Land Cover Cloud Analytics: from Global Services to Regional Insights

Land Cover Cloud Analytics: from Global Services to Regional Insights

DOI:https://doi.org/10.52939/ijg.v18i6.2451

Land cover serves as an important constraint for all kinds of land use and human activities, and as an indicator for natural as well as societal dynamics on Earth’s surface. Mapping and monitoring land cover (LC) is a key objective of space-based earth observation (EO). Since the beginning of regular global satellite imagery coverage with the Landsat platforms, huge repositories of EO data have been built up, recently complemented by a range of other platforms, e.g. ESA’s Sentinel family.

Many of these EO archives now are available under open licenses, facilitating the development of derivative data products like global land cover as presented by Tsendbazar et al., in [1] and Mora et al., in [2]. Original imagery as well as data products have enormous storage requirements and are subject to continuous expansion, enhancement and updating. Cloud environments therefore are the only feasible architectures where land cover archives and data products can be maintained and made readily accessible for users worldwide.

Land cover, as just one example of EO based products, is being widely used for mapping and establishing context for spatial planning [3], environmental monitoring, and land use decisions. In addition, it serves as an integrated indicator for various spatial processes, and as such is of interest whenever a deeper assessment and understanding of regional geographies is required. Spatial analysis aims at generating information in support of decision making, and this information created from a ‘densification’ of data is the main intended output from geospatial methods and tools. This study aims at demonstrating the power of cloud based spatial analysis, leveraging the online availability of EO based land cover data products. It uses the ‘World Land Cover 30m BaseVue’ data set originally created and published by MDA, and available inter alia through Esri’s Living Atlas cloud infrastructure.

Land cover (LC) not only is an integrative indicator for factors like climate, substrate and the interaction of ecological processes, but also is well suited to monitoring change induced by global (climate) processes [4] as well as regional human activities. Land cover also is a factor determining the local potential for human habitation and economic activities, including its role in complex biogeo climatic feedback cycles. While in extreme cases LC is essentially ‘nil’ – considering barren rock or soil or the special case of water, snow and ice, in most cases it requires a scale-dependent and systems-oriented definition. This is particularly true for multi-level vegetation covers as well as seasonal agricultural patterns. Like most mapping tasks, the schema or classification employed to identify area categories depends on the purpose of the resulting representation. This paper is entirely focused on remote sensing based and thus phenological acquisition [5] [6] somewhat restricting the types of LC units which can be distinguished. On the other hand, this approach facilitates work with globally homogeneous seamless data sets easily available through online services. The work performed towards this paper aims at determining the analytical potential and constraints for regionalized insights into LC distribution and dynamics. These again are considered highly relevant and factually indispensable for the monitoring of impacts from direct human action as well as climate change. This literally ‘top-down’ approach of global land cover monitoring is demonstrated and discussed using a Central Asia case study within the borders of the Kyrgyz Republic.

2. Materials and Methods

2.1 Study Area

This exploratory study was conducted covering the territory of the Kyrgyz Republic in Central Asia (Figure 1). As an approx. 200000 km² country with a rich topography spanning different climates from sub-mountain, mountain, high-mountain and nival zones, it is home to a correspondingly diverse set of land cover categories well suited for exploring the performance and quality of global LULC data sets. From an agricultural perspective Park et al., [7] explore the cropland suitability of Kyrgyz lands, while other studies like Liu et al., [8] include climate change scenarios. The general methodology of correlating land cover with other spatial variables however is considered transferable to any other region.

2.2 Objectives

Analyzing the distribution of land cover and its correlation with topographic parameters at a ground resolution of 25m pixels corresponds to the original spatial resolution of imagery as well as digital elevation models (DEM), and at the same time demonstrates the power and potential of extending this analysis over huge swaths of land and potentially the entire globe, instead of limited local study areas. Within this meta-objective of assessing and validating the global scope and analytical efficiency of creating geospatial insights from a combination of cloud-based data sets with fully scalable methods and tools, this case study explores the topographic patterns of land cover across the Kyrgyz Republic.

Figure 1: Topographic context of Kyrgyzstan as a study area

For the first time not only the balance of land cover categories for this country is established, but also LC distribution over topographic parameters. These examples are intended to showcase the potentials arising from combining, aggregating, and analyzing easily accessible global data sets with cloud-based geospatial methods.

2.3 Data Bases

This study is focused on (nearly) globally available online data sets accessible as online services and available for analysis. These products became feasible with the advent of sensors like MODIS, AVHRR and ETM which for the first time enabled the building of a ‘global view’ of the state and condition of Earth’s surface. The increase in open access to remote sensing date up to dekameter range created a decisive impulse for the move from project-based and regional needs-driven LULC classification towards global supply-oriented coverages. The ‘Global Land Survey’ [9] using Landsat ETM+ (after the full opening of Landsat archives, and based on earlier precursors) set the stage for the emergence of freely accessible worldwide LC data. It is important to mention, however, that automatic classification of remote sensing data is not the only path to land cover data. The European CORINE program initiated in 1985 started out with manual classifications of Landsat imagery into 44 categories with an original minimum mapping unit of 25ha and a focus on environmental monitoring [9]. Subsequently this was transitioned into the Sentinel – focused Copernicus Land Monitoring Service.

The geospatial cloud analytics case study presented in this data set is predominantly based on two global data sets, therefore the general approach would be applicable and reproducible essentially anywhere. These data sets represent land cover and elevation, respectively, and as due to their worldwide availability (except for polar regions) provide a seamless and complete coverage of Kyrgyzstan. The ‘World Land Cover 30m BaseVue’ data set [10], see Table 1, is a commercial product using a multitemporal, semi-automated supervised classifier (https://www.maxar.com/products/basevue-lulc). The original capture dates were April 2014 to June 2014 with Landsat 8 with continuous updates until August 2020 for the data set used by the authors at https://arcg.is/Py4ei. Available as a premium service on the Living Atlas cloud platform (https://livingatlas.arcgis.com) like the other products discussed in this section it has a proven record of supporting analyses anywhere on the globe in the 25/30m resolution dimension. These LC data are accessible as an image service with query, identify, export and raster function capabilities and contain cell values according to the following class definitions.

Table 1: Land cover class definitions according to [10]

Figure 2: BaseVue land cover service with enlarged sample detail

Figure 3: Land Cover categories within Kyrgyzstan - dominated by barren land, shrubs and grassland, with minor proportions of forests, agricultural land and water/snow/ice

For the presentation of analyses highlighted in this paper the classes 7 and 8 have been consolidated under ‘Agriculture’, the minimal amount of ‘Wetlands’ has been added to ‘Water’ and 20 and 21 were combined into ‘Urban’. Classes 10, 13, 14, and 15 were not present within the study area. Figure 2 shows the original view and standard symbology zoomed into the study area with a local more detailed sample around the Kyrgyz capital city Bishkek.

Statistically aggregating the LC classes on the BaseVue service within Kyrgyzstan’s boundaries shows more than 80% of the land mass covered by only three approximately equal categories, barren land, shrub / scrub and grassland. Only 6% are available for agriculture, and area exceeded by water / ice / snow. Forest cover is minimal, and built-up land is concentrated in a few major settlements.

While this kind of summary statistics is easily tabulated, for further analysis and the development of change scenarios it is important to allocate LC categories to specific topographic, climatic and human access contexts – this question will be addressed further below. The map in Figure 2 shows a lot of detail and serves as a starting point for high resolution multi-thematic analyses, but does not really provide a quick and crisp overview of spatial LC patterns across the country. For more effective communication of the ‘big picture’ a generalized view was created with a hexagonal standard mapping unit of 10km². Within each hexagon, the dominant ‘majority’ class was selected by zonal statistics and assigned a unique value according to the legend provided. While BaseVue 2013 has been used exclusively in the analysis presented below, multiple alternative data sets now are available with somewhat different characteristics and are referenced here for context: most recently, a ten class global land use/land cover (LULC) data set based on Sentinel 2 for the year 2020 has been generated at 10 meter resolution [11]. It distinguishes classes water – trees – grass – flooded vegetation – drops – scrub/shrub – built area – bare ground – snow/ice – clouds and claims an overall accuracy of 86% against a validation set [12]. It was produced by a deep learning model trained using over 5 billion hand-labeled Sentinel-2 pixels, sampled from over 20,000 sites distributed across all major biomes of the world. The underlying deep learning model uses 6 bands of Sentinel-2 surface reflectance data: visible blue, green, red, near infrared, and two shortwave infrared bands. To create the final map, the model is run on multiple dates of imagery throughout the year, and the outputs are composited into a final representative map of 2020 (https://arcg.is/0Kemuy). In addition, the processing workflow has been established in a way to allow future global as well as regional replication based on updated sensor data, thus facilitating updating and change detection. Acknowledging that the quest for global LC data sets is already dating back at least three decades, the original GLC can be considered a baseline data set which only much more recently has been made accessible as an online service. The table below provides a quick look at the evolution of monitoring global land cover since then (Table 2).

Figure 4: 10km² hexagons assigned the majority class within each hexagon

Table 2 : Overview of select global land cover data sets

The below presented exemplary analyses, however, are not predominantly focused on change monitoring and detection, even though this has been the main objective for the creation of the above introduced coverages. The authors rather want to emphasize the potential for analysis across thematic domains, like:

- LC distribution and dynamics across different morphometric features,

- LC relative to population densities and human activities,

- LC differences within and between national jurisdictions.

These kinds of cross-domain analyses are facilitated by global data sets similarly accessible like the above introduced LC services, in particular the Airbus WorldDEM used in subsequent analyses (https://www.intelligence-airbusds.com/imagery/reference-layers/worlddem/). To work with topography across the study area in the following analyses, the global multi-resolution terrain elevation service from the Living Atlas was accessed at https://arcg.is/feefK at the (25m) resolution of the WorldDEM data.

2.4 Data Analysis

Multi-thematic analysis long has been a mainstay set of methods within the field of spatial analysis. Exploring correlations and systematic interdependencies between coverages of spatial data is an important starting point for understanding spatial distributions of observations – like land cover. Having a clear picture of the impact of independent spatial variables like terrain, zonal climate factors and human action on land use and land cover ultimately helps with modeling distributions [13] as well as developing scenarios for anticipating change. Applicable overlay methods depend on the type of data involved. Metric data sets, like elevation vs precipitation lend themselves to simple correlation quantifying a degree of interdependence – although due to the typical presence of a high degree of spatial autocorrelation inferential statistics have limited value. At the core of exploratory analyses in this paper is the traditional map algebra approach of zonal analysis, generating descriptive statistics of a metric, continuous spatial variable like terrain elevation within the spatially discrete zonal categories of e.g. land cover.

Outlined already in the foundational book by Tomlin [14], zonal analysis per se obviously does not qualify as a novel approach in geospatial analysis. Within the context of this paper however it is applied to demonstrate the immense added value derived from two recent developments: the services-based open access to global data sets with high spatial and temporal resolution, and the emancipation of processing frameworks from personal workstations towards cloud computing. The combination of these developments significantly lowers the hurdles for exploratory analysis of land cover distributions and dynamics on e.g. national scales without having to go through the previously required enormous efforts of data preparation and staging of analyses. This evolution of analytical frameworks is tightly connected with the establishment of Spatial Data Infrastructures [15] [16] and Digital Earth twins [17], as demonstrated e.g. through the European INSPIRE Spatial Data Infrastructure presented by Minghini et al [2].

A spatial resolution of 25m was explicitly defined for all subsequent analysis steps, fully leveraging the nominal 25m resolution for terrain elevation and reasonably close to the 30m original Landsat ETM resolution underlying the BaseVue land cover service. Values of derivatives like slope therefore must be considered within the constraints of this resolution.

3.Results

To demonstrate the power and potential of cloud-based multi-thematic analysis over large regions covering entire countries, the BaseVue 2013 land cover data set zones were analyzed against topographic variables at a 25m spatial resolution. For full appreciation of the data volume involved and the scale of these cross-tabulations of categorized data through zonal operations it shall be kept in mind that each layer (land cover, elevation and slope) include approx. 320 million data points each. The entire workflow has been implemented in the Esri ‘ecosystem’, leveraging data from the Living Atlas infrastructure, using ArcGIS Pro for analytical steps and ArcGIS Online for presentation including storymapping.

>3.1 Hypsometric Distribution of Land Cover Categories

The summary table presented as Figure 5 summarizes the zonal analysis of BaseVue LC classes against WorldDEM elevation, aggregated into 100m elevation steps and cut off above 5000m. The column labeled ‘HYPSOMET’ shows the rather unusual hypsometric curve of this country. A peak around 1600m elevation highlights the huge lake area of ‘Issyk Kul’ in the northeastern region, and the dominant elevations between 3000m and 3500m are characteristic for the large tracts of land with only very limited economic potential as seasonal pastures.

Figure 5: Hypsometric distribution of land cover categories across Kyrgyzstan

The coloured graphic bars within the chart represent the relative per-LC-class elevation distribution and have to be interpreted with reference to the total share of these classes as indicated in Figure 3. Agricultural uses and human settlements dominate only below 1500m elevation, from there a few forests and a lot of grassland and shrub in dryer areas take over with barren land dominating above 3500m followed by nival high mountain environments. This overview could of course be translated back into a multivariate selection map (not shown here for space reasons) highlighting e.g. low lying semi arid and currently unused areas which potentially could be made available for agriculture through irrigation. Or those shrublands can be identified offering opportunities for afforestation, or for conversion into productive grasslands. Clearly these would be naïve approaches requiring finetuning with additional factors and constraints to deliver any kind of useful policy recommendations. Climate variables, soils, feasibility of irrigation, access to markets and other criteria would have to be included with only slope inclination addressed as an additional factor below – as the main thrust of this paper aims at demonstrating the application of regional analysis through cloud computing and leveraging of online LC services, not pursuing specific development and policy questions.

3.2 Slope Patterns of Land Cover Categories

Topographic gradients, commonly referred to as ‘slope’ are another important constraint for the development of land uses and the assessment of regional potentials. In Figure 6 again the entire national land mass was classified in 5° slope categories using the 25m WorldDEM with direct use of server-side processing. The frequency distribution including sizable proportions of rather steep slopes – nearly half the country is in the steeper than 20° bracket. Again the slope frequencies in the respective columns have to be interpreted relative to those LC categories as quantified in Figure 3. Besides the obvious flat area and gentle slope preferences of settlements and agriculture (and of course water bodies) grassland due to its use as pastures dominates hilly slopes. Shrub and barren land depends less on slope than on substrate and humidity and therefore does not exhibit a clear slope frequency profile, with forests remaining on steeper slopes while ice and snow of course are determined mostly by elevation. Again, slope serves as but one factor in a multi-thematic land cover analysis and demonstrates the enormous benefits derived from openly and readily accessible global (digital terrain) data sets for a broad range of analyses.

Figure 6: Distribution of land cover categories across 5° slope classes. Second column from left represents overall slope frequency within the country

4. Discussion and Conclusions