Conference Agenda

Rice is the most important food crop in Asia and the mapping and monitoring of paddy rice fields is an important task in the context of food security, food trade policy, water management and greenhouse gas emissions modelling. Asia’s biggest rice producers are facing increasing pressure in terms of food security due to population and economic growth while agricultural areas are confronted with urban encroachment and the limits of yield increase. At the same time demand for rice imports is increasing, spurred by global population growth.

Despite the importance of knowledge about rice production the countries official land cover products and rice production statistics are of varying quality and sometimes even contradict each other. Available remote sensing studies focused either on time-series analysis from optical sensors or from Synthetic Aperture Radar (SAR) sensors. We try to address the sensor specific limitations by proposing a paddy rice mapping approach that combines medium spatial resolution, temporally dense time-series from the optical MODIS sensors and high spatial resolution time-series from the Sentinel-1 A/B SAR sensors.

We developed a method to use MODIS time-series and a one-class classifier to create medium resolution rice maps [1]. In a next step we used these medium resolution rice maps to mask Sentinel-1 Interferometric Wide Swath images, which limits the amount of data to process and allows efficient rice mapping over larger areas. The high resolution rice masks are then created by segmentation of multi-temporal SAR images into objects, from which backscatter time-series are derived and classified. We created 10m resolution rice-maps that also allow seasonality extraction, given enough Sentinel-1 acquisitions. This method allows concurrent, accurate and high resolution mapping of paddy rice areas from freely available data. Results of our paddy rice classification will be presented for selected study sites in Asia.

1. Clauss, K.; Yan, H.; Kuenzer, C. Mapping Paddy Rice in China in 2002, 2005, 2010 and 2014 with MODIS Time Series. Remote Sensing. 2016, 8, 434.

The need for accurate information on the state and evolution of tropical forest types at regional and continental scales is widely recognized, particularly to analyze the forest diversity and dynamics, to assess degradation and deforestation processes and to better manage these natural resources (Achard et al. 2014).

A few global and continental land cover or forest cover products have been derived from Landsat satellite imagery at 30m resolution: they either contain detailed thematic information without temporal dynamics (Chen et al. 2015, Giri and Long 2010) or contain information on forest-cover changes over long time periods (10 to 30 years) without thematic classes such as the discrimination of evergreen forests (Kim et al. 2014, Hansen et al. 2013, Potapov et al. 2015).

The objective of this study is to map undisturbed evergreen and semi-deciduous forests at 30m resolution over the full tropical humid domain and to better characterize the changes and disturbances which occurred during the last 33 years in these forests. Therefore we exploited the full archive of Landsat imagery between 1984 and 2016 and developed a pixel-based automatic methodology which includes four steps: (i) pre-processing of the Landsat time series with cloud masking and filtering of sensor artefacts, (ii) single-date image classification (driven by a large spectral library) into three basic classes (evergreen forest, vegetative non-forest cover and poorly/non vegetated cover), (iii) creating forest/non-forest maps for three epochs based on the occurrence of non-forest classes, and (iv) production of a final map of detailed forest types based on the temporal succession of observed basic classes from 1984 to 2016.

The resulting map includes six classes: undisturbed forest cover, old and young vegetation regrowths, old or recent deforested areas (during last 10 years), recently disturbed areas (during last 3 years) and other land cover. This map at 30 m resolution allows the identification of small linear features such as gallery forests and of small disturbance events such as skid trails and logging decks. The use of a 33-year Landsat time series allows (i) to identify most deforestation and degradation events (when > 0.1 ha) that occurred during this period, (ii) to provide the dates of the forest disturbances, and (iii) to considerably reduce the confusion that usually occurs with small scale agricultural fields (shifting cultivation, tree plantation and irrigated crops…). Finally we characterize the deforested and disturbed classes by providing their timing and occurrence (date of first and last events, number of events).

The accuracy of the forest map was assessed over Africa from an independent sample of reference data (3830 plots) created through visual expert interpretation of Landsat imagery at several dates and finer resolution satellite imagery with an overall agreement of 90%. The pan-tropical map and the accuracy assessment results over Africa will be presented at the conference.

It is intended in the future to adapt and apply the methodology on Sentinel-2 data for a better characterization of forest-cover disturbances.

Remote sensing is an important tool for understanding the spatial and temporal dynamics of forest disturbances over large areas. The most recent developments in disturbance detection algorithms utilize dense time series information that enables the detailed characterization of abrupt and gradual disturbance events. These advances allow the mapping of a wide variety of disturbance agents, including harvest, fire, blowdown, and insect attacks. However, current algorithms have primarily been developed and tested in the forest ecosystems of North America, which are characterized by relatively homogeneous coniferous forests, little to none management, and medium to large-scale disturbance patches. Those algorithms have not been validated for mapping forest disturbances in Europe yet, where forest ecosystems are much more variable in terms of species composition, landscape structure, and forest management. Consequently, our aim was to evaluate current disturbance detection algorithms for mapping forest disturbance in Europe by i) comparing spectral-temporal characteristics of forest disturbances across a sample of five protected forest areas in the temperature forests of central Europe; and by ii) comparing disturbance characteristics inside the protected forests to the surrounding unprotected forests in order to understand the compound effect of natural and management disturbances on spectral-temporal characteristics. We utilized dense Landsat time series from the USGS and ESA archives covering the time period 1985 to 2016. We mapped forest disturbance characteristics (i.e., magnitude and duration) using state-of-the-art time series tools in conjunction with random forests classification and a set of photo-interpreted reference plots. Primary results show that disturbances were detected with high accuracies (>90% overall accuracy) across all sites. Spectral-temporal characteristics varied substantially within and outside protected forests. In particular, unmanaged areas showed more long-term disturbances – likely related to blowdown and following bark beetle mortality – whereas disturbances in managed forests were mainly short and of high magnitude. However, temporal patterns of disturbances were similar inside and outside the protected areas, suggesting that spectral disturbance patters outside protected forests are superimposed by a strong management signal. Our analysis improves the understanding of forest disturbances – and how to map them using remote sensing – for European temperate forests, a forest ecosystem of high economic and ecological value. As such, this study paves the way for using long and dense time series of optical satellite data, i.e. Landsat and Sentinel-2, for mapping and understanding forest dynamics across Europe.

Wetlands are amongst the planet‘s most productive ecosystems providing a wealth of ecosystem services, e.g., nutrition, flood control and protection, or support of biodiversity, to name a few. Nevertheless, wetlands are exposed to multiple threats due to climate change, agricultural pressure, hydrological modifications, fragmentations, etc. Thus, a consistent mapping and monitoring of global wetland ecosystems is very important to track changes and trends aiming at supporting wetland conservation and sustainable management. Although EO data are ideal for large scale inventorying of wetlands, the large diversity of them makes remote detection particular challenging. This diversity and resulting challenge has been tackled by many researchers applying different sensors (optical and radar) and mapping techniques to delineate wetland from non-wetland areas. A global and homogenous inventory, however, is still not available and scope of ongoing research.

Herein, we present an innovative and operational water and wetness product building on data from Sentinel-1 SAR and Sentinel-2 MSI complemented with historical data from the Landsat missions. Rather than trying to detect wetlands in the ecological sense, we derive wetlands in the physical meaning identifying the wetness of the underlying land surface.

Using a hybrid sensor approach, i.e., combining optical and radar observations, provides a more robust wetland delineation with optical imagery being more sensitive to the vegetation cover and radar imagery to soil moisture content. Additionally, the higher frequency of observations stemming from the combined data streams contributes to a better characterization of seasonal dynamics which is important so that seasonal and temporary changes do not lead to false conclusions of the overall long-term trend in wetland extent. Within the domain of optical remote sensing, the identification of wetlands is based on the enhancement of the spectral signature using bio-physical indices sensitive to water and wetness and subsequent derivation of a water and wetness probability index. The radar-based algorithm builds on geophysical parameters, surface soil moisture dynamics and water bodies, derived from historical Envisat ASAR and Sentinel-1 backscatter time series to identify permanent/temporary wet and flooded areas. In addition, it is possible to identify flooded vegetation according to the double-bounce scattering principle in densely vegetated wetlands. The non-flood prone areas are masked using the Height Above Nearest Drainage (HAND) index. After the separate processing of the optical and radar imagery, the data is fused into a combined water and wetness product. With our methods we aim at detecting the current status of wetland areas, but also to capture the historic evolution taking into account the past 25 years in a fully automated manner.

The above described methods are currently applied for several large regional sites throughout Africa within the GlobWetland Africa project and for the Pan-European production of the “water and wetness” High Resolution Layer of the Copernicus Land Monitoring Service. We will further present a thorough validation of the product for different wetland ecosystems and discuss remaining issues, mainly due to global data availability and coverage.