Monitoring Land Use and Land Use Change (LULUC) through remote sensing is a common approach to generating necessary data for quantifying anthropogenic impacts on the Earth’s system. Land monitoring through remote sensing has traditionally been challenging due to the cost of acquiring satellite imagery and commercial software to conduct remote sensing analysis [1
] and due to the high level of technical skill required to pre-process and analyze imagery and conduct a robust land assessment [2
Remote sensing data are now used in many national and international land assessments, such as national forest inventories (e.g., France, Italy, Switzerland, USA) and the European Land Use and Land Cover Survey (LUCAS) [2
]. These assessments have followed a multi-phase sampling approach, in which the first phase involves visual interpretation of Very High Resolution (VHR) satellite imagery, and the second phase is devoted to collecting ground-based data in the field. The use of remote sensing data during the first phase enables national and international experts to efficiently assess a large number of sites and quantify the area of land currently and historically allocated to broad land use and land cover categories. Subsequently, ground-based data gathered during the more time-demanding second phase allow experts to develop a more detailed understanding of land characteristics and variability within broad land use and land cover categories. The data resulting from the two phases are synergistic, as Phase 2 data can be used to estimate uncertainties within the spatial extent and area estimation of land use and land cover categories, while the latter can be used to extrapolate more detailed land characteristics (e.g., vegetation types, carbon stocks, etc.) from a relatively small number of field sites to the landscape level drawing upon the much larger number of sites assessed in the first phase. This method has also been adopted by countries to quantify their LULUC with low uncertainty and to address their need to report to the United Nations Framework Convention on Climate Change (UNFCCC) [4
The field of remote sensing has undergone significant changes over the past decade that have helped make land monitoring more cost efficient and technically feasible for non-remote sensing experts. The cost of many type of satellite imagery has decreased; the free accessibility of imagery has increased [6
]; and several non-commercial software packages have been developed to facilitate the analysis of imagery. The United States Geological Survey (USGS) and Google are two institutions at the forefront of these developments.
The USGS has spearheaded significant changes through its adoption of an open data policy with the Landsat imagery archive in 2008. Landsat is the world’s largest and longest running archive of satellite imagery, with imagery acquisition dating back to 1972. Its global coverage and frequent imagery acquisition (16-day revisit time) make the archive highly suitable for supporting land change monitoring. Prior to 2008, the cost of Landsat imagery ranged from USD20–4000 depending on the image format, the sensor and the imagery acquisition date. Four years after making the archive freely available, the number of image downloads each month rose from under 3000 to over 250,000, exponentially increasing the use of satellite imagery for land monitoring [10
Google released Google Earth in 2005, a virtual globe desktop software, enabling users to freely view medium, high and very high spatial resolution satellite imagery. The best imagery available for each site is automatically loaded in Google Earth’s user-friendly software that allows users to zoom into any place on Earth while the software seamlessly manages ten petabytes of geographic information [11
]. Google Earth software offers users a level of vertical integration that was not previously available within the field of remote sensing. In the past, companies that developed image processing software were distinct from those that launched satellites and acquired satellite imagery, while the service providers with the technical expertise that used the software and imagery for remote sensing analysis often represented a third company. Thus, an individual interested in LULUC monitoring often needed to consult at least three different companies to achieve one’s Earth observation objectives. Google changed this paradigm by streamlining imagery acquisition and arduous processing (such as geo-referencing and mosaicking) to make satellite imagery ready for visual interpretation within a simple application. Google Earth has enabled many users wishing to monitor land (e.g., foresters, conservationists, indigenous leaders, etc.) to bypass the need for remote sensing experts to monitor their land. By 2011, Google Earth had been downloaded over one billion times while global Internet usage had reached around 2.2 billion worldwide, making Google Earth the world’s most popular geospatial application [12
Numerous free and open source software platforms for land monitoring have been developed by building upon Google Earth and its freely accessible archive of satellite imagery. Examples of software facilitating the analysis of land use, land cover and other land characteristics include: (1) Geo-Wiki; (2) GLCF Labeling Tool; (3) LACO-Wiki; (4) SkyTruth; (5) TimeSync; (6) Tomnod and (7) VIEW-IT. An overview of the scope and functionality of the different software is provided in Table 1
Geo-Wiki, GLCF Labeling Tool, LACO-Wiki, TimeSync and VIEW-IT are designed to facilitate visual interpretation of land cover and/or land use data primarily for the purpose of map validation at any scale, local to global, but these tools can be used for other purposes, as well [14
]. In contrast, Skytruth and Tomnod were developed to collect very specific types of information, such as the spatial extent of land degradation due to mountain-top removal mining or natural disasters [19
]. Many of these free and open source tools (e.g., Geo-Wiki, VIEW-IT, Skytruth and Tomnod) have been developed by non-governmental organizations or academic institutions to enable the crowdsourcing of volunteered geographic information (VGI) by non-remote sensing experts [8
]. However, GLCF Labeling Tool, LACO-Wiki and TimeSync were designed mainly for remote sensing experts [16
All of these software packages draw upon one or two archives of satellite imagery: one with VHR imagery (in Google Maps or DigitalGlobe’s web mapping interface) and/or one archive of medium-(Landsat) or coarse-resolution (MODIS) imagery. VIEW-IT, GLCF Labeling Tool and TimeSync also display automatically generated time series of vegetation indices and the results of other image processing algorithms.
Despite the aforementioned developments, significant challenges remain for monitoring LULUC through remote sensing. Google Maps and Google Earth and the software mentioned above facilitate visual interpretation through VHR satellite imagery (one meter or less). However, the small geographic scope and the irregular time intervals of VHR imagery acquisition limit its use for national and subnational assessments for land monitoring (including LULUC), particularly in areas with persistent cloud cover, phenological changes or rapid change. While the Landsat archive may excel in these aspects with its global geographic coverage and bi-monthly imagery acquisition, its relatively low spatial resolution (30 m) makes it difficult or impossible to identify small features in imagery and small-scale changes within landscapes that may be of interest to land monitors. Despite these limitations that impact visual interpretation of the imagery, Landsat can be extremely useful for (semi-)automated imagery analysis.
Furthermore, the application of semi-automated classification methods (e.g., pixel-based, object-based) on VHR images to develop national, regional or global maps has proven to be challenging for a number of reasons. These challenges include: (i) the high cost associated with VHR imagery; (ii) their low spatial extent (a few hundreds of km2
]; (iii) their relatively low availability due to their low temporal resolution and lack of global coverage [21
]; (iv) the variation of radiometric properties among sensors; (v) the influence of acquisition conditions (i.e., Sun-scene-sensor angles) [22
] and (vi) classic atmospheric perturbations (e.g., cloud, fires) [21
]. All of these challenges of land use monitoring are exacerbated when assessing land use change and attempting to acquire imagery from multiple years with acquisition conditions that are sufficiently consistent to enable the comparison of land characteristics.
Here, we present Collect Earth, a free and open source software developed by the Food and Agriculture Organization of the United Nations (FAO) to facilitate the collection, management and analysis of land data. Like its predecessors, Collect Earth also enables expert and non-expert users to draw upon Google technology to freely access and visually interpret satellite imagery for data collection. Collect Earth geo-synchronizes the visualization and use of imagery of varying spatial and temporal resolutions, including DigitalGlobe, SPOT, Sentinel 2, Landsat and MODIS imagery within Google Earth, Bing Maps and Google Earth Engine [26
]. Collect Earth differs from previously existing land monitoring tools by offering access to: (a) multiple archives of VHR satellite imagery that can support the assessment of land use and land cover dynamics; (b) graphical representations of inter-annual and intra-annual vegetation indices generated with Landsat and MODIS imagery in Google Earth Engine (GEE), new technology for cloud-based, automated processing of satellite imagery; and (c) built-in data analysis tools through an integration with Saiku Analytics. Collect Earth also differs from previous land monitoring software in that (d) it offers a robust data collection framework that is fully customizable by non-experts; and (e) it streamlines the use of probability sampling statistics. Collect Earth accesses three archives of satellite imagery that have an expansive coverage and collectively enable users to assess any area in the world. However, where supplementary VHR imagery has been acquired, such imagery can be imported into Google Earth (Pro) in numerous formats and immediately used for a land assessment with Collect Earth.
In remote sensing, the terms visual interpretation and photo interpretation refer to human interpretation of two-dimensional images to use visual elements, such as tone, shape, pattern, texture and shadow, to identify objects within satellite or aerial imagery [29
]. While previously available free and open source land assessment software packages facilitate basic visual interpretation, Collect Earth draws upon Google Earth, GEE and Bing Maps to enable land assessment through augmented visual interpretation. Images from multiple years are supplemented by seasonal and multi-year graphs of several indices that are automatically generated by scripts within the GEE Code Editor (e.g., Landsat 8 32-day Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), Enhanced Vegetation Index (EVI), Moderate Resolution Imaging Spectroradiometer (MODIS) 16-DAY NDVI and Landsat 7 Monthly NDVI Composite). The visual interpretation of these indices in GEE Code Editor, in conjunction with VHR and HR satellite imagery, enable users to assess current land attributes and trends over the past 16 years more comprehensively than otherwise possible through basic visual interpretation.
Landsat 7 and MODIS, which are used to generate the vegetation index time series, were both launched in 1999. To support users interested in conducting a longer-term land assessment, Collect Earth also has functionality to indicate the presence of Landsat 5 Annual Imagery. Landsat 5 is the world’s longest-operating Earth observation satellite; however, its images are neither spatially contiguous nor available in the Earth Engine archive at regular time intervals. For each site assessed, Collect Earth generates a graphic within Earth Engine showing the number of Landsat 5 images available in the archive and their acquisition date to guide users seeking imagery dating back to 1984.
Another key difference in using augmented visual interpretation and a probabilistic sampling design to assess the land use is that it does not involve nor require any modelling or extrapolation to produce the final results. This is substantially different from common remote sensing techniques, which generally assess land through the extrapolation of a subset of training sites over an entire study area [30
]. Local visual interpretation of VHR imagery is often used for training and/or validation purposes of extrapolated maps of land cover or tree cover [7
]. In essence, Collect Earth combines these two steps by exclusively using data previously reserved for training/validation to directly produce results at a national, regional or global scale.
This paper presents a detailed overview of Collect Earth’s structure and functionality. To illustrate the application of the tool and a possible customization, an example of land monitoring in Papua New Guinea is presented in Section 3
. The advantages and potential limitations of using Collect Earth for these and other assessments are subsequently discussed.
3. Application Example
This section is based on work undertaken by the Papua New Guinea Forest Authority (PNGFA) in the context of its preliminary national assessment of LULUCF with Collect Earth. The assessment was conducted between October 2013 and May 2014.
More than twenty-five government officers participated in the assessment, all of whom had extensive local knowledge and professional backgrounds in forestry and silviculture, but little or no prior GIS or remote sensing experience. Although other LULUCF assessments in Papua New Guinea have relied upon a smaller number of individuals with established GIS skills, the PNGFA team found that the use of Collect Earth software facilitated the involvement of officers with other professional backgrounds. After five days of training on the use of Collect Earth, the officers worked for varying lengths of time ranging from 5 to 64 days and completed the majority of the data collection in a four-month period. The average number of plots assessed per day by each officer working full-time (at least seven hours) was 46, with individual averages ranging from 13 to 91 plots. After several days of experience working with the Collect Earth, several officers were able to assess over 100 plots in one day.
PNGFA’s main limitation during this first assessment with Collect Earth was the Internet speed, as this directly impacts the amount of time required for imagery in Google Earth, Bing Maps and Earth Engine to load before image interpretation can occur. Standardizing land use interpretation among a large number of officers was also a significant challenge. PNGFA has subsequently addressed these issues by strengthening the Internet bandwidth and by adjusting their methodology and work space to facilitate collaboration during Collect Earth assessments.
In the 2013/2014 assessment presented in this application example, PNGFA officers recorded 45 different attributes for each of the 25,279 stratified systematic sampling plots assessed in Papua New Guinea. Land use change data were collected within a reference period from the early 1970s–2013. Land use changes between 2001 and 2013 were assessed with Collect Earth, while the assessment of land use changes prior to 2001 was based on local knowledge. For the purposes of this application example, we use a subset of the data that focuses on only five attributes and 2240 plots within three provinces: Milne Bay, North Solomon and West New Britain. The data presented here cannot be considered as final results and are solely used for illustrative purposes.
3.1. Data Collection Form
PNGFA’s data collection form was customized to gather information in a manner consistent with IPCC guidelines, thus enabling PNGFA to use the resulting data to address some of its data needs for reporting to the United Nations Framework Convention on Climate Change [44
]. Figure 2
shows a sample data collection form used in Papua New Guinea for gathering land use and land use change data, as well as land cover information.
3.2. Sampling Design and Project Properties
The sampling design for the Papua New Guinea assessment was generated in QGIS with a probabilistic sampling design, enabling area estimations. In Papua New Guinea’s stratified-systematic sampling design, smaller provinces measuring less than 5000 km2 (with the exception of National Capital District) have more intensive sampling, with four-times as many plots per square kilometer as the country’s larger regions. The plot size is consistent with the minimum mapping area required to apply Papua New Guinea’s national forest definition.
The Papua New Guinea LULUCF assessment with Collect Earth uses square-shaped plots of one hectare containing 25 sampling points (Figure 3
). Each sampling point therefore corresponds to approximately 4% of the plot. Data are analyzed at the plot level, while the sampling points within the plot can be used to quantify and characterize land cover within the plot. For example, canopy cover percentage within the plot can be measured to apply the canopy cover threshold of the national forest definition.
provides an example of the sampling design in West New Britain province of PNG (Figure 3a), the project properties (Figure 3b), as well as a picture of the plot area from Google Panoramio [45
]. Google Panoramio pictures are available to users free of charge through Google Earth and allow users to exploit an additional visualization tool to assess land.
3.3. Augmented Visual Interpretation
presents an illustration of an LULUCF assessment based on the IPCC Guidelines for National Greenhouse Gas Inventories [40
]. In Google Earth, the plot falls in an oil palm plantation in February 2011. The current land use is therefore ‘cropland’. Using the historic VHR imagery in Google Earth (without accessing any of the additional sources that Collect Earth provides), the user can observe that the land use of the same plot was ‘forest’ in March 2001, but partially cleared sometime before September 2001; by June 2003, the land use of the plot has changed to ‘cropland’. Thus, a forest-to-cropland land use change has occurred.
After this initial assessment in Google Earth, the user can begin evaluating the information in the additional archives Collect Earth accesses. Figure 5
presents the geo-synchronized view of the plot at different in scales in Bing Maps (Figure 5
a), GEE (Figure 5
b) and GEE Code Editor (Figure 5
The user can observe that the image in Bing Maps (dating June 2011) corroborates the observations made in Google Earth.
Loading the Landsat 7 Annual Greenest-Pixel Composite for 2014 in GEE (Figure 5
b), we can observe that the plot is at the edge of a cropland plantation and forest and that the current land use for 2014 is indeed ‘cropland’.
In an assessment like PNGFA’s, the inter-annual vegetation indices in the GEE Code Editor (Figure 5
c) can guide users to hone in on periods of significant change, as opposed to loading and reviewing imagery from every year of the reference period. Forest clearings can easily be detected, as well as harvests within cropland.
3.4. Data Analysis and Visualization Using the Built-In Saiku Analytics
presents examples of analytical queries in Saiku for (a) current land use composition; (b) land use change; (c) current land use of historic forest area; and (d) current cropland subdivisions within historic forest area in three provinces of Papua New Guinea. The bar graph, table and pie chart in Figure 6
were generated within Saiku.
a illustrates that land categorized as ‘forest’ according to the IPCC land representation framework occupies the largest area in the three provinces.
An overview of land use conversions from the early 1970s–2013 is presented in Figure 6
b as a land use change matrix. Initial IPCC land use categories are listed in the columns, while current land use categories are presented in rows. For example, of the 962 initial forest plots in West New Britain, 880 plots remained as forest, while 82 plots were converted to cropland by 2013. The land use conversion presented in Figure 6
is an example of one such forest-to-cropland conversion in West New Britain. Approximately three quarters of forest-to-cropland conversions in the province occurred within lands currently categorized as oil palm plantations. In all three provinces, the conversion of forest to other land uses constitutes a substantial portion of all land use changes. Countries can use this type of land use change matrix for international reporting.
c focuses on the first column of the West New Britain land use change matrix, presenting the current land use composition of 2001 forest lands in pie chart format. For example, the plots in West New Britain that were converted from forest to cropland are shown in light orange. The vast majority of West New Britain’s 2001 Forest lands remained forest in 2013. In terms of forest changes to other land uses, Milne Bay, West New Britain and North Solomons present three different profiles: low-, medium- and high-level forest change.
Most forest changes that are observed in this illustrated example are forest-to-cropland conversions. Figure 6
d presents these land use changes in greater detail by looking at the land use subdivisions (which were developed by the PNGFA). In this example, it provides, where possible, the type of agriculture ‘forest lands’ have been replaced by and their relative proportion. The land use composition and the land use subdivisions of an area are critical for establishing a baseline against which future changes can be assessed. Information on historical land use subdivision, such as that provided in Figure 6
d, can also be useful for understanding the drivers of deforestation or other forms of land use change.
Large-scale land assessments have often been conducted with several approaches: compiling national maps [46
], using sampling-based approaches [47
] or developing exhaustive maps from remote sensing products [9
]. Mayaux et al. (2005) [49
] underlined that sampling approaches may yield more accurate results in the case of forest distribution if based on large numbers of small units. However, existing sampling-based products [46
] are based on a small number of large units extracted from Landsat scenes. Consequently, recent efforts were more focused on the improvement of spatially-exhaustive remote-sensing products, where experts use commercial, specialized GIS or image analysis software that run semi-automated algorithms for detecting and categorizing changes in spectral signatures in the satellite imagery of a landscape [9
Commonly-used semi-automated approaches to land use and land cover (LULC) assessment may include key steps such as: (1) the selection of remotely-sensed data; (2) the determination of a suitable classification system; (3) the selection of training samples; (4) image pre-processing; (5) the selection of a suitable classification approach; (6) image segmentation and feature extraction; (7) post-classification processing; and (8) accuracy assessment [51
]. Such LULC assessments can incur expenses throughout this process for acquiring satellite imagery and commercial GIS or image analysis software and also for human resources to contract remote sensing experts to conduct image pre-processing and analysis.
Google Earth, Google Earth Engine and Bing Maps, as well as the existing land assessment and map validation tools listed in Table 1
can be used individually to support one or several steps outlined above for semi-automated image analysis. However, Collect Earth is currently the only tool that can simultaneously access all three of the aforementioned archives, thus enabling users to draw upon the strengths of each, while also reducing the cost and time devoted to image acquisition (Step 1). As spatial and temporal gaps in freely available VHR imagery are inevitable, Collect Earth’s integration with GEE helps users to easily fill these gaps with information from coarser resolution Sentinel 2, Landsat and MODIS imagery.
Collect Earth’s integration with multiple archives of VHR imagery can facilitate the generation of data for training samples (Step 3) and accuracy assessment (Step 8), reducing the potential expenditure by $16–$25 USD/km2
for imagery ranging from 0.6 to 4-m resolution [52
]. The information gathered with VHR imagery is extremely important for the typical LULC assessment that uses semi-automated algorithms because a relatively small amount of information from sites directly observed is used to train an algorithm that will classify vast areas and a relatively large number of sites that have not been directly observed.
In contrast, Collect Earth provides a framework for users to go beyond this more limited use of visual interpretation. Through augmented visual interpretation with Collect Earth, users can simultaneously analyze imagery of multiple scales and base their entire assessment on the same activity that often underpins only training and the accuracy assessment portions of LULC studies. Thus, while only 5%–10% of an area might be directly assessed in a typical LULC assessment with semi-automatic algorithms, Collect Earth and the input sampling design guide users to assess 100% of the sites that are used in the calculation of statistics regarding land use, land cover and land dynamics. This allows users to avoid classical uncertainties and biases related to extrapolations of mapping-based methods.
When conducting land use or land cover change analysis with high resolution imagery, the cost of analysis can be $160–$250 USD/km2
, ten-times greater than the cost of the imagery because of the large amount of image pre-processing (Step 4) required [52
]. Pre-processing can include geometric rectification, radiometric calibration, atmospheric correction and topographic correction [51
]. Some of these tasks are conducted by Google and DigitalGlobe when they add new images to their archives. Collect Earth and augmented visual interpretation enable users to skip pre-processes and reduce expenses associated with this step by accessing pre-processed imagery and by facilitating a methodology that can easily be applied by non-remote sensing experts with minimal pre-processing.
Lastly, conducting an LULC assessment with using Collect Earth or other free software, including those listed in Table 1
, can reduce expenditure on commercial software licenses.
The application of augmented visual interpretation with Collect Earth for LULUCF assessment in Papua New Guinea is one of numerous potential applications of the software (Table 2
). Regardless of whether Collect Earth is used to conduct a base assessment, to facilitate on-going monitoring, to gather information to train land assessment algorithms or to generate data to assess the accuracy of existing maps, the software’s user-friendly interface and robust framework can broaden the engagement of individuals with little or no prior GIS and remote sensing experience. In conjunction with Google Earth, Bing Maps, Google Earth Engine and Saiku, Collect Earth can facilitate the assessment of land use, land cover and land dynamics by government agencies, non-for-profit organizations, academic institutions, field experts or other individuals.
Nevertheless, Collect Earth also has some limitations. The accuracy and robustness of an assessment depend heavily on applying an appropriate sampling design and sampling intensity to adequately capture the variability of the land characteristics being assessed. Furthermore, the point-sampling methodology is a non-exhaustive spatial cover (depending on the sampling intensity), thus limiting the full variability of the land that can be classified and measured.
The augmented visual interpretation approach presented here using Collect Earth is currently based on optical data. Although the powerful combination of sources of information that Collect Earth make available for a land monitoring assessment (including the Landsat Greenest pixel products) usually provides cloud-free information, occasional (partial or full) cloud cover over a sampling plot remains a challenge.
Another limitation is that the Internet is required to access the imagery and satellite-derived data that are used by Collect Earth. In cases where only Google Earth imagery is required, a low or medium speed Internet connection can be adequate. However, to quickly visualize and process years of MODIS, Landsat and Sentinel 2 imagery, high speed Internet is necessary.
Finally, when multiple users are working on the same assessment, it is crucial that a clear methodological framework for augmented visual interpretation is established to ensure consistency.
With nearly three quarters of the Earth’s surface impacted by human activity, it is more important than ever that countries, organizations, communities and individuals are cognizant of current, past and future land characteristics. Improved land monitoring by both remote sensing and non-remote sensing experts through augmented visual interpretation can enable a broader array of actors to take an active role in monitoring lands currently impacted by human activities. The application of Collect Earth in Papua New Guinea illustrates how the software can be used at the national and subnational level. A recent assessment of trees, forests and land use in drylands has demonstrated that Collect Earth can also be used to conduct rapid land assessments at the global level [53
]. Collect Earth is not only a tool for land monitoring, but it can also support land use planning, management, transparency and accountability at multiple scales.
Collect Earth makes a substantial contribution to a significant trend that has been observed in the field of remote sensing over the past ten years by improving access to freely available satellite imagery and making imagery analysis more accessible to non-remote sensing experts. By altering the inputs of Collect Earth, such as the data collection form, sampling design and plot size, users can easily configure Collect Earth to address specific land monitoring purposes, such as landscape restoration, reporting for REDD+, national forest inventories, disaster assessments and humanitarian work, livestock and rangeland management, etc. (Table 2
), with a multi-temporal and multi-scale approach.
The most significant innovation of Collect Earth is that it enables anyone to conduct a robust land assessment of any area of the world using free and open source tools, VHR satellite imagery freely accessible online and augmented visual interpretation.
For users who wish to learn more about Collect Earth or use the software for a land assessment, the following supplementary materials are available: 1. Collect Earth User Manual (version 1) [39
]; 2. Collect Earth installation file for Windows operating systems [54
]; 3. Collect Earth installation file for Mac operating systems [54
]; and 4. Collect Earth—Papua New Guinea customization (CEP) file [54