Mapping pan-European land cover using Landsat spectral-temporal metrics and the European LUCAS survey

Highlights

• Landsat-8 spectral-temporal metrics robust predictors of pan-European land cover

• Using one year and three years of Landsat-8 data yielded similar overall accuracies.

• Accuracy of artificial land and croplands increased using three years of Landsat data

• LUCAS data useful for pan-European land cover mapping

• Results more consistent and accurate than CORINE in 12 classes

Abstract

This study analyzed, for the first time, the potential of combining the large European-wide land survey LUCAS (Land Use/Cover Area frame Survey) and Landsat-8 data for mapping pan-European land cover and land use. We used annual and seasonal spectral-temporal metrics and environmental features to map 12 land cover and land use classes across Europe. The spectral-temporal metrics provided an efficient means to capture seasonal variations of land surface spectra and to reduce the impact of clouds and cloud-shadows by relaxing the otherwise strong cloud cover limitations imposed by image-based classification methods. The best classification model was based on Landsat-8 data from three years (2014–2016) and achieved an accuracy of 75.1%, nearly 2 percentage points higher than the classification model based on a single year of Landsat data (2015). Our results indicate that annual pan-European land cover maps are feasible, but that temporally dynamic classes like artificial land, cropland, and grassland still benefit from more frequent satellite observations. The produced pan-European land cover map compared favorably to the existing CORINE (Coordination of Information on the Environment) 2012 land cover dataset. The mapped country-wide area proportions strongly correlated with LUCAS-estimated area proportions (r = 0.98). Differences between mapped and LUCAS sample-based area estimates were highest for broadleaved forest (map area was 9% higher). Grassland and seasonal cropland areas were 7% higher than the LUCAS estimate, respectively. In comparison, the correlation between LUCAS and CORINE area proportions was weaker (r = 0.84) and varied strongly by country. CORINE substantially overestimated seasonal croplands by 63% and underestimated grassland proportions by 37%. Our study shows that combining current state-of-the-art remote sensing methods with the large LUCAS database improves pan-European land cover mapping. Although this study focuses on European land cover, the unique combination of large survey data and machine learning of spectral-temporal metrics, may also serve as a reference case for other regions. The pan-European land cover map for 2015 developed in this study is available under https://doi.pangaea.de/10.1594/PANGAEA.896282.

Introduction

Land cover is a key variable influencing the Earth's energy balance, the hydrological and carbon cycle, and the provisioning of natural resources and habitat (Bonan, 2008; Brovkin et al., 2004; Foley et al., 2005). Satellite imagery has long been used for mapping land cover from space (Townshend et al., 1991). The first systematic observations of the Earth land surface became available with the launch of Landsat-1 in 1972, which opened new opportunities for automated land cover mapping (Gordon, 1980; Landgrebe, 1976). In the 1990s and early 2000s, a number of operational land cover programs emerged based on Landsat and Landsat-like data, e.g. CORINE (Coordination of Information on the Environment) land cover in Europe (Büttner, 2014), NLCD (National Land Cover Database) in the USA (Homer et al. 2004, Jin et al., 2013, Vogelmann et al., 2001), EOSD (Earth Observation for Sustainable Development of Forests) in Canada (Wulder et al., 2008), NCAS-LCCP (National Carbon Accounting System-Land Cover Change Project) in Australia (Furby, 2002), and PRODES (Programa Despoluição de Bacias Hidrográficas) in Brazil (INPE, 2002). Over the last decades, tremendous progress has been made in developing up-to-date and accurate land cover information. The availability of high quality, ready-to-analyze medium resolution sensor data (Drusch et al., 2012; Wulder et al., 2012) and advances in IT infrastructures and cloud computing have made it possible to produce regional to global land cover maps in a relatively short period of time (Gong et al., 2013; Xiong et al., 2017).

In Europe, CORINE provides the most comprehensive and detailed maps of land cover and land use change at medium spatial resolution (Büttner, 2014). The CORINE program was established by the European Commission to create harmonized geographical information on the state of the environment in the European Community. CORINE land cover is mapped in 44 classes using a minimum mapping unit of 25 ha for areal phenomena and a minimum width of 100 m for linear phenomena. The first CORINE land cover map was created for the reference year 1990 using satellite data from 1986 to 1998. The project involved 25 countries, and it took ten years to complete. Subsequent land cover updates were produced for the years 2000, 2006, and 2012. Over the years, the product chain has evolved and production times have decreased. The most recent CORINE land cover 2012 involved 39 countries and took two years to complete. CORINE's approach builds on harmonized protocols and guidelines that are executed independently by each country. This way, countries can integrate their national efforts and inventory resources and make the best use of local expert knowledge. The disadvantage is that variable mapping approaches may be employed and that the majority of countries heavily rely on visual interpretation of high-resolution satellite imagery. Only a few countries apply semi-automatic solutions combining national in-situ data and satellite image analysis. Considering the increasing demand for consistent and up-to-date land cover information, there is a strong need to develop and test automated approaches for mapping pan-European land cover.

The increasing and free availability of ready-to-analyze medium resolution imagery has brought new opportunities for the automated mapping of land cover and land cover change over large areas (Hansen and Loveland, 2012). Traditional image-based classification approaches require cloud-free or near-cloud free imagery, and they often go along with heavy user-interaction and visual interpretation methods. Scaling image-based approaches over large areas is technically and logistically challenging, particularly in areas that have a lot of clouds. Mass-processing of Landsat and Landsat-like data has become easier with per-pixel approaches that integrate multiple (potentially all) cloud-free reflectance values for a given pixel and time period, and produce seamless mosaics across image boundaries (Roy et al., 2010).

A variety of per-pixel approaches exist that may be categorized as follows: composites based on the best pixel observation within a given time period (Griffiths et al., 2012; Hermosilla et al., 2015; Roy et al., 2010), spectral-temporal metrics based on statistical moments (Azzari and Lobell, 2017; Potapov et al., 2015), and phenological metrics derived from fitting time series functions (Schwieder et al., 2016; Zhu and Woodcock, 2014a). All these methods have advantages and disadvantages; sometimes they are used in combination. Composites select the best pixel observations for a given time period using a set of rules and optimization functions. The goal is to reduce atmospheric effects while retaining original reflectance observations. Spectral-temporal metrics do not (necessarily) yield observed reflectance. Rather they describe the spectral distribution and variance of a pixel over a year or season. Spectral-temporal metrics are a simple means to express seasonal land cover dynamics related to vegetation growth and land use, and they have been used to map land cover (Azzari and Lobell, 2017; Hansen et al., 2011), tree species (Thompson et al., 2015), and forest disturbances (Potapov et al., 2015). Time series approaches yield a detailed model of land surface phenology because they directly estimate parameters that describe the timing, shape, and periodicity of the land surface changes. In this respect, time series methods are superior to spectral-temporal composites, but they also have higher requirements regarding data density and computing power. Research is ongoing to evaluate how to scale time series methods effectively across very large areas. In this study, we build on spectral-temporal metrics to automate land cover classification across Europe.

Collecting reference data for training and validation is an essential component in land cover mapping. Ideally, a combination of photo-interpretation and field work is employed that covers the expected range of land cover and environmental conditions (Strahler et al., 2006). If reference data are used for validation, then a probability sampling design is required to ensure statistical estimates of map accuracy (Stehman, 1997). Since reference data collection takes a major effort, there is a paucity of reference datasets at the regional (country to continental) and global scale. For global scale validation, tremendous progress has been made through crowdsourcing activities (Fritz et al., 2017) and probability-based samples of high-resolution imagery (Pengra et al., 2015). These approaches fill a crucial gap, but they also have trade-offs concerning interpretation uncertainty (Foody and Boyd, 2013) and sample size (Banaszkiewicz et al., 2014). Countries that support national land cover programs often rely on photo-interpretation and field survey data (Vogelmann et al., 2001; Homer et al., 2004; Wickham et al., 2010), which are usually not publicly available.

The European Land Use/Cover Area frame Survey (LUCAS) provides a unique source of publicly available land cover survey data. LUCAS is coordinated by Eurostat, the statistical office of the European Commission, with the aim to inform decision makers and the general public about changes in management and coverage of the European territory (Martino and Fritz, 2008). LUCAS collects information on land cover and land use, agro-environmental, and soil data every three years in the European Union since 2006. The survey is conducted in two phases. In the first phase, land cover is recorded via photo-interpretation at a systematic sampling grid with points spaced 2 km apart. In the second phase, a stratified sample is visited in the field. This sample represents the largest harmonized in-situ land cover record in Europe. In 2012, the in-situ measurements involved 750 field surveyors. The main purpose of LUCAS is statistical estimation. The survey was not designed for mapping. However, the potential benefits are high, if LUCAS data can be successfully used to train modern machine learning classification algorithms. Past remote sensing studies that incorporated LUCAS data have been limited to statistical estimates (Gallego and Bamps, 2008), map validation (EEA, 2006), and semi-automated training of classification models over smaller regions (Mack et al., 2017). No study has utilized LUCAS for automated, pan-European land cover mapping.

The objective of this study was to develop a method for mapping pan-European land cover by automated classification of Landsat spectral-temporal metrics using the largest European-wide land cover survey LUCAS. We evaluate the mapping approach at the continental and regional scales and specifically address the following research questions:

• How does data density impact classification accuracy?

• How do environmental predictors improve classification accuracy?

• How does the automated land cover map compare to the existing CORINE product?