Mangrove forests were naturally distributed in coastal areas where slow-moving waters allow fine sediments to settle and may prevent coastal soil erosion during hurricane seasons in the warm coastlines of tropical oceans all over the world. It also provides a wide range of ecological and socioeconomic services for human beings and habitats for various wildlife forms, such as shorebirds, crab-eating monkeys, and fishing cats [1–3]. Extensive development of coastal areas in many developing countries by drastically transforming mangrove forests to other land uses such as aquaculture and built-up areas has constantly ignored such ecological and socioeconomic services of mangrove ecosystems [4–6]. The total area of mangrove forests in the world was drastically declined from approximately 16.5 million ha in 1980s to 15.2 million ha in 2000s [7, 8].
The destruction was mainly driven by the conversion of mangrove forests to shrimp farms and coastal urbanization to meet the human needs of socioeconomic development [9–11]. The destruction of mangrove forests has reduced their viability and the quality of socioeconomic and ecological services, consequently creating environmental problems, such as land degradation, soil erosion, and loss of habitats of animal and plant species [12–15]. Thus, information of the status of spatial distributions of mangrove forests is deemed critical for forest managers and scientists to evaluate current land-use management policies.
The estuary complex in the Gulf of Fonseca, Central America sharing the boundaries of El Salvador, Honduras, and Nicaragua is an ideal study site. The region characterized by a diverse landscape of mangrove forests, marshes, mudflats, and lagoons connected with the Pacific Ocean is thus significant for biodiversity conservation. However, uncontrolled aquaculture development, especially shrimp farming, started in early 1980s  that provided much demanded socioeconomic development and employment for the local community in the region  has driven mangrove deforestation, constantly degrading ecological and socioeconomic services of mangrove forests causing environmental issues, including water pollution, land degradation, and changes in hydrological regime [18–20].
Therefore, understanding of the spatial density and current distribution of mangrove forests in the study region is deemed important to the land-use assessment and rehabilitation of mangrove ecosystems.
Remote sensing has long been recognized as one of the most efficient tools for evaluating land-use/land-cover (LULC) because it can provide spatiotemporal data in a timely manner at the local, regional, and global scales. The low-resolution remote sensing, such as moderate resolution imaging spectroradiometer (MODIS) and advanced very high resolution radiometer (AVHRR), has advantages for LULC monitoring at a regional scale because it has a wide coverage and is free of charge for data acquisition. However, it seems infeasible in the context of the study region owing to the mixed-pixel issue of small and fragmented patches of mangrove forests. This study used Rapideye data processed by using the support vector machines (SVM) [21–23], which is a machine learning classifier [24, 25], for mapping the density status and spatial distribution of mangrove forests in the Gulf of Fonseca, Central America. The advantages of using Rapideye imageries for monitoring mangrove forests in the region were that this satellite sensor has ability to acquire high-resolution data (5 m resolution) over a 77 km wide swath with five spectral bands in which the red-edge wavelength (690–730 nm) between red and near-infrared wavelengths could be found useful for discriminating vegetation types. The red-edge band is strongly correlated with foliar chlorophyll content and vegetation stress [26, 27].
In this study, aside from mapping the density status of mangrove forests we also tested if the red-edge band of Rapideye satellite data could improve the mangrove mapping results in the study region. The data were processed using SVM because this classifier can perform complex classification tasks, including mangrove forests and spectral mixture problems [28–32]. The SVM demonstrates to produce more accurate classification outcome than traditional classifiers, such as maximum likelihood, decision tree, and neural networks [33, 34]. Because our hypotheses also tried to bridge a research gap if the contribution of Rapideye red-edge band to the classification results of mangrove density in the study region can be confirmed, four numerical schemes of band combination were tested: scheme-1 (bands 1–3, 5 excluding the red-edge band 4), scheme-2 (bands 1–5), scheme-3 (bands 1–3, 5 incorporating with the normalized difference vegetation index, NDVI), and scheme-4 (bands 1–3, 5 incorporating with the normalized difference red-edge index, NDRI). To avoid biases when comparing the classification results across different schemes, the same training pixels in the group of candidate pixels were extracted from the ground reference dataset for classification and the same pixels were extracted from the group of pixels for accuracy assessment.
The objective of this study was to delineate the density of mangrove forests from Rapideye data in the Gulf of Fonseca, Central America using SVM. The results produced from this study could led to the provision of an unique agro-ecological insight of the status and distributions of dense mangrove, sparse mangrove, agriculture (e.g., cropping areas and orchards), fallowed land, and mudflat, which may essentially contribute to devise successful plans for aquaculture development and conservation of mangrove ecosystems in the study region.
2 Study region
We selected the Gulf of Fonseca, Central America to investigate the density of mangrove forests from Rapideye data (Figure 1). The region bordered by El Salvador, Honduras, and Nicaragua is an estuarine complex characterized by a subtropical climate and saline-affected conditions that are suitable for aquaculture development and shrimp culture. This study region was diversified by landscapes and rich in biodiversity because of presence of globally important ecosystems of mangrove forests, marshes, mudflats, and lagoons. It is thus significant for ecosystem conservation and eco-tourism. However, due to rapid population growth and meeting basic needs of the rural livelihoods and socioeconomic development, many parts of the region have been transformed to aquaculture fields, especially shrimp farms . The conversion of mangrove forests to aquaculture and shrimp farms without taking into account long-term impacts on sustainable development has consequently led to mangrove deforestation and environmental issues .
3 Data collection
Two Rapideye images (level 3A) covering the estuary complex (Gulf of Fonseca) acquired on 4 and 14 March 2012 were used in this study. The data have five spectral bands: band 1 (blue), band 2 (green), band 3 (red), band 4 (red-edge), and band 5 (near infrared). The red-edge band spectrally located between the red and near-infrared bands is able to provide useful information to characterize plant types and land-cover abundance. The advanced spaceborne thermal emission and reflection radiometer (ASTER) digital elevation model (DEM) (30 m resolution) was used for masking out high elevation areas. The 2011 El Salvadorian LULC map (scale: 1/125,000) provided by the Salvadoran Ministry of Agriculture, the 2002 Honduran LULC map (scale: 1/250,000) obtained from the Honduras National Institute of Forest Conservation and Development, the 2009 LULC map of the Gulf of Fonseca , the 2011 global LULC map (from the Land Cover Type Yearly L3 Global 500 m SIN Grid Product collected from NASA), and topographic maps were also used as reference data sources for crosschecking and accuracy assessment of the classification results.
Because of the unavailability of updated LULC maps covering the study region for accuracy assessment of the classification results, this study counted on the ground reference data created after the field surveys using Google Earth imagery, currently available LULC maps, and the information recorded from the field work conducted in 2011. We segmented the Rapideye image into a number of classes using the Isodata clustering algorithm. The results were overlaid with the existing LULC maps and compared with the high resolution Google Earth imagery for crosschecking. The ground reference map was created by digitizing homogenous sites associated with each land-cover class. This map was then rasterized into a grid of 5 m resolution (Figure 2) and used for selecting training pixels for SVM classification and accuracy assessment of the classification results.
The flowchart shows a three-step procedure of data processing adopted in this study (Figure 3): (1) data pre-processing including image ortho-rectification and non-vegetated area masking, (2) image classification using SVM, and (3) accuracy assessment of the classification results based on the ground reference data, followed by the testing of the significant difference among the four numerical schemes using the Z-test statistic.
4.1 Data pre-processing
The ortho-rectification of Rapideye images was processed to remove distortions caused by terrain and image tilt effects for the purpose of creating a planimetrically correct image. This process was carried out using the ASTER DEM (30 m resolution) and 30 ground control points uniformly chosen from distinct features throughout the study region using LULC and topographic maps. The images were registered to the Universal Transverse Mercator (UTM) system (datum zone 16 N) and mosaicked. The results indicated that the geometric accuracy of the systematically corrected images was less than 5 m (1 sigma).
Two vegetation indices (NDVI and NDRI) were calculated to test the relative contribution of the red-edge band in various mapping schemes. These indices were calculated as follows: (1) (2)
where ρred (630-685 nm), ρrededge (690-730nm), and ρnir (760-850 nm) are red, red-edge and near-infrared bands, respectively. The NDVI  is sensitive to the amount of greenness biomass, and thus is widely adopted for monitoring changes in vegetation growth. Because the reflectance of green vegetation changes abruptly in the wavelength region of 670–780 nm due to the effect of strong chlorophyll absorption and leaf internal scattering so that NDRI is useful for studying the chlorophyll content as a measure of vegetation growth condition [39, 40].
Because mangroves are naturally grown in intertidal coastlines between the land and sea, the areas higher than 30 m could be masked out using the ASTER DEM. The non-vegetated areas (i.e., water bodies and built-up areas) were also masked out using the mean NDVI, where its value was smaller than 0. The masking process was to limit our analysis to vegetated areas that are lower than 30 m.
4.2 Image classification
Based on the visual interpretation of the Google Earth imagery, the available LULC maps, and the data collected from the field surveys, six classes were identified for SVM classification (Figure 4), including: dense mangrove (dark purple tones), sparse mangrove (brighter purple tones), agriculture such as cropping areas and orchards (orchid tones), fallowed land (white tones), and mudflat (dark grey tones).
The SVM is a nonparametric classifier based on the statistical learning theory . This method classifies data by finding the optimal hyperplane that separates all data points of one class from those of the other class. Assume that the data for training is set of points xi corresponding to their categories yi, where xi ∈ R and yi = ±1. For a linear case of two classes, class 1 is represented as +1 and class 2 represented as –1. The equation of a hyperplane is expressed as w · x + b = 0, where x is a point lying on the hyperplane, w is a parameter determining the hyperplane orientation, and b is the offset of the hyperplane from the origin. The data must fall on the proper side of the hyperplane of either yi · xi + b ≥ +1 or yi · xi + b ≤ –1. These two hyperplanes can be rewritten as yi (w · xi + b) –1 ≥ 0. Therefore, if an optimal hyperplane exists, the training data are separable. The margin between the two planes is 2/¦w¦ so that the optimal hyperplane is required to minimize this distance 2/¦w¦ using Lagrange multipliers [41, 42]. Details about SVM can be found in the textbook of the nature of statistical learning theory .
For a nonlinear case, this algorithm can be extended using kernel functions K(xi, w) to make the computation easier in feature space by replacing the vector product in w · x + b = 0 and yi(w · xi + b) –1 ≥ 0. In this study, the radial basis function (RBF) was used because it demonstrates to produce more accurate results in classification of hyperspectral data than other functions such as linear and polynomial kernels . The RBF is expressed as follows: (3)
where γ > 0 is the kernel parameter, xi is a set of samples (xi ∈ Rn) is an n-dimensional vector and yi is the classl abel.
The training samples selected out of the synchronous collection of the ground reference data were sorted out to train SVM for classification based on Rapideye data. To avoid biases due to selecting the same training pixels for SVM classification and accuracy assessment of the classification results, the ground reference data were partitioned randomly into two different groups of pixels, namely group-1 for training and group-2 for accuracy assessment. From the group of training pixels (group-1), a total of 12,000 pixels (2,000 for each class) were randomly extracted to train SVM for classification.The Jeffries-Matusita distance (JM) (values from 0–2) was used to verify spectral separability between two classes , calculated as follows: (4)
where B is the Bhattacharyya distance , expressed as follows: (5)
where m1, m2 and σ1, σ2 are the class means and class variances, respectively. The JM value of 2 indicates a complete separation between two classes, and the lower values indicate the higher possibility of having cases of misclassification. A refinement process was eventually applied to merge non-mangrove classes (i.e., agriculture, fallowed fields, and mudflat) into a newer class, namely others.
4.3 Accuracy assessment
The classification results were verified using the ground reference data (i.e., group of pixels for accuracy assessment). For each class, a total of 2,000 pixels were randomly extracted to compare with those from the classification maps using the confusion matrix. Kappa coefficient and overall accuracy were calculated to measure the classification accuracy. Other parameters including producer and user accuracies were also computed to measure per-class classification accuracy. The Z-test statistic based on Kappa statistic [46, 47] was also calculated to test if the classification results produced from the four band combination schemes were significantly different. Assuming Rapideye data has a normal distribution, the test statistic (Z) becomes: (6)
where k1 and k2 are the kappa coef?cients obtained from the results of accuracy assessment between two band combination schemes, and their associated variances (σ1, σ2). Introducing a confidence level of 95% the two band combination schemes differ significantly if Z >1.96.
5.1 Spectral information of different training classes
The JM distance measure applied to the training samples randomly extracted from the ground reference data indicated that SVM could achieve satisfactory spectral separation among proposed land-cover classes (Table 1). The mean JM values derived from the five spectral bands for pairs of classes were generally greater than 1.6. The results thus suggested that the selected classes could be appropriate for SVM classification. There were some exceptions that indicated the lower JM values, including sparse mangrove–forest (JM = 0.73), sparse mangrove–mudflat (JM=1.02), forest–mudflat (JM=1.33), and mudflat–agriculture (JM=1.31). It was thus noticeable that distinction between the sparse mangrove and forest might not be successful in SVM classification due to spectral confusion.
5.2 Distribution of mangrove forests and accuracy assessment results
The classification results achieved from SVM classification of the 2012 Rapideye data showed the spatial distribution of mangrove forests in the Gulf of Fonseca (Figure 5). The mangrove forests generally sheltered the coastlines and thin fringes of estuaries of the Gulf of Fonseca. In general, the comparable spatial distributions of mangrove forests between four different band combination schemes were salient. The dense mangroves were more concentrated in the upper part of the region bordered by El Salvador and Honduras because in this region there were several natural reserves that were strictly protected by the local governments for biodiversity conservation. The majority of sparse mangrove forests was found in the lower part of the region because a number of shrimp farms have long been present in the coastal waters of mangrove forests for the purpose of aquaculture development in the study region.
The classification results associated with different band combination schemes were compared with the ground reference data to evaluate the accuracy of SVM classification. The comparisons were facilitated by a way that 2,000 pixels for each class (i.e., dense mangrove, sparse mangrove, and others) were randomly extracted from the ground reference data (group-2) to compare with those obtained from the classification maps produced by four different numerical schemes. The agreement across these four schemes can be generally confirmed with the overall accuracies and Kappa coefficients that are generally higher than 96% and 0.9, respectively, in all cases (Table 2). Of a total of 2,000 pixels retrieved to determine the accuracy in each class, the class having a lower producer accuracy level was observed for the sparse mangrove class due to spectral confusion based on the JM distance. The spectral confusion was often present due to encountering mixed-pixel problems in areas where road and canal networks have been well-developed for transportation contributing to the possible classification errors.
When comparing the classification performance among the four numerical schemes, the highest performance was found by using scheme-2 (bands 1–5). The overall accuracy and Kappa coefficient were 97.0% and 0.95, as compared to the corresponding values obtained from scheme-1 (bands 1–3, 5, without the red-edge band 4) that were 96.2% and 0.94, respectively. The stability of the red-edge band contributing to the improvement of classification results was evidenced by comparing the two schemes associated with the inclusion of NDVI and NDRI. The slightly better overall accuracy of 96.9% and Kappa coefficient of 0.95 can be found in scheme-4 (bands 1–3, 5 and NDRI), as compared to the corresponding values of 96.1% and 0.94, respectively, in scheme-3 (bands 1–3,5 and NDVI). Hence, the results can be confirmed by the fact that the better classification results can be achieved by using schemes 2 and 4, which incorporated the red-edge band.
The Z-test statistic was applied to determine if the classification results produced from schemes 2 and 4 that incorporated the red-edge band were significantly different with those produced from schemes 1 and 3 that excluded the red-edge band for SVM classification. The results of this pairwise test (between two confusion matrices) indicated that the classification results obtained from schemes 2 and 4 with red-edge band were significantly different from schemes 1 and 3 without the red-edge band.This could be evidenced by that the Z-test values were greater than 2.1 or more, greater than the critical value of 1.96 (Table 3). The Z-test values obtained for the pairs of schemes 1 and 2, and those of schemes 1 and 4 were 2.51 and 2.14, whereas the Z-test value achieved for the pair of schemes 1 and 3 was simply 0.28. Similarly, the Z-testvalues of 2.79 and 2.42 were obtained for the pairs of schemes 2 and 3, and schemes 3 and 4, respectively. The pair of schemes 2 and 4 had a Z-test value of 0.37 only. From these findings, it might be concluded that the contribution of red-edge band improved SVM classification results of mangrove density in the study region.
It was worthwhile to compare advantages and disadvantages of SVM applied to Rapideye data for classification of mangrove forests associated with the red-edge band. The SVM applied to Rapideye data confirmed the validity of this classification method for mapping the mangrove density in the Gulf of Fonseca with the overall accuracy and Kappa coefficient higher than 96% and 0.9, respectively. The advantages of SVM were that this algorithm is relatively flexible, allowing to use unlimited training samples, provided that such samples are expected to be pristine. In this study, the training samples were randomly extracted from the ground reference image. Given that SVM is a nonlinear algorithm, enabling us to account for a wide range of data types, such as nonlinear vegetation indices (i.e., NDVI and NDRI), which can be used without priori assumptions about the data. Once SVM is fully trained, the classification can be well-performed. However, disadvantages of using SVM for classification rested on that the choice of a kernel function and its parameter values also affected the level of classification accuracy. In this study, the selection of RBF kernel was made based on literature review. With this clue, the SVM algorithm appeared to be a good one for mapping mangrove forests based on Rapideye data in the study region.
The classification outcome clearly indicated that the incorporation of the red-edge band information into the SVM classification algorithm improved only approximately 0.83% and 0.82% of the overall accuracies for scheme-2 that incorporated the red-edge band and scheme-4 that incorporated NDRI as compared to schemes 1 and 3. However, the Z-test statistic for testing the significant difference between two confusion matrices revealed that the contribution of the red-edge band to the classification results of mangrove forests in Central America was significant, given that the Z-test values was higher than 1.96. This contribution could be explained that the increase of the classification accuracy was attributed to the larger sensitivity of the red-edge band than the near-infrared band for the vegetation cover differentiation. As indicated in the classification results, the sparse mangrove class yielded higher producer accuracy levels associated with schemes 2 and 4 both of which incorporated the red-edge band, when compared to those associated with schemes 1 and 3 both of which did not incorporate the red-edge band.
It was however noticed that in this study the classification results produced by the pixel-based SVM algorithm and the accuracy assessment were carried out using the randomly extracted pixels synchronous with the ground reference data based on the ground reference data which may had inherent discrepancies to some extent. Further investigations of the contribution of the red-edge band to the improvement of LULC classification should thus be carried out using different classifiers to confirm the importance of red-edge band for LULC classification.
The findings achieved from this study confirmed the validity of SVM for density mapping of mangrove forests in the Gulf of Fonseca, Central America. The mapping results compared with the ground reference data indicated the overall accuracies and Kappa coefficients generally higher than 96% and 0.9, respectively. The slightly higher levels of classification accuracy were observed for schemes 2 and 4 both of which incorporated the red-edge band in the classification. Although the Z-test results statistically supported the inclusion of red-edge band in the classification of mangrove forests in the study region, the mapping results showed that such an incorporation improves only 0.83% and 0.82% of the overall accuracies for these two schemes. Overall, this study could provide quantitative information in regard to the status and density distributions of mangrove forests in the study region, which may be important for natural resources managers to evaluate their current management practices of mangrove ecosystems in regard to how to formulate a better long-term management strategy.
This study was supported by the Taiwan International Cooperation and Development Fund (ICDF-101-011).The financial support was gratefully acknowledged. The field work would not have been possible without the help of staff from the Ministries of Agriculture of El Salvador, Nicaragua, and Honduras.
Costanza, R. Visions, values, valuation, and the need for an ecological economics. Bioscience 2001, 51. Google Scholar
Giri, C.; Ochieng, E.; Tieszen, L.L.; Zhu, Z.; Singh, A.; Loveland, T.; Masek, J.; Duke, N. Status and distribution of mangrove forests of the world using earth observation satellite data. Global Ecol. Biogeogr. 2011, 20, 154-159. Web of ScienceCrossrefGoogle Scholar
Duke, N.C.; Meynecke, J.O.; Dittmann, S.; Ellison, A.M.; Anger, K.; Berger, U.; Cannicci, S.; Diele, K.; Ewel, K.C.; Field, C.D. , et al. A world without mangroves? Sci 2007, 317, 41-42. CrossrefGoogle Scholar
Rahman, A.F.; Dragoni, D.; Didan, K.; Barreto-Munoz, A.; Hutabarat, J.A. Detecting large scale conversion of mangroves to aquaculture with change point and mixed-pixel analyses of high-fidelity modis data. Remote Sens. Environ. 2013, 130, 96- 107. Web of ScienceCrossrefGoogle Scholar
Alongi, D.M. Present state and future of the world’s mangrove forests. Environ. Conserv. 2002, 29, 331-349. Google Scholar
FAO. The world’smangroves 1980-2005: Fao forestry paper 153; Food and Agriculture Organization of the United Nations, Rome, 2007. Google Scholar
Spalding, M.; Kainuma, M.; Collins, L. World atlas of mangroves. Earthscan: 2010. Google Scholar
Duke, N.C.; Meynecke, J.-O.; Dittmann, S.; Ellison, A.M.; Anger, K.; Berger, U.; Cannicci, S.; Diele, K.; Ewel, K.C.; Field, C.D. , et al. A world without mangroves? Science 2007, 317, 41-42. CrossrefGoogle Scholar
Kirui, K.B.; Kairo, J.G.; Bosire, J.; Viergever, K.M.; Rudra, S.; Huxham, M.; Briers, R.A. Mapping of mangrove forest land cover change along the kenya coastline using landsat imagery. Ocean Coast. Manage. 2013, 83, 19-24. Web of ScienceCrossrefGoogle Scholar
Vannucci, M.Mangrove management and conservation: Present and future; United Nations University Press, Tokyo, Japan, 2004.Google Scholar
Walters, B.B.; Rönnbäck, P.; Kovacs, J.M.; Crona, B.; Hussain, S.A.; Badola, R.; Primavera, J.H.; Barbier, E.; Dahdouh-Guebas, F. Ethnobiology, socio-economics and management of mangrove forests: A review. Aquat. Bot. 2008, 89, 220-236.CrossrefWeb of ScienceGoogle Scholar
FAO. Global production statistics 1950-2005; Food and Agriculture Organization of the United Nations, Rome, Italy, 2005. Google Scholar
Tobey, J.A.; Clay, J.; Vergne, P.; Center, U.o.R.I.C.R. Maintaining a balance: The economic, environmental and social impacts of shrimp farming in latin america. Coastal Resources Center: 1998. Google Scholar
Stonich, S.C.; Bort, J.R.; Ovares, L.L. Globalization of shrimp mariculture: The impact on social justice and environmental quality in central america. Society&Natural Resources 1997, 10, 161-179.Google Scholar
Dewalt, B.R.; Vergne, P.; Hardin, M. Shrimp aquaculture development and the environment: People, mangroves and fisheries on the gulf of fonseca, honduras. World Development 1996, 24, 1193-1208. CrossrefGoogle Scholar
Boser, B.E.; Guyon, I.; Vapnik, V. In A training algorithm for optimal margin classiers, The Fifth Annual Workshop on Computational Learning Theory, 1992; ACM Press, New York, USA. Google Scholar
Vapnik, V.N. The nature of statistical learning theory. Springer: 1995. Google Scholar
Cortes, C.; Vapnik, V. Support-vector network. MLear 1995, 20, 273-297. Google Scholar
Foody, G.M.; Mathur, A. A relative evaluation of multiclass image classification by support vector machines. Geoscience and Remote Sensing, IEEE Transactions on 2004, 42, 1335-1343. CrossrefGoogle Scholar
Haykin, S. Neural networks: A comprehensive foundation. Prentice Hall.: Upper Saddle River, NJ, 1994. Google Scholar
Collins, W. Remote sensing of crop type and maturity. Engineering 1978, 44, 43-55. Google Scholar
van der Meer, F.D.; de Jong, S.M. Imaging spectrometry:: Basic principles and prospective applications. Springer: 2011. Google Scholar
Kavzoglu, T.; Colkesen, I. A kernel functions analysis for support vector machines for land cover classification. IJAEO 2009, 11, 352-359.Google Scholar
Singh, S.K.; Srivastava, P.K.; Gupta, M.; Thakur, J.K.; Mukherjee, S. Appraisal of land use/land cover of mangrove forest ecosystem using support vector machine. Environmental Earth Sciences 2013, 71, 2245-2255.Web of ScienceGoogle Scholar
Kanniah, K.; Sheikhi, A.; Cracknell, A.; Goh, H.; Tan, K.; Ho, C.; Rasli, F. Satellite images for monitoring mangrove cover changes in a fast growing economic region in southern peninsular malaysia. Remote Sensing 2015, 7, 14360.CrossrefWeb of ScienceGoogle Scholar
Otukei, J.R.; Blaschke, T. Land cover change assessment using decision trees, support vector machines and maximum likelihood classification algorithms. IJAEO 2010, 12, 27-31. Google Scholar
Stonich, S.C. The promotion of non-traditional agricultural exports in honduras: Issues of equity, environment and natural resource management. Development and Change 1991, 22, 725- 755. CrossrefGoogle Scholar
Southworth, J.; Munroe, D.; Nagendra, H. Land cover change and landscape fragmentation—comparing the utility of continuous and discrete analyses for a western honduras region. Agriculture, Ecosystems & Environment 2004, 101, 185-205. CrossrefGoogle Scholar
CCAD. Land-use/land-cover map of the gulf of fonseca; Central American Commission for the Environment and Development, Central America: 2009. Google Scholar
Rouse, J.W.; Haas, R.H.; Schell, J.A.; Deering, D.W. In Monitoring vegetation systems in the great plains with erts, Proceedings of the Third ERTS Symposium, Washington DC, 1973; Washington DC, pp 309-317. Google Scholar
Jensen, J.R. Remote sensing of the environment: An earth resource perspective. Pearson Prentice Hall: 2007. Google Scholar
Arfken, G. Lagrange multipliers. In Mathematical methods for physicists, 3rd ed.; Academic Press: Orlando, FL, 1985; pp 945-950. Google Scholar
Sundaram, R.K. A first course in optimization theory. Cambridge University Press: 1996. Google Scholar
Richards, J.A.; Jia, X. Remote sensing digital image analysis: An introduction. Springer, New York, USA: 2006. Google Scholar
Bhattacharyya, A. On a measure of divergence between two statistical populations defined by their probability distributions. Bulletin of the Calcutta Mathematical Society 1943, 35, 99-109. Google Scholar
Congalton, R.G.; Oderwald, R.G.; Mead, R.A. Assessing landsat classification accuracy using discrete multivariate analysis statistical techniques. Photogrammetric engineering and remote sensing. 1983, 49, 1671-1678. Google Scholar
Congalton, R.G.; Green, K. Assessing the accuracy of remotely sensed data: Principles and practices. CRC Press: 2009. Google Scholar
About the article
Published Online: 2017-06-09
Citation Information: Open Geosciences, Volume 9, Issue 1, Pages 211–220, ISSN (Online) 2391-5447, DOI: https://doi.org/10.1515/geo-2017-0018.
© 2017 Nguyen-Thanh Son et al.. This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 3.0 License. BY-NC-ND 3.0