Next Article in Journal
A Geo-Tagged COVID-19 Twitter Dataset for 10 North American Metropolitan Areas over a 255-Day Period
Previous Article in Journal
Measurements of LoRaWAN Technology in Urban Scenarios: A Data Descriptor
 
 
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Data Descriptor

A Disease Control-Oriented Land Cover Land Use Map for Myanmar

1
Department of Geographical Sciences, University of Maryland, College Park, MD 20742, USA
2
Department of Earth System Science, Tsinghua University, Beijing 100084, China
3
Department of Geosciences, Texas Tech University, Lubbock, TX 79409, USA
*
Author to whom correspondence should be addressed.
Submission received: 12 May 2021 / Revised: 8 June 2021 / Accepted: 11 June 2021 / Published: 13 June 2021
(This article belongs to the Section Spatial Data Science and Digital Earth)

Abstract

:
Malaria is a serious infectious disease that leads to massive casualties globally. Myanmar is a key battleground for the global fight against malaria because it is where the emergence of drug-resistant malaria parasites has been documented. Controlling the spread of malaria in Myanmar thus carries global significance, because the failure to do so would lead to devastating consequences in vast areas where malaria is prevalent in tropical/subtropical regions around the world. Thanks to its wide and consistent spatial coverage, remote sensing has become increasingly used in the public health domain. Specifically, remote sensing-based land cover/land use (LCLU) maps present a powerful tool that provides critical information on population distribution and on the potential human-vector interactions interfaces on a large spatial scale. Here, we present a 30-meter LCLU map that was created specifically for the malaria control and eradication efforts in Myanmar. This bottom-up approach can be modified and customized to other vector-borne infectious diseases in Myanmar or other Southeastern Asian countries.
Dataset License: CC-BY-4.0

1. Summary

Accurate accounts of land cover/land use (LCLU) in the form of LCLU maps are critical sources of information that allow stakeholders to make informed decisions. Many LCLU datasets, such as the Moderate Resolution Imaging Spectroradiometer (MODIS) MCD12 Land Cover product [1] and the National Land Cover Database (NLCD) [2], are geared towards serving general needs (i.e., land planning, forest management, etc.) and are generally created based on global algorithms (i.e., algorithms that are consistently applied across the areas of interest). However, there are situations where specific LCLU classification schemes are required and these circumstances are generally associated with the efforts that are implemented on smaller spatial scales. As a result of such specific needs, the LCLU maps produced using global top-down algorithms are likely to be insufficient for use locally. Recently, publicly available datasets and algorithms designated to specific LCLU classes, such as croplands and water bodies, have been increasingly offered with a high readiness for scientific usage. As these datasets and algorithms are designed for specific purposes and are often associated with uncertainty metrics (allowing the mapping results to be adjusted locally), they are likely to be more reliable in representing the LCLU classes of interest than the datasets that are derived using top-down methods. Leveraging such datasets and algorithms that utilize local knowledge about the issues in question could be the key to the development of bottom-up LCLU maps that serve various special and localized needs.
LCLU patterns have been shown to be important for disease epidemiology and can be conveniently determined using remote sensing data. Land cover determines habitat suitability for hosts and vectors [3] and remote sensing data are commonly used to determine vector/host animal habitats. Disease risk variation, however, is not only determined by the land cover but also by the land use, which helps to understand the places that people visit and use for certain activities or during certain times of the day or year [4], and contributes to the determination of human distribution and the contact between humans and vectors [3]. LCLU maps and landscape metrics that have been derived using remote sensing have been used to understand the patterns and transmission of several diseases, including malaria, Lyme disease, cholera, dengue, Zika, etc. [5,6,7,8,9,10]. Most of such studies have used the mapped vegetation derived from remote sensing to understand the factors associated with vector presence and modeled human transmission risk, relying on the fact that most vector-borne disease transmission is dependent partly on vegetation characteristics [5]. For example, some studies have used LCLU maps at multiple spatial resolutions to predict the disease transmission risk or to study the role of LCLU in disease transmission based on the links between vector habitat preferences, LCLU classes, and disease incidence [6,10]. In contract, others have associated LCLU based landscape metrics such as edge density, patch density, and proportion of land use with disease risk [9].
The Southeastern Asian country of Myanmar, home to a population of more than 53 million people, carries one of the highest malaria burdens within the Greater Mekong region, despite having only 4% of its population [11,12]. Although Myanmar has successfully reduced the number of malaria cases by ~82% between 2012 and 2017 [13], the emergence of drug-resistant parasites in the region in recent years [14] has made Myanmar a critical battleground for the global fight against malaria. The use of satellite data and LCLU information has been demonstrated to be crucial in understanding landscape-scale malaria exposure and disease transmission. Previous work has established some linkages between the LCLU types and malaria risks. Specifically, deforestation and agricultural development have been shown to influence mosquito species abundance and malaria incidence through changes to the ecology of vector species as well as the nature and stage of the agricultural development determining patterns of human contact with vectors [7,15,16].
In addition to capturing and highlighting the areas where malaria exposure and transmission are more likely, LCLU maps could also be of great importance for malaria treatment efforts because they could provide information about the distribution of human settlements at large spatial scales. This is especially crucial for rural villages with low economic well-being and limited access to care [16,17]. However, creating LCLU maps that offer meaningful information about village distribution across the country is challenging because (1) it requires the mapping to be implemented at a 30-meter or higher resolution to ensure the successful mapping of remote villages (according to our personal observations, many remote villages in Myanmar are fairly small in extent and consist of a few buildings with small footprints), which are known to serve as the remnant pools of malaria [18], and (2) villages are quite similar to many bare ground features (e.g., riverbanks, bare mountain tops) in terms of their spectral signature. There are several existing LCLU datasets (Table 1) that provide information about human settlements that cover Myanmar at a 30-meter or higher resolution, including the SERVIR land cover dataset for the Greater Mekong region (hereafter referred to as “SERVIR Mekong land cover dataset”) [19], the Global Human Settlement Layer (GHSL) dataset [20], and the Global Human Built-up And Settlement Extent (HBASE) dataset [21]. However, due to either their multipurpose nature (SERVIR Mekong land cover) or global algorithms (GHSL and HBASE), their performance in identifying remote villages in Myanmar is limited (as shown in the Data Records section).
In this paper, we present an LCLU map of Myanmar, which is the first 30 m LCLU map designed specifically for the control and elimination of malaria in Myanmar. Being an application-oriented LCLU map, it was created following a bottom-up approach, which allows for the LCLU classes that are important for malaria transmission to be mapped with relatively higher confidence. In addition, this approach is inherently more flexible than the traditional methodologies used to produce LCLU maps, because (1) the classes and their priorities can be modified based on application-specific needs and (2) there is freedom to adopt the best available datasets and algorithms. Due to this flexibility, even though the produced LCLU map is only for the year 2016, given proper inputs, such LCLU representations could be produced for other years to develop a consistent time-series of LCLU maps (most of the classes with potentially high interannual variability in our classification scheme, such as managed forests and villages, can be updated annually). Moreover, similar LCLU maps could be developed for other vector-borne diseases based on their respective epidemiology or disease transmission characteristics. Additionally, due to its strong focus on remote villages, this methodology could potentially be used in other applications that are unrelated to public health but have equally strong demands on the accurate mapping of villages. This may include international humanitarian efforts, considering Myanmar has one of the highest refugee populations by country of origin, due to long-lasting and ongoing domestic conflicts [22,23].

2. Data Description

The presented dataset (https://doi.org/10.1594/PANGAEA.921126) [24] has been released through PANGAEA (https://www.pangaea.de/), an open data repository designated for the publication of datasets related to geoscience and environmental science. The dataset contains two data files. One is the LCLU map in a single GeoTiff file, and the other is a Microsoft Excel table listing the classes contained in the LCLU map. Figure 1 shows an overview of the entire LCLU map, with zoomed-in views of four different locations across Myanmar. Figure 2 shows the distribution of the ten classes in terms of the total area as mapped by the LCLU map across Myanmar.
We conducted a visual comparison between our map and the following three existing 30 m datasets that contain settlement-related classes (Table 1): the SERVIR Mekong land cover dataset [19], the GHSL dataset [20], and the HBASE dataset [21]. The comparison (Figure 3) shows that our product provides a much more spatially comprehensive representation of the classes related to human presence than all of the other datasets, especially in remote rural areas, where the villages typically consist of bare ground and small individual buildings. The difference observed in the mapping of human settlements between our product and the existing products is not surprising considering we adopted a locally developed algorithm to specifically map the village class.

3. Methods

This LCLU map is designed to support malaria research and the operational interventions for malaria elimination in Myanmar. Hence, we designed a classification scheme to include the land cover classes associated with the habitat suitability of the malaria vector—the Anopheles spp. mosquito—and the land use classes that can potentially increase human exposure to malaria. Our LCLU map was developed using several published LCLU datasets, modified ancillary data (for the depression and bare surface classes), and an independently mapped class (village). The combination of data sources inevitably led to several classes overlapping at certain locations. This meant that there was a need for a mapping hierarchy to ensure that each individual pixel of the final LCLU map belonged to a single class. Therefore, the classification scheme, which is discussed in detail subsequently, describes not only the classes that are included in the map but also the hierarchy of the mapped classes.
The methodology of our map development can be broken down into three major steps (Figure 4). First, the roads in Myanmar were manually digitized, leading to a much more complete and up-to-date road network dataset for Myanmar compared with other existing public datasets that we were aware of. Second, the resultant road network dataset was used in conjunction with a set of publicly available satellite imagery and ancillary data or their derivatives to map the village extent in Myanmar. Finally, the LCLU map was assembled based on the mapped village extent and a series of additional datasets.

3.1. Road Digitization

Road networks are an important indicator for understanding population distribution and human mobility. However, after conducting a visual assessment of VHR imagery and existing road network data in Myanmar, we determined that there are limited road network data available for Myanmar, especially in mountainous regions. From April 2019, as the basis of our effort using the administrative boundaries downloaded from the Myanmar Information Management Unit (MIMU, http://geonode.themimu.info, accessed in April 2019), we digitized roads using VHR Google Earth images based on the Myanmar Roads dataset that was provided by the Humanitarian OpenStreetMap Team [30]. Notably, we have added road networks in vast areas of Myanmar (Figure 5, center) and individual townships (select areas in different parts of the country are shown in Figure 5, panels 1–4). Townships are the third-level administrative divisions below districts (level two) and states/regions/union territories (level one). We digitized over roughly 25,000 km of additional roads.

3.2. Village Mapping

In Myanmar, rural populations are particularly susceptible to malaria because of their relatively lower economic status and access to care [17]. Therefore, understanding the distribution of the settlements, especially those that are remote, has huge implications for developing effective interventions. It is for that reason that our LCLU map includes a village class designated primarily to rural settlements that are not characterized by impervious surfaces (hence not mapped by impervious surface datasets such as GIMS [26]). Hoffman-Hall, Loboda, Hall, Carroll, and Chen [18] developed a random forest-based village mapping algorithm for Ann, a township in the Rakhine State of Myanmar. Considering Ann is only a small part of Myanmar (0.9% of the total area of Myanmar), an improved algorithm was needed to account for the great heterogeneity in environmental, socio-demographic conditions, and spatial patterns across the country. Here, building upon the foundation proposed by Hoffman-Hall, Loboda, Hall, Carroll, and Chen [18], we developed an updated village mapping algorithm. We adopted 26 variables, which were identified by Hoffman-Hall, Loboda, Hall, Carroll, and Chen [18] to be important for village mapping. We also added a slope variable, as our preliminary analysis confirmed its influence on village distribution. Table 2 lists the total of 27 variables that were fed into our random forest models. Most of these variables were derived based on corresponding Landsat imagery, i.e., one cloud-free Landsat composite for the dry–cold season (February–April 2016) and one for the dry–hot season (November 2016–January 2017). All of the variables were processed at a spatial resolution of 30 m.
At the country level, Myanmar has large variations in terrain. Based on our field experience, the contextual relationships of the individual buildings in the villages located in the mountain regions are evidently different to those in the plains, i.e., villages are generally larger and denser in the plains, whereas in the mountains, villages are much sparser, smaller, and more irregularly shaped. Based on this information, we created two training sample pools for the village class, separating mountain and plain topography. Using the SRTM DEM data, we divided the entire country of Myanmar into two terrain classes (mountain: ≥1000 m; plain: <1000 m, using an elevation threshold we selected empirically).
We used the village point dataset published by MIMU to guide our selection of training sample points. Covering the majority of Myanmar, with more than 47,000 points, this is the most comprehensive village point dataset to our knowledge. Every village point in the MIMU dataset was buffered by 500 m (a threshold value we chose empirically), after which 2000 sample points were randomly generated with the mountain and plain areas each having 1000 points. Then, we visually examined each of the 2000 random points using VHR imagery hosted on Google Earth to see whether that point was actually placed within the village. Owing to the differences in the spatial extents of the individual villages and the inaccuracy of some of the MIMU village points, a considerable number of the random points were placed outside of villages (e.g., in croplands or forests adjacent to villages). These out-of-bound random points were preserved and used as the training sample for the “non-village” class, while those random points correctly located in villages were considered as the training sample for the “village” class.
All of the training sample points (village + non-village) were used to extract the values from the independent variables listed in Table 2, resulting in two training sample pools, one for the mountain region and one for the plain region. Each of these training sample pools was fitted to a random forest model, which was subsequently applied to the testing data (created utilizing all of the independent variables for the entire country) to identify villages across Myanmar. Additional training points for each class were added iteratively upon visual examination of classification outputs. The final training dataset consisted of 19,790 points (village: 464; non-village: 19,326) for the mountain region and 3456 points (village: 215; non-village: 3241) for the plain region (the number of training sample points for the mountain region is considerably larger than that for the plain region due to the added difficulty of characterizing villages in areas with high topographic variability). A quantitative assessment of the mapping accuracy of the village class was performed as part of the combined LCLU map.

3.3. LCLU Map Assembly

LCLU map classes are chosen to be representative of the landscape. Additionally, based on our understanding of the malaria distribution and spread, as well as our field knowledge of the environmental and demographic conditions in Myanmar, several classes (Table 3) are also important in the context of malaria transmission. For example, we included depressions, defined as areas where standing water periodically or sporadically emerges after rain or floods, because they are a known breeding habitat of mosquitoes [37,38]. We included croplands because they cover a substantial proportion of the landscape and, in addition to flooded paddy fields being a mosquito habitat [39], they are also an important interface where mosquito–human transmission is likely to happen, as agriculture is the main source of income for Myanmar, accounting for more than a third of its gross domestic product [40]. Similarly, managed forests were also included in our LCLU map because they usually coincide with plantations (based on our field knowledge), which are another common environment for occupation-related exposure to malaria [12,41]. The impervious surface class, characterized by larger towns/cities and large manmade structures such as roads, was included because its association with malaria distribution and spread may be different from that in villages, where living conditions and access to care are more limited.
In the end, a total of ten classes were identified for the LCLU classification scheme (Table 3). These were, subsequently, assembled from separate data sources based on the following order of priority: perennial water > impervious surface > villages > croplands > managed forests > natural forests > ephemeral water > depressions > bare surfaces > shrub/grass. The priorities were determined empirically based on the expert opinion of in-country medical researchers and public health professionals on the general knowledge regarding the relevance of the class for malaria transmission. If a pixel met the criteria of multiple classes simultaneously, it was ultimately classified into the class with the highest priority. For example, if a pixel was mapped as impervious surface by GIMS, it would be eventually merged into the impervious surface class, unless it simultaneously met the criteria for perennial water, in which case it would be merged into the perennial water class.

4. Validation

We assessed the map classification accuracy using a stratified random sampling design guided by the “good practices” described by Olofsson et al. [44]. We used the open-source stratified area estimator design tool available on the SEPAL platform (https://sepal.io/), a cloud-based platform developed by the United Nations Food and Agriculture Organization Forestry Department’s Open Foris team, to implement the sampling design and generate the stratified random points. The dataset was reclassified in ArcMap to exclude the depressions class. The remaining map classes, except depressions, were used as strata. The stratified area estimator design tool calculates the overall sample size and the strata sample size using the equations described by Olofsson, Foody, Herold, Stehman, Woodcock, and Wulder [44]. The expected user’s accuracies for the perennial water, villages, croplands, natural forests, and bare surfaces strata were assigned 0.75, while the impervious surface, managed forests, ephemeral water, and shrub and grass strata were assigned 0.65. The desired standard error of the map’s overall accuracy was set to 0.013. A total of 1179 samples were created with at least 75 samples per class to ensure adequate samples of rare classes [44]. The samples were buffered to 30 m × 30 m in ArcMap and reviewed against Google Earth VHR imagery.
Although the land cover classifications represent distinct categories, a combination of various LCLU classes is frequently found within a 30-meter pixel, particularly for the classes with small footprints (e.g., impervious surface), or the classes with an expected mix of different land covers (e.g., villages). Additionally, the general lack of multi-temporal VHR data creates ambiguity in the exact extent of a class at a specific point in time (the point when the class was mapped vs. when the class is assessed) particularly for highly dynamic classes (e.g., perennial or ephemeral water), but also for many other classes. To account for this potential ambiguity and mixed pixels, we adopted a linguistic scale developed by Gopal and Woodcock [45] for fuzzy accuracy assessments of classified maps. The scale ranges from 1 to 5, where (1) is “absolutely wrong”, (2) is “understandable but wrong”, (3) is “reasonable or acceptable answer”, (4) is “good answer”, and (5) is “absolutely right” [46]. An interpreter reviewed each sample against each stratum according to the five-level scale. The responses were recorded using Microsoft Access.
Two fuzzy accuracy assessment functions, MAX and RIGHT [46], were calculated to assess the map classification. The MAX function considers correctness if the interpreter properly identified the sample as a five, or “absolutely right.” The RIGHT function considers correctness if the appropriate class is identified as a three or higher, (i.e., “reasonable or acceptable answer”, “good answer”, or “absolutely right”). Overall, the weighted fuzzy accuracy of the map classification is 69.22% (Table 4).

5. User Notes

We would like to demonstrate two possible scenarios where our LCLU map could be used, among others.
Scenario 1: Using our LCLU map to identify high-risk areas. Under this scenario, it is recommended that users first assess the relationship between each of the classes of our LCLU map and the specific infectious disease in question. The identified relationships can be used to determine the weight that is assigned to each class, which ultimately can serve as the foundation of a system that highlights the high-risk areas. If interannual changes in the distribution of the high-risk areas are of interest, users can apply the algorithm outlined in this paper to the data acquired in other years and generate similar LCLU-based risk maps for other years.
Scenario 2: Using our LCLU map to identify remote villages. As our intercomparison shows (Figure 3), our LCLU map offers one of the highest levels of accuracy for identifying remote villages among the existing datasets. Users can integrate our map with other datasets, such as OpenStreetMap road networks and DEM, to identify villages that are isolated and might need special attention in disease monitoring and eradication. Similar to Scenario 1, users can create annual LCLU maps for multiple years based on which spatio-temporal distribution of villages can be tracked across Myanmar.

6. Limitation and Caveat

While our LCLU map performs well identifying villages in general (as indicated in Table 4), it is unable to capture individual buildings. In other words, standalone buildings that are not part of a perceivable settlement are likely to be misclassified into other classes. This is, however, not surprising considering that the resolution of Landsat imagery is 30 m. For the delineation of the footprints of individual buildings, VHR imagery is likely required.
We would like to point out that the performance of the cropland class is likely to be strongly subject to seasonality. This is because Myanmar has a tropical monsoon climate, which means there is great seasonality in the precipitation patterns across the country. During the monsoon season, floods are prevalent in Myanmar, which may cause the flooded croplands to spectrally resemble water bodies. Therefore, for applications that require high accuracy in the delineation of croplands, users are encouraged to incorporate additional data with higher temporal resolutions, using our LCLU map as the baseline map.

Author Contributions

Conceptualization, T.V.L. and D.C.; methodology, T.V.L., D.C., and A.H.-H.; software, D.C., V.S., J.H., and A.H.-H.; validation, A.B.; investigation, D.C., V.S., J.H., A.H.-H., and Y.L.; resources, Q.Y.; data curation, D.C.; writing—original draft preparation, D.C., V.S., and A.B.; writing—review and editing, T.V.L., A.H.-H., Q.Y., and Y.L.; visualization, D.C. and A.B.; supervision, T.V.L.; project administration, T.V.L.; funding acquisition, T.V.L. All authors have read and agreed to the published version of the manuscript.

Funding

This work was supported by the NASA Applied Sciences Program under grant #80NSSC18K0326, the NASA Land-Cover and Land-Use Change Program under grant #NNX17AI06G, and the NIH International Center of Excellence in Malaria Research (ICEMR) Program under grant #5U19AI129386-04.

Data Availability Statement

The scripts written for this project are in Python, R, and IDL. We do not limit the access to the scripts. All scripts that were directly related to the production of this dataset have been publicly released through GitHub (https://github.com/dchengeo/Myanmar_LCLU_map/). A README file has also been uploaded through GitHub, which includes brief descriptions of the uploaded scripts. The stratified estimator design tool available on the SEPAL platform is available on sepal.io and GitHub (https://github.com/openforis/accuracy-assessment/tree/master/aa_design). The users are welcome to contact the corresponding author for the details of code implementation.

Conflicts of Interest

The authors declare no conflict of interest.

References

  1. Friedl, M.; Sulla-Menashe, D. MCD12Q1 MODIS/Terra+Aqua Land Cover Type Yearly L3 Global 500m SIN Grid. 2015. [Google Scholar] [CrossRef]
  2. Homer, C.; Dewitz, J.; Jin, S.; Xian, G.; Costello, C.; Danielson, P.; Gass, L.; Funk, M.; Wickham, J.; Stehman, S.; et al. Conterminous United States land cover change patterns 2001–2016 from the 2016 National Land Cover Database. ISPRS J. Photogramm. Remote Sens. 2020, 162, 184–199. [Google Scholar] [CrossRef]
  3. Vanwambeke, S.O.; Bennett, S.N.; Kapan, D.D. Spatially disaggregated disease transmission risk: Land cover, land use and risk of dengue transmission on the island of Oahu. Trop. Med. Int. Health 2011, 16, 174–185. [Google Scholar] [CrossRef] [Green Version]
  4. Lambin, E.F.; Tran, A.; Vanwambeke, S.O.; Linard, C.; Soti, V. Pathogenic landscapes: Interactions between land, people, disease vectors, and their animal hosts. Int. J. Health Geogr. 2010, 9, 54. [Google Scholar] [CrossRef] [Green Version]
  5. Beck, L.R.; Lobitz, B.M.; Wood, B.L. Remote sensing and human health: New sensors and new opportunities. Emerg. Infect. Dis. 2000, 6, 217–227. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  6. Bhunia, G.S.; Kesari, S.; Chatterjee, N.; Kumar, V.; Das, P. Localization of kala-azar in the endemic region of Bihar, India based on land use/land cover assessment at different scales. Geospat. Health 2012, 6, 177–193. [Google Scholar] [CrossRef] [Green Version]
  7. Stefani, A.; Dusfour, I.; Corrêa, A.P.S.; Cruz, M.C.; Dessay, N.; Galardo, A.K.; Galardo, C.D.; Girod, R.; Gomes, M.S.; Gurgel, H. Land cover, land use and malaria in the Amazon: A systematic literature review of studies using remotely sensed data. Malar. J. 2013, 12, 1–8. [Google Scholar] [CrossRef] [Green Version]
  8. Akter, R.; Naish, S.; Hu, W.; Tong, S. Socio-demographic, ecological factors and dengue infection trends in Australia. PLoS ONE 2017, 12, e0185551. [Google Scholar] [CrossRef]
  9. Weinstein, J.S.; Leslie, T.F.; von Fricken, M.E. Spatial associations between land use and infectious disease: Zika virus in Colombia. Int. J. Environ. Res. Public Health 2020, 17, 1127. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  10. King, R.J.; Campbell-Lendrum, D.H.; Davies, C.R. Predicting geographic variation in cutaneous leishmaniasis, Colombia. Emerg. Infect. Dis. 2004, 10, 598. [Google Scholar] [CrossRef] [PubMed]
  11. Mu, T.T.; Sein, A.A.; Kyi, T.T.; Min, M.; Aung, N.M.; Anstey, N.M.; Kyaw, M.P.; Soe, C.; Kyi, M.M.; Hanson, J. Malaria incidence in Myanmar 2005–2014: Steady but fragile progress towards elimination. Malar. J. 2016, 15, 503. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  12. Zaw, M.T.; Thant, M.; Hlaing, T.M.; Aung, N.Z.; Thu, M.; Phumchuea, K.; Phusri, K.; Saeseu, T.; Yorsaeng, R.; Nguitragool, W.; et al. Asymptomatic and sub-microscopic malaria infection in Kayah State, eastern Myanmar. Malar. J. 2017, 16, 138. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  13. WHO. Help Prevent Malaria—WHO Myanmar Newsletter Special; WHO: Geneva, Switzerland, 2018. [Google Scholar]
  14. Nyunt, M.H.; Shein, T.; Zaw, N.N.; Han, S.S.; Muh, F.; Lee, S.-K.; Han, J.-H.; Thant, K.Z.; Han, E.-T.; Kyaw, M.P. Molecular evidence of drug resistance in asymptomatic malaria infections, Myanmar, 2015. Emerg. Infect. Dis. 2017, 23, 517. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  15. Yasuoka, J.; Levins, R. Impact of deforestation and agricultural development on anopheline ecology and malaria epidemiology. Am. J. Trop. Med. Hyg. 2007, 76, 450–460. [Google Scholar] [CrossRef] [Green Version]
  16. Hoffman-Hall, A.; Puett, R.; Silva, J.; Chen, D.; Baer, A.; Han, K.T.; Han, Z.Y.; Thi, A.; Htay, T.; Thein, Z.W.; et al. Malaria exposure in Ann Township, Myanmar as a function of land cover and land use: Combining satellite earth observations and field surveys. GeoHealth 2020, 4. [Google Scholar] [CrossRef] [PubMed]
  17. Han, S.M.; Rahman, M.M.; Rahman, M.S.; Swe, K.T.; Palmer, M.; Sakamoto, H.; Nomura, S.; Shibuya, K. Progress towards universal health coverage in Myanmar: A national and subnational assessment. Lancet Glob. Health 2018, 6, e989–e997. [Google Scholar] [CrossRef] [Green Version]
  18. Hoffman-Hall, A.; Loboda, T.V.; Hall, J.V.; Carroll, M.L.; Chen, D. Mapping remote rural settlements at 30 m spatial resolution using geospatial data-fusion. Remote Sens. Environ. 2019, 233, 111386. [Google Scholar] [CrossRef]
  19. Saah, D.; Tenneson, K.; Poortinga, A.; Nguyen, Q.; Chishtie, F.; Aung, K.S.; Markert, K.N.; Clinton, N.; Anderson, E.R.; Cutter, P.; et al. Primitives as building blocks for constructing land cover maps. Int. J. Appl. Earth Obs. Geoinf. 2020, 85, 101979. [Google Scholar] [CrossRef]
  20. Corbane, C.; Pesaresi, M.; Kemper, T.; Politis, P.; Florczyk, A.J.; Syrris, V.; Melchiorri, M.; Sabo, F.; Soille, P. Automated global delineation of human settlements from 40 years of Landsat satellite data archives. Big Earth Data 2019, 3, 140–169. [Google Scholar] [CrossRef]
  21. Wang, P.; Huang, C.; Brown de Colstoun, E.C.; Tilton, J.C.; Tan, B. Global Human Built-up and Settlement Extent (HBASE) Dataset from Landsat. 2017. [Google Scholar] [CrossRef]
  22. UNHCR. Refugee Data Finder. Available online: https://www.unhcr.org/refugee-statistics/ (accessed on 23 February 2021).
  23. Ullah, A.A. Rohingya refugees to Bangladesh: Historical exclusions and contemporary marginalization. J. Immigr. Refug. Stud. 2011, 9, 139–161. [Google Scholar] [CrossRef]
  24. Chen, D.; Baer, A.; He, J.; Hoffman-Hall, A.; Shevade, V.; Ying, Q.; Loboda, T.V. Land cover land use map for Myanmar at 30-m resolution for 2016. 2020. [Google Scholar] [CrossRef]
  25. Sexton, J.O.; Song, X.-P.; Feng, M.; Noojipady, P.; Anand, A.; Huang, C.; Kim, D.-H.; Collins, K.M.; Channan, S.; DiMiceli, C.; et al. Global, 30-m resolution continuous fields of tree cover: Landsat-based rescaling of MODIS vegetation continuous fields with lidar-based estimates of error. Int. J. Digit. Earth 2013, 6, 427–448. [Google Scholar] [CrossRef] [Green Version]
  26. Brown de Colstoun, E.C.; Huang, C.; Wang, P.; Tilton, J.C.; Tan, B.; Phillips, J.; Niemczura, S.; Ling, P.-Y.; Wolfe, R.E. Global Man-made Impervious Surface (GMIS) Dataset from Landsat. 2017. [Google Scholar] [CrossRef]
  27. Thenkabail, P.S.; Knox, J.W.; Ozdogan, M.; Gumma, M.K.; Congalton, R.G.; Wu, Z.; Milesi, C.; Finkral, A.; Marshall, M.; Mariotto, I. Assessing future risks to agricultural productivity, water resources and food security: How can remote sensing help? Photogramm. Eng. Remote Sens. 2012, 78, 773–782. [Google Scholar]
  28. Tucker, C.J. Red and photographic infrared linear combinations for monitoring vegetation. Remote Sens. Environ. 1979, 8, 127–150. [Google Scholar] [CrossRef] [Green Version]
  29. Ying, Q.; Hansen, M.C.; Potapov, P.V.; Tyukavina, A.; Wang, L.; Stehman, S.V.; Moore, R.; Hancher, M. Global bare ground gain from 2000 to 2012 using Landsat imagery. Remote Sens. Environ. 2017, 194, 161–176. [Google Scholar] [CrossRef] [Green Version]
  30. OpenStreetMap Contributors. OpenStreetMap. Available online: https://data.humdata.org/dataset/hotosm_mmr_roads/ (accessed on 1 February 2021).
  31. Giglio, L.; Schroeder, W.; Justice, C.O. The collection 6 MODIS active fire detection algorithm and fire products. Remote Sens. Environ. 2016, 178, 31–41. [Google Scholar] [CrossRef] [Green Version]
  32. Schroeder, W.; Oliva, P.; Giglio, L.; Csiszar, I.A. The New VIIRS 375m active fire detection data product: Algorithm description and initial assessment. Remote Sens. Environ. 2014, 143, 85–96. [Google Scholar] [CrossRef]
  33. DeVries, B.; Huang, C.; Lang, M.W.; Jones, J.W.; Huang, W.; Creed, I.F.; Carroll, M.L. Automated quantification of surface water inundation in wetlands using optical satellite imagery. Remote Sens. 2017, 9, 807. [Google Scholar] [CrossRef] [Green Version]
  34. Vermote, E.; Justice, C.; Claverie, M.; Franch, B. Preliminary analysis of the performance of the Landsat 8/OLI land surface reflectance product. Remote Sens. Environ. 2016, 185, 46–56. [Google Scholar] [CrossRef] [PubMed]
  35. Gao, B.-C. NDWI—A normalized difference water index for remote sensing of vegetation liquid water from space. Remote Sens. Environ. 1996, 58, 257–266. [Google Scholar] [CrossRef]
  36. Crist, E.P.; Cicone, R.C. A physically-based transformation of Thematic Mapper data—The TM Tasseled Cap. IEEE Trans. Geosci. Remote Sens. 1984, GE-22, 256–263. [Google Scholar] [CrossRef]
  37. Fillinger, U.; Sombroek, H.; Majambere, S.; van Loon, E.; Takken, W.; Lindsay, S.W. Identifying the most productive breeding sites for malaria mosquitoes in The Gambia. Malar. J. 2009, 8, 62. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  38. Chen, T.; Zhang, S.; Zhou, S.-S.; Wang, X.; Luo, C.; Zeng, X.; Guo, X.; Lin, Z.; Tu, H.; Sun, X.; et al. Receptivity to malaria in the China–Myanmar border in Yingjiang County, Yunnan Province, China. Malar. J. 2017, 16, 478. [Google Scholar] [CrossRef] [PubMed]
  39. Richards, E.E.; Masuoka, P.; Brett-Major, D.; Smith, M.; Klein, T.A.; Kim, H.C.; Anyamba, A.; Grieco, J. The relationship between mosquito abundance and rice field density in the Republic of Korea. Int. J. Health Geogr. 2010, 9, 32. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  40. FAO. Myanmar at a Glance. Available online: http://www.fao.org/myanmar/fao-in-myanmar/myanmar/en/ (accessed on 17 August 2020).
  41. Singhasivanon, P.; Thimasarn, K.; Yimsamran, S.; Linthicum, K.; Nualchawee, K.; Dawreang, D.; Kongrod, S.; Premmanisakul, N.; Maneeboonyang, W.; Salazar, N. Malaria in tree crop plantations in south-eastern and western provinces of Thailand. Southeast Asian J. Trop. Med. Public Health 1999, 30, 399–404. [Google Scholar] [PubMed]
  42. Pickens, A.H.; Hansen, M.C.; Hancher, M.; Stehman, S.V.; Tyukavina, A.; Potapov, P.; Marroquin, B.; Sherani, Z. Mapping and sampling to characterize global inland water dynamics from 1999 to 2018 with full Landsat time-series. Remote Sens. Environ. 2020, 243, 111792. [Google Scholar] [CrossRef]
  43. Hansen, M.C.; Potapov, P.V.; Moore, R.; Hancher, M.; Turubanova, S.A.; Tyukavina, A.; Thau, D.; Stehman, S.V.; Goetz, S.J.; Loveland, T.R.; et al. High-Resolution Global Maps of 21st-Century Forest Cover Change. Science 2013, 342, 850–853. [Google Scholar] [CrossRef] [Green Version]
  44. Olofsson, P.; Foody, G.M.; Herold, M.; Stehman, S.V.; Woodcock, C.E.; Wulder, M.A. Good practices for estimating area and assessing accuracy of land change. Remote Sens. Environ. 2014, 148, 42–57. [Google Scholar] [CrossRef]
  45. Gopal, S.; Woodcock, C. Theory and methods for accuracy assessment of thematic maps using fuzzy sets. Photogramm. Eng. Remote Sens. 1994, 60, 181–188. [Google Scholar]
  46. Woodcock, C.E.; Gopal, S. Fuzzy set theory and thematic maps: Accuracy assessment and area estimation. Int. J. Geogr. Inf. Sci. 2000, 14, 153–172. [Google Scholar] [CrossRef]
Figure 1. Overview of the presented LCLU map for Myanmar. Insets 1-4 shows zoomed-in views of 4 locations, respectively.
Figure 1. Overview of the presented LCLU map for Myanmar. Insets 1-4 shows zoomed-in views of 4 locations, respectively.
Data 06 00063 g001
Figure 2. Distribution in terms of total area of the ten classes contained in the presented LCLU map.
Figure 2. Distribution in terms of total area of the ten classes contained in the presented LCLU map.
Data 06 00063 g002
Figure 3. A comparison between the current product and three existing land cover products in terms of the representation of human presence (SERVIR Mekong [19]: urban and built-up; HBASE [21]: built-up and settlement; GHSL [20]: built-up; current product: impervious surface and villages). All other classes irrelevant to human presence in the compared products are masked out (indicated in black). The SERVIR Mekong land cover map that was compared and displayed was produced for 2016 (the same year as our product).
Figure 3. A comparison between the current product and three existing land cover products in terms of the representation of human presence (SERVIR Mekong [19]: urban and built-up; HBASE [21]: built-up and settlement; GHSL [20]: built-up; current product: impervious surface and villages). All other classes irrelevant to human presence in the compared products are masked out (indicated in black). The SERVIR Mekong land cover map that was compared and displayed was produced for 2016 (the same year as our product).
Data 06 00063 g003
Figure 4. Workflow for the development of the presented dataset. The purple boxes with dashed line boundaries indicate the three major technical steps (corresponding to the subsections of Methods). The blue and green boxes indicate the major data inputs and intermediate outputs, respectively. The orange box represents the final output, which is the produced LCLU map. Abbreviations: a Shuttle Radar Topography Mission (SRTM), a digital elevation model (DEM), a Moderate Resolution Imaging Spectroradiometer (MODIS), a Visible Infrared Imaging Radiometer Suite (VIIRS), the Landsat Vegetation Continuous Fields (Landsat VCF) project [25], Very High Resolution (VHR), the global surface water dynamics (GSWD) product, the Global Man-made Impervious Surface (GMIS) product [26], the Global Food Security Analysis-Support (GFSAD) product [27], the Global Forest Change (GFC) product [25], the normalized difference vegetation index (NDVI) [28], the global bare ground gain (GBG) product [29].
Figure 4. Workflow for the development of the presented dataset. The purple boxes with dashed line boundaries indicate the three major technical steps (corresponding to the subsections of Methods). The blue and green boxes indicate the major data inputs and intermediate outputs, respectively. The orange box represents the final output, which is the produced LCLU map. Abbreviations: a Shuttle Radar Topography Mission (SRTM), a digital elevation model (DEM), a Moderate Resolution Imaging Spectroradiometer (MODIS), a Visible Infrared Imaging Radiometer Suite (VIIRS), the Landsat Vegetation Continuous Fields (Landsat VCF) project [25], Very High Resolution (VHR), the global surface water dynamics (GSWD) product, the Global Man-made Impervious Surface (GMIS) product [26], the Global Food Security Analysis-Support (GFSAD) product [27], the Global Forest Change (GFC) product [25], the normalized difference vegetation index (NDVI) [28], the global bare ground gain (GBG) product [29].
Data 06 00063 g004
Figure 5. The digitized road network (in red lines) in comparison to the OpenStreetMap road network [30] (in blue lines) of Myanmar, with edits to the dataset made as necessary. Panels one through four show comparisons of the digitized road network and the OSM road network of (1) Bhamo Township in the Kachin State, (2) Demoso Township in the Kayah State, (3) Hakha Township in the Chin State, and (4) Kyunhla Township in the Sagaing Region. These four regions were selected because they were among the areas that have the highest number of roads (in distance) added and represent different parts of the country. Administrative boundaries are from MIMU.
Figure 5. The digitized road network (in red lines) in comparison to the OpenStreetMap road network [30] (in blue lines) of Myanmar, with edits to the dataset made as necessary. Panels one through four show comparisons of the digitized road network and the OSM road network of (1) Bhamo Township in the Kachin State, (2) Demoso Township in the Kayah State, (3) Hakha Township in the Chin State, and (4) Kyunhla Township in the Sagaing Region. These four regions were selected because they were among the areas that have the highest number of roads (in distance) added and represent different parts of the country. Administrative boundaries are from MIMU.
Data 06 00063 g005
Table 1. List of datasets whose performance in mapping settlement-related classes were compared.
Table 1. List of datasets whose performance in mapping settlement-related classes were compared.
DatasetLCLU TypeName of
Settlement-Related Classes
Spatial ResolutionYear
SERVIR MekongMulti-classUrban and built-up30 mAnnually
between 1987 and 2018
HBASESingle classBuilt-up and settlement30 m2010
GHSLSingle classBuilt-up30 m2013–2014
Current ProductMulti-class(1) Impervious surface, (2) villages30 m2016
Table 2. List of variables used by the random forest models and their brief descriptions. For index calculation formulas, refer to Hoffman-Hall, Loboda, Hall, Carroll, and Chen [18]. Dry–cold and dry–hot seasons refer to February–April 2016 and November 2016–January 2017, respectively, in Myanmar.
Table 2. List of variables used by the random forest models and their brief descriptions. For index calculation formulas, refer to Hoffman-Hall, Loboda, Hall, Carroll, and Chen [18]. Dry–cold and dry–hot seasons refer to February–April 2016 and November 2016–January 2017, respectively, in Myanmar.
VariablePeriod
Covered
Data SourceUnitOriginal
Resolution
Description
Distance to MODIS Active Fire2016MODIS active fire product [31]MeterVectorEuclidean distance to MODIS active fire points.
Distance to VIIRS Active Fire2016VIIRS active fire product [32]MeterVectorEuclidean distance to VIIRS active fire points.
Distance to RoadsN/AUpdated road network (described previously)MeterVectorEuclidean distance to all roadways from OpenStreetMap and our digitized road network.
Distance to 3rd Order or Greater WaterwayN/ASRTM DEMMeter30 mEuclidean distance to 3rd order or greater waterway identified using SRTM DEM following Hoffman-Hall, Loboda, Hall, Carroll, and Chen [18].
Distance to WaterN/ANorth America surface water map [33]Meter30 mEuclidean distance to all waterbodies.
NIR Mean Occurrence (Dry–Cold)Dry–Cold SeasonLandsat imageryUnitless30 mMean values of 3 × 3 kernels based on the NIR band of Landsat 8.
Landsat 8 Band 7 SWIR2 (Dry–Cold)Dry–Cold SeasonLandsat imageryUnitless30 mLandsat 8 Band 7 (SWIR2) for the dry–cold season.
Landsat 8 Band 10 TIRS1 (Dry–Cold)Dry–Cold SeasonLandsat imageryKelvin30 mLandsat 8 Band 10 (TIRS1) for the dry–cold season.
Landsat 8 Band 11 TIRS2 (Dry–Cold)Dry–Cold SeasonLandsat imageryKelvin30 mLandsat 8 Band 11 (TIRS2) for the dry–cold season.
NBR2 (Dry–Cold)Dry–Cold SeasonLandsat imageryUnitless30 mNormalized Burn Ratio 2 (NBR2 [34]) for the dry–cold season.
NDVI (Dry–Cold)Dry–Cold SeasonLandsat imageryUnitless30 mNDVI for the dry–cold season.
NDWI6 (Dry–Cold)Dry–Cold SeasonLandsat imageryUnitless30 mNormalized Difference Water Index using SWIR1 band (NDWI6 [35]) for the dry–cold season.
NDWI7 (Dry–Cold)Dry–Cold SeasonLandsat imageryUnitless30 mNormalized Difference Water Index using SWIR2 band (NDWI7 [35]) for the dry–cold season.
Tasseled Cap Wetness (Dry–Cold)Dry–Cold SeasonLandsat imageryUnitless30 mTasseled Cap Wetness (TCW [36]) for the dry–cold season.
Texture: NIR Mean Occurrence (Dry–Hot)Dry–Hot SeasonLandsat imageryUnitless30 mMean values of 3 × 3 kernels based on the NIR band of Landsat 8.
Landsat 8 Band 7 SWIR2 (Dry–Hot)Dry–Hot SeasonLandsat imageryUnitless30 mLandsat 8 Band 7 (SWIR2) for the dry–hot season.
Landsat 8 Band 10 TIRS1 (Dry–Hot)Dry–Hot SeasonLandsat imageryUnitless30 mLandsat 8 Band 10 (TIRS1) for the dry–hot season.
Landsat 8 Band 11 TIRS2 (Dry–Hot)Dry–Hot SeasonLandsat imageryUnitless30 mLandsat 8 Band 11 (TIRS2) for the dry–hot season.
NBR2 (Dry–Hot)Dry–Hot SeasonLandsat imageryUnitless30 mNBR2 for the dry–hot season.
NDVI (Dry–Hot)Dry–Hot SeasonLandsat imageryUnitless30 mNDVI for the dry–cold season.
NDWI6 (Dry–Hot)Dry–Hot SeasonLandsat imageryUnitless30 mNDWI6 for the dry–cold season.
NDWI7 (Dry–Hot)Dry–Hot SeasonLandsat imageryUnitless30 mNDWI7 for the dry–cold season.
Tasseled Cap Wetness (Dry–Hot)Dry–Hot SeasonLandsat imageryUnitless30 mTCW for the dry–hot season.
Seasonal Difference in Tasseled Cap Wetness2016Landsat imageryUnitless30 mAbsolute value of TCW difference between the dry–hot and dry–cold seasons.
ElevationN/ASRTM DEMMeter30 mTerrain elevation directly based on the SRTM DEM.
SlopeN/ASRTM DEM%30 mTerrain slope calculated based on the SRTM DEM.
Tree Cover2015Landsat VCF tree cover [25]%30 mTree cover based on Landsat VCF dataset.
Table 3. List of classes of the LCLU map and the corresponding input data and technical summaries.
Table 3. List of classes of the LCLU map and the corresponding input data and technical summaries.
ClassDefinitionInput DataTechnical Summary
Perennial WaterConsistent water surface with low seasonal or interannual variabilityGSWD [42]Pixels mapped by GSWD as “permanent water” (water bodies with consistent extent between 1999 and 2018) or “water gain” (water bodies that emerged between 1999 and 2018).
Impervious SurfaceMan-made surface such as buildings and concrete ground surface that is different from bare groundGMIS [26]Pixels mapped by GMIS as having impervious proportion values of larger than 1%.
VillagesAggregation of buildings in rural areas, built on bare groundMapped previouslyVillage extent as mapped by our random forest-based village mapping algorithm.
CroplandsCroplandsGFSAD [27]Pixels mapped by GFSAD as croplands.
Managed ForestsForests that show signs of disturbances in the near past (i.e., since 2000)GFC [43]Pixels mapped by GFC as forest loss between 2000 and 2016.
Natural ForestsForests that do not show signs of disturbance in the near past (i.e., since 2000)Landsat VCF Tree Cover [25]Pixels mapped by Landsat VCF product as having tree cover values of larger than 40%.
Ephemeral WaterWater surface with high seasonal or interannual variabilityGSWD [42]Pixels mapped by GSWD as “stable seasonal”, “high frequency”, “dry period”, or “wet period”. These four classes correspond to areas where high levels of seasonal or interannual variability between land and water took place.
DepressionsAreas with high levels of curvature and are likely hotspots for standing water after rain or floodsSRTM DEMPixels whose curvature values are between −10 and −1 and flow accumulation values are greater or equal to 3. Curvature and flow accumulation were calculated based on SRTM DEM using the corresponding tools in ArcGIS.
Bare SurfacesAreas with limited tree cover and low NDVI values (<0.5)GBG [29]Pixels mapped by GBG as bare ground or have NDVI values of less than 0.5 (a threshold value chosen empirically). The NDVI values were calculated based on a Landsat cloud-free composite created for the dry–cold season in 2016 for entire Myanmar.
Shrub/GrassAreas with limited tree cover but high NDVI values (≥0.5)N/AAll pixels that were not mapped as the classes above were classified into this class.
Table 4. Results of MAX and RIGHT functions.
Table 4. Results of MAX and RIGHT functions.
Map StrataTotal Samples (n)Correct
Samples
Defined by MAX
Function
(n)
MAX
(M)
Function
Correct
Samples
Defined by RIGHT
Function
(n)
RIGHT
(R)
Function
Improvement in
Accuracy
when Fuzziness Is
Considered
(R-M)
Area Weights
Perennial water754458.67%5674.67%16.00%0.007
Impervious surface753445.33%5066.67%21.33%0.002
Villages753344.00%6384.00%40.00%0.003
Croplands1529059.21%11877.63%18.42%0.190
Managed forests752432.00%4864.00%32.00%0.048
Natural forests39325665.14%33284.48%19.34%0.489
Ephemeral water751925.33%3546.67%21.33%0.029
Shrub and grass184179.24%6233.70%24.46%0.229
Bare surfaces752330.67%3141.33%10.67%0.003
Total117954045.80%79567.43%21.63%
Total weighted
accuracy
48.20% 69.22%21.02%
Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Share and Cite

MDPI and ACS Style

Chen, D.; Shevade, V.; Baer, A.; He, J.; Hoffman-Hall, A.; Ying, Q.; Li, Y.; Loboda, T.V. A Disease Control-Oriented Land Cover Land Use Map for Myanmar. Data 2021, 6, 63. https://doi.org/10.3390/data6060063

AMA Style

Chen D, Shevade V, Baer A, He J, Hoffman-Hall A, Ying Q, Li Y, Loboda TV. A Disease Control-Oriented Land Cover Land Use Map for Myanmar. Data. 2021; 6(6):63. https://doi.org/10.3390/data6060063

Chicago/Turabian Style

Chen, Dong, Varada Shevade, Allison Baer, Jiaying He, Amanda Hoffman-Hall, Qing Ying, Yao Li, and Tatiana V. Loboda. 2021. "A Disease Control-Oriented Land Cover Land Use Map for Myanmar" Data 6, no. 6: 63. https://doi.org/10.3390/data6060063

Article Metrics

Back to TopTop