2017 and 2018 have been used to analyse the radar response with respect to the crop evolution.

Preliminary results show significant correlation between the SAR observables and the crop growth stage while further research is currently undertaken to derive correspondence with crop biophysical parameters.

2. Introduction
According to the Food and Agriculture Organization FAO of the United Nations, Peru is the world’s second biggest asparagus producer. It is the third most exported agricultural product of Peru after grapes and mango [1,2]. The crop is highly recognized as a source of jobs in several regions of the country.

In this sense, the businesses in charge of the fields are constantly searching for ways to optimize the inputs resources and to improve crops yield. The capability of the Sentinel 1 satellite to systematically acquire data allows its use as a monitoring tool to measure the variables affecting the asparagus fields.

3. Test site and datasets

The fields are located near the city of Trujillo in the north of Peru as shown in figure 1. Since there are no dramatic changes in the average temperature, it is possible to grow two campaigns per year.

Single look complex data have from Sentinel 1 for the 2017 and 2018 years have been pre-processed in two different incident angles to analyse the multitemporal dual polarimetric sensor response to the crop changes.

4. Methodology

The measurements on the field provide a tool to make an interpretation of the single look complex images acquired and processed over the study area. Several parameters of the radar satellite response have been analysed, including the available backscatter signals as well as its ratio. Similarly, the parameters obtained after performing the dual polarimetric decompositions such as the alpha/entropy decomposition are analysed against the ground truth provided by the farmers.

5. Preliminary results

Initial outcomes of the analysis confirm the potential of sentinel 1 for monitoring the fields. The temporal trends of the radar observables can be correlated to the crop phenology provided by the farmers in the fields.

The following statements have been tested and confirmed:

• The HV channel shows remarkable sensitivity to the vegetation growth, including the double interaction between radiated energy with the soil and the asparagus stalks growing vertically at the beginning of the season (figure 2, right side).
• The backscatter ratio shows correlation with the increase in the canopy density and in general with the presence of vegetation in the fields
• The alpha angle obtained after dual polarization decomposition allows to clearly determine moments of harvest and the beginning of a new season.
• Deviations from the normal behaviour of the time series can be visually identified, therefore, they can be automatically detected. This was proved, for instance, with irregular peaks in the VV channel during a torrential rain period in the study area.
• The temporal trends of the different variables present a repetitive behaviour during the four seasons considered, thus, enabling the possibility to create crop models that replicate the crop evolution based on historical satellite and ground truth data.
• Simplified manual algorithms show the potential to retrieve the most important crop stages, such as the beginning of a new season, the vertical growth of asparagus spears, the establishment of a fern and the harvest dates.
• Initial investigations also show the possibility to identify irrigated from non-irrigated parcels. This happens considering that the site is mostly sand, thus, the potential differentiation comes from separating wet and dry sand.

Using these preliminary findings, current research focuses on the design of an automatic classification tool to retrieve the crop growth stage, based on the ground and satellite data from previous campaigns. Additional polarimetric features, such as the multitemporal dual pol Wishart dissimilarity measure, are also being tested and will be included in the tool in order to increase the number of crop stages retrieved as well as the accuracy of its performance.

6. Current and future research

Since the initial step to analysing correctly the crop is to understand the interaction between the electromagnetic energy radiated from the sensor with the different moments of the crop evolution, a tool previously used for this purpose and used in other crops is being tested.

An adaptation of a quad polarimetric change detector has been done to the case of dual pol data, with the aim of evaluating the radar changes caused by the changes on the structure and dielectric constant of the crop during its evolution [3]. Multitemporal analysis based on change detection facilitates the understanding of the radar backscatter response [4].

The tool was originally designed as a quad polarimetric methodology to visualize the presence of the scattering mechanisms at any point in time during the season [4]. It is based on the comparison of the polarimetric information described by the covariance matrix of a target in two different acquisitions. This comparison, equivalent to change detection, describes what scattering mechanisms have increased and / or decreased from one acquisition to the other.
The application for the dual pol case is currently under development, understanding the physical meaning of the results obtained considering the limitations of the C2 matrix to fully characterize the polarimetric response of a target.

On the other, as part of the project to which this study is part of, data assimilation techniques will be studied in the coming months to integrate optical data and agronomical modelling to the asparagus crop, so that a complete solution for monitoring the crop and forecasting crop yield can be developed.


7. Acknowledgement

This research was supported by the Project EO4cultivar, led by Environment Systems Ltd and co-funded by the UK Space Agency. The EO4cultivar project is an international collaboration project with partners in the UK, Peru and Colombia through the International Partnership Programme (IPP) of the UK Space Agency.

References

[1] Rios L., “Agro-industries characterization and appraisal: Asparagus in Peru,” Food and Agriculture Organization of the United Nations (FAO), 2007.

[3] A. Marino and A. Alonso-Gonzalez,” Optimisations for different change models with polarimetric SAR,” EUSAR 2018; 12th European Conference on Synthetic Aperture Radar, Aachen,
Germany, 2018, pp. 1-5.

[4] C. Silva and A. Marino and J. M. Lopez-Sanchez and I. Cameron, Monitoring agricultural fields using an optimisation of the difference of covariance matrices for POLSAR, (IGARSS), 2018

Area A - Board 40

At northern latitudes above 63° N, the revisit time of Sentinel-2 can result in an image every two to three days. The availability of image time series which capture phenological changes should lead to improved identification of alpine vegetation classes; previously, access to few images and spectral similarity among the classes have been problematic for mapping of these classes. Among the new methods positioned to take advantage of the increased amount of data are machine learning and data assimilation methods such as Hidden Markov Models. In this project, a Hidden Markov Model and a Random Forest model have been applied and the results compared. A Sentinel-2 data time series with eleven granules from May to September of 2018 were used in the northern alpine area of Abisko, Sweden. The Hidden Markov Model determines classes based on stages and transition probabilities through time and recursively determines the final class. Due to the focus of Hidden Markov Models on distinct phenological stages, particular attention has been paid to identification of the class “snowbed vegetation”, as it goes through the stages of being snow-covered, to exposed soil, and eventually (depending on the environmental conditions) to a vegetated state of moss or alpine grasses. In addition, this study examined the impact of conducting versus not conducting a topographic normalization for alpine vegetation classification particularly in the context of using a time series of Sentinel-2 images and the classification algorithms named here. The results indicate that when an increased number of images are classified using Random Forest without topographic normalization, the impact on class accuracy is less than when using Hidden Markov Model. When comparing the classification results from applying a Hidden Markov Model and Random Forest, the Hidden Markov Model requires a substantially increased number of reference data plots as compared to the Random Forest classification, which may outweigh any small gains in accuracy achieved.

Area A - Board 41

The Copernicus Sentinels provide data that can profoundly impact vegetation phenology studies. In particular the high spatial and spectral resolution of Sentinel-2 (S2) data opens the gates to unprecedented opportunities to overcome some of the difficulties faced in remotely sensed phenological research based on the previous generation of satellite (e.g. MODIS, LANDSAT, AVHRR, … ). Moreover, the integration of S2 and Sentinel-3 (S3), which has a higher temporal resolution at the expense of a coarser spatial resolution, can further enhance these advances by complementary balancing the specific limitations of the two missions.
Currently, assessments of the accuracy in the estimation of key phenological events (i.e. start of season, end of spring growth, start and end of senescence, …) based on S2 and S3 data are still scarce, mainly because of the lack of consistent near-surface monitoring networks that could provide ground-truth data.
In this study, we take advantage of continuous ground NDVI measurements collected through a broad observation network encompassing Alpine and Mediterranean ecosystems in Europe. The ground observation network includes a wide range of vegetation communities like alpine grasslands, deciduous needle-leaf forests and tree-grass ecosystems. This comprehensive ground NDVI dataset is compared to Sentinel-derived vegetation indexes.
Our objective is to understand whether S2 and S3 vegetation indexes (VIs, e.g. NDVI) can accurately describe the seasonal trajectory of vegetation development in these ecosystems and thus provide accurate estimates of phenological events. In particular we compare phenological events derived from S2 and S3 time series to phenological events extracted from ground NDVI trajectories. We discuss the results of this comparison in the light of different temporal and spatial resolutions provided by S2 and S3, the effect of cloud cover and the effect of landscape heterogeneity. Finally, we evaluate the potential of using Sentinel data to detect the impact of climate extremes on ecosystems phenology and seasonal dynamics.

Area A - Board 42

The circumpolar boreal belt comprises about one-third of the global forest area and constitutes one of the largest biome in the world. These regions are particularly vulnerable to climate change. There is a strong consensus that warming of boreal regions is unfolding faster than anywhere else on Earth. Such changes are affecting substantially the structure and functioning of ecosystems and are expected to further amplify as we move into the future. The progressive rise in CO2 fertilization effect and the concurrent reduction in thermal limitations to plant growth have led to a positive trend in primary productivity over most of the boreal regions. This is also reflected in an increase of green leaf density, as expressed by the Leaf Area Index (LAI). Several global satellite LAI products are now available covering the past three decades. However, differences across these dataset in varying sensitivity to spectral variations, satellite orbit drift, as well as the different retrieval algorithms, may lead to divergent vegetation dynamics that are difficult to validate properly. Here we provide a systematic inter-comparison of three satellite LAI products and explore the reasons of such discrepancies. The AVHRR-based GIMMS-NDVI version 3g (LAI3g), the Land Long-Term Data Record (TCDR) and the Long-term Global Mapping (GLOBMAP) series are compared for the period 1982-2011 over high northern latitudes (>50◦N). We focus on the growing season average LAI as key diagnostic metric and evaluate his temporal trend for each single product. Since there are no adequate ground networks for the systematic measurement of LAI within the study area, we use alternative satellite dataset as benchmark in a “process-oriented” perspective. Specifically, the evaluation of LAI has been carried out through direct comparison against an independent proxy of aboveground biomass as captured by the Vegetation Optical Depth (VOD) retrieved from passive microwave sensors. Ancillary information in the form of land cover and land cover change maps are also used to better understand ecosystem responses. Results show that trends in vegetation activity from TCDR LAI product are consistent with VOD estimates, considering the diversity of definitions. Conversely, LAI from GLOBMAP and VOD show divergent patterns for 60% of the study area. Greening is especially pronounced in the TCDR data, conversely GLOBMAP shows a slightly negative trend. LAI3g shows a reasonably good agreement with TCDR while different patterns with respect to GLOBMAP. Finally, we investigate how different boreal ecosystems have responded to environmental changes during the last three decades. TCDR records – the ones that are more consistent with the other dataset – support the hypothesis that there is a generally widespread “greening” of non-forest vegetation. Results emphasize that spatial patterns in the greening of the high northern latitudes are not consistently represented by the diverse satellite products and that a careful usage of these products should be taken to make larger assessments of ecosystem-response to climate change over the last three decades.

Area A - Board 43

This study investigates the long term phenology and albedo record for the ABoVE domain, using the MODIS and VIIRS BRDF, Albedo, and NBAR products (MCD43 and VNP43) at 500 m and 1 km spatial resolutions. The Arctic-Boreal Vulnerability Experiment (ABoVE) is a large spatial and temporal scale research project designed to build knowledge, over a decade, of the vulnerability and resilience of Arctic ecosystems and associated human systems. The targeted Arctic-Boreal Region (ABR) is in Alaska and western Canada, where the effects of climate change are unfolding rapidly. The Moderate Resolution Imaging Spectroradiometer (MODIS) MCD43 data record extends from 2000 to present, and the Visible Infrared Imaging Radiometer Suite (VIIRS) VNP43 data record provides continuity since 2012. Together these sensors provide a substantial temporal record of the surface energy budget through information on surface BRDF and albedo at moderate spatial resolution and daily temporal resolution that can complement extensive ABoVE field-based data collection and airborne remote sensing campaigns in the ABR. With the continuous MODIS and VIIRS view angle-corrected reflectance (Nadir BRDF Adjusted Reflectance, NBAR) data, Land Surface phenology (LSP) can also be monitored with far greater precision.

This study analyzes the seasonal phenology of the region and explores the corresponding albedo changes. Four phenological dates (greenup onset, maturity onset, senescence onset, and dormancy onset) are retrieved from daily NBAR-EVI (Enhanced Vegetation Index) to determine the dynamics of the shoulder and growing seasons over the past two decades. Changes in phenology reflect alteration of regional surface/atmosphere energy budgets that occur via changes to surface albedo – and particularly those occurring on shoulder seasons – which is an important factor for the ABR region since it directly relates to rates of permafrost warming and thawing. Alterations to the ABR energy balance affect warming/cooling feedbacks that propagate to the global climate system. Excellent continuity is demonstrated in the phenological and albedo metrics provided by the MODIS and VIIRS records, showing that analysis of long-term trends across the satellite records provides meaningful results.

Area A - Board 44

The High Arctic Svalbard is characterized by a short and intense growing season, where even small changes in the timing of the growing season highly influences the plant production and the population dynamics of most animals, birds, and insects. Previous we have processed clear-sky time-series of MODIS data to characterize the growing season for the 2000-2017 periods and revealed huge variability between the years in the onset of the growing season in and plant productivity. We have also revealed a strong link between the sea-ice dynamic and seasonal NDVI. The vegetation on Svalbard is often dominated by lichens or bryophytes with a different phenological cycle compared with vascular plants. Due to the coarse resolution to MODIS data this creates many uncertainties in the growing seasons mapping. With Sentinel-2 data we are able to characterize the seasonal cycle on vegetating type level and deal with this problem. With Sentinel-2 data we concentrate the mapping of central parts of Svalbard at about 78°N, which are the warmest areas on the island with a mean July temperature of 6.5 °C. Due to the location close to the North Pole and the polar orbit of Sentinel-2, data are obtained every day, or even twice a day, at around 12 UTC. Due to this frequent cloud cover a reliable cloud-mask has to be developed to remove all the noisy data before the interpolation to daily data. Then, at such northern location the solar elevation angle is low, creating shadows in the topographic complex area. In this presentation we show how we deal with these challenges in processing of a clear-sky time-series. Then, we use this clear-sky time-series to explore which wavebands best identify different phenophases at different vegetation types, as validated from RGB time-laps cameras (‘phenocams’) in field. The results indicate that NDVI is good enough to map the onset of the growth in vascular plant dominated areas, while other band combination (indices) is better in other vegetation types and to characterize autumn phenophases.

Area A - Board 45

Photosynthesis or gross primary production (GPP) is the uptake of carbon through land ecosystems. Reliable global estimates of GPP are needed to quantify changes in the land carbon sink. GPP has been estimated using optical satellite observations, which however may be frustrated by cloud cover or weak illumination conditions. On the other hand, microwave satellite observations are weather and illumination independent and allow global daily retrievals of vegetation optical depth (VOD), a quantity related to biomass and relative water content. Thus, in theory, VOD can also serve as a proxy for GPP, but relationships between VOD and GPP are still unclear.

Here, we assess the potential of estimating GPP globally from VOD derived from various microwave frequencies (L, C, X, Ku-band) and sensors, including the Advanced Scatterometer (ASCAT), the Soil Moisture Ocean Salinity (SMOS) mission, and the Advanced Microwave Scanning Radiometers (AMSR-E and AMSR2). We compared VOD with (1) GPP estimates from FLUXNET eddy covariance sites, (2) with the upscaling-based FLUXCOM GPP product, and with (3) observations of sun-induced fluorescence (SIF) from the Global Ozone Monitoring Experiment-2 (GOME-2), which is a direct proxy for GPP.

Our results [1] show widespread positive correlations between VOD and GPP, which are of similar magnitude as correlations between SIF and GPP. In deciduous forests, the change of VOD (ΔVOD) has a higher correlation with GPP than the original VOD signal, which is indicative for changes in leaf tissue. This suggests that VOD and ΔVOD can be jointly used to estimate GPP. Based on this result, we then globally predicted GPP from VOD and ΔVOD by training generalized additive models (GAM) to GPP measured in situ at FLUXNET sites. The best VOD-based GPP estimates reached higher performances in predicting monthly GPP anomalies than a SIF-based GPP estimate. Our results demonstrate that VOD from microwave satellites can complement optical satellite data in long-term, global monitoring of photosynthesis on temporal resolutions down to one day. In our presentation, we will show the potential of the new GPP dataset for mapping phenological variability and changes.

[1] Teubner, I.E., Forkel, M., Jung, M., Liu, Y.Y., Miralles, D.G., Parinussa, R., van der Schalie, R., Vreugdenhil, M., Schwalm, C.R., Tramontana, G., Camps-Valls, G., & Dorigo, W.A. (2018). Assessing the relationship between microwave vegetation optical depth and gross primary production. International Journal of Applied Earth Observation and Geoinformation, 65, 79-91, doi: 10.1016/j.jag.2017.10.006

Area A - Board 46

The health and extent of world’s forests play a very significant role as part of the studies on the Biosphere. Indeed, the estimate of the amount of Earth’s biomass gives important insights into the residual resources and capacity of regeneration of our Planet. Among the others, the Amazon rainforest deserves a special attention, first for its large extent and bio-diversity, but especially because deforestation phenomena have recently increased at an alarming rate, jeopardizing its ecosystem and influencing climate changes.
Spaceborne systems provide an accurate and cost-effective solution for monitoring forests coverage at a global scale. Optical and laser sensors are widely used for mapping forests coverage and structure as well as Synthetic Aperture Radar (SAR) sensors. In particular, the latter, given their daylight independence and capability to penetrate clouds, represents a unique solution for monitoring vegetated areas. Indeed, tropical rainforests as the Amazon, are characterized by a wet season of about six months per year, during which trees are constantly hidden under a thick layer of clouds, and therefore not visible by optical sensors.
Developed in the framework of the Living Planet Fellowship 2018 named HI-FIVE (High-Resolution Forest Coverage with InSAR & Deforestation Surveillance), the presented work aims at providing a strategy for the exploitation of ESA long-term database for the systematical monitoring of the worlds’ forest extent. Thanks to its short revisit time and large coverage, the Sentinel-1 mission is a remarkable data source for current and future scientific applications.
The proposed methodology is based on the combined exploitation of the detected backscatter and of the correlation between SAR images from Sentinel-1 interferometric time-series. In particular, we exploit the interferometric coherence to estimate the evolution of temporal decorrelation of the observed target on the ground and use a decreasing exponential model to fit the data. The retrieved model’s parameters show a highly informative content and can be used as input observables for the derivation of an ad-hoc classification algorithm based on machine learning.
In the final work, we will show the analysis of the retrieved parameters with respect to each of the considered classes (forest, non-forest, and man-made), as well as their relationship and impact on the classification. The results are provided over a test site situated in the Rondonia state, Brazil, where severe deforestation phenomena take place and an up-to-date mapping is therefore highly recommended. Moreover, we will assess the advantages and drawbacks of using Sentinel-1 repeat-pass interferometry for classification purposes. Finally, we will introduce the potential of combining this approach with high-resolution forest/non-forest maps from bistatic TanDEM-X data for an effective monitoring of on-going deforestation.

Area A - Board 47