Improving land use/ land cover classification utilizing a

0

Available online freely at www.isisn.org

Bioscience Research Print ISSN: 1811-9506 Online ISSN: 2218-3973

Journal by Innovative Scientific Information & Services Network

RESEARCH ARTICLE BIOSCIENCE RESEARCH, 2018 15(4):4049-4060. OPEN ACCESS

Improving land use/ land cover classification utilizing a hybrid method of decision trees and artificial neural networks

R.S. Morgan1 and M. Faisal2 1Soils and Water Use Department, Agricultural and Biological Research Division, National Research Centre, Dokki Cairo, Egypt 2Drainage Research Institute, National Water Research Center, El-Kanater El-Khairia, Cairo, Egypt *Correspondence: [email protected] Accepted: 05 Dec.2018 Published online: 31 Dec. 2018

Monitoring the changes in land use/ land cover (LU/LC) has become increasingly important for sustainable management of the natural resources to overcome the problems of haphazard development, deteriorating environmental quality, loss of fertile agricultural lands and destruction of wetlands. Comparable to the areas located at the northern part of the Nile Delta, EL-Manzala Lake and its surroundings are continuously subjected to numerous threats. These threats include converting parts of the lake into other land uses such as the fish ponds located to the south and south-east of the lake. The lake water is also subjected to pollution from the inflow of agriculture drainage water. Hence, the necessity arises to have a permanent monitoring strategy for this area to enable decision makers to develop plans to overcome these threats. The aim of this work is to develop a hybrid method that integrates decision trees (DT) and artificial neural networks (ANN) to increase the accuracy of classification in a selected area covering El-Manzala Lake and its surrounding. In this study, the DT was designed first in a manner that the classes can be separated and the DT results are then input into an ANN. Two satellites data were integrated as the input of the classification. The data included the reflectance of the Sentinel 2 satellite bands from the blue to the shortwave bands (discarding the red edge bands), and the emission of the thermal bands of the Landsat 8 satellite. The images of both satellites were acquired at the same date. Moreover, theses data were used in the format of band ratios and specific principal component analysis (PCAs). Utilizing this approach the overall accuracy of the classification could reach 98.6% while it could only reach 70.2 % when processing the data using the DT only and has been lower using ANN. Keywords: land use/ land cover - decision trees - neural networks – hybrid method

INTRODUCTION Classification is data mining task that predicts

categorical class labels (Kim, 2016). Decision tree is machine learning technique, which divides the dataset to reach a homogenous classification of a target variable (Beucher et al., in press). It is a non-parametric classifier that does not require any prior statistical assumptions to be made regarding the distribution of data. The basic structure of the

decision tree consists of one root node, a number of internal nodes, branches and finally a set of terminal nodes (Otukei and Blaschke, 2010). The internal node represents a test on an attribute, each branch represents an outcome of the test, and each terminal node represents a class. The topmost decision node is called root node and corresponds to the best predictor. These techniques have substantial advantages for

http://www.isisn.org/

mailto:[email protected]

Morgan and Faisal Hybrid DT and NN method for LU/LC classification

Bioscience Research, 2018 volume 15(4): 4049-4060 4050

remote sensing classification problems because of their flexibility, intuitive simplicity, and computational efficiency (Friedl and Brodleyf, 1997). However, the method has some drawbacks, including instability, the non-smooth nature of the decision boundary, and the possibility of over fitting (Kim, 2016). Artificial neural network (ANN) is an information processing paradigm that is inspired by the way biological nervous systems, such as the brain, process information (Basu et al., 2010). An ANN is formed from hundreds of single units (artificial neurons) connected with coefficients (weights), which constitute the neural structure. These neurons are organized in layers and connected in a network and working in unison to solve specific problems (Agatonovic-Kustrin and Beresford, 2000). The main characteristics of neural networks are that they have the ability to learn complex nonlinear input-output relationships, use sequential training procedures, and adapt themselves to the data (Basu et al., 2010). Recently decision tree is becoming one of the most widely used methods for land use/ land cover classification of remote sensing prevailing other methods (Friedl and Brodleyf 1997 and Pal and Mather, 2003 and Kandrika and Roy, 2008 and Punia et al., 2011 and Chasmer et al., 2014 and Ghulam et al., 2014 and Zhang et al., 2017). Nevertheless, the potentiality of ANN for land use/ land cover classification has been reported by Morgan et al., (2015) which recommended further exploration of the use of ANN when classifying a similar area located in the central Nile Delta area. However, they reported that the overall accuracy did not exceed 65 %. Both DT and ANN can be improved using hybrid methods that combine the two methods in a way that save the advantages of each method in order to construct a more effective method (Zhou and Chen, 2002). Zhao (2004) classified the common types of these hybrid methods into the following: 1) Transformational approaches in which a DT is designed first, and then transformed into an ANN; 2) the DT is Embed into a ANN which can be considered a kind of modular neural network with each module being a small DT; 3) Embed the NNs into a DT, it is a DT with each non-terminal node containing an expert neural network.

The objective of this work is to develop a transformational hybrid approach utilizing DT and ANN for land use\ land cover classification in a selected area in the north east of Egypt using remote sensing data. In this approach the decision tree was used to help in determining the

topology of neural networks. MATERIALS AND METHODS

Study area The study area is located at the north eastern

part of the Nile Delta of Egypt between

coordinates 30o

51\ 2.2” N, 31o

43’ 20.1” E and 31o

31’ 20.1” N, 32o 18’ 46.7” E, (Figure 1).

Figure1: Location map of the study area

It covers an area of about 2800 km2 and

includes the area between the Bahr El-Baqar drain in the east and the extends southward Mediterranean Sea to the north and it for about 50 km from the coast. The Damietta Branch of Nile River in the west outlined part of its western boundary. The maximum temperature in the study

area varies from 20.1oC to 35.3

oC while the

minimum temperature ranges between 7.5oC to

20.2oC. Total precipitation varies from 29.2 to 72.3

mm (Elnaggar, 2013). The topography of the north-eastern Nile Delta is characterized by a low relief, and its surface slopes gently towards the northerly direction, while it holds a rolling shape



towards the south, where the ground climbs up to a moderately elevated plateau. The vegetation in the study area is dominated by rice, cotton, maize and some vegetables such as tomato, potato, cucumber and few orchard trees (Ezzeldin et al., 2016). El-Manzala Lake is located on the north-eastern edge of the Nile Delta and considered the largest of the delta lakes. The water quality of the lake declined due to the inflow of agricultural, domestic and industrial drainage water, mainly located in the southern section of the lake. Other environmental degradation factors that affected the lake included the overgrowth aquatic vegetation. Large parts of the study area used to be occupied by sabkhas and salt marshes. Nevertheless, it have been subjected to agricultural expansion projects planned by the government and led to the conversion of large areas of the sabkhas and salt marshes into agricultural areas.

Currently, there is a need for an effective and accurate approach to monitor the spatial and temporal changes in agricultural lands in the eastern part of Nile-Delta in Egypt to provide decision makers with reliable information in order to establish rational land use policies (Elnaggar, 2013).

Data collection and pre-processing This study used both Sentinel-2 (S2) and

Landsat 8 images collected at the same date on 7/10/2017. The area was covered by two images of each satellite. S2 was developed by the European Space Agency (ESA) and provides high spatial, spectral and temporal resolution images of the earth. The satellites provide a set of 13 spectral bands, four bands at 10m, six bands at 20m and three bands at 60m spatial resolution (Clevers & Gitelson 2013 and Vajsova & Aastrand 2015). On the other hand, Landsat 8 represents collaboration between NASA and USGS and, acquires global moderate resolution measurements of the Earth's surface in the visible, near-infrared, short wave, and thermal

infrared wavelength. The satellite carries a two-sensor payload, the Operational Land Imager (OLI) and the Thermal Infrared Sensor (TIRS). It provides a set of 11 spectral bands all of 30 m spatial resolution except the panchromatic band at 10m and the two thermal bands at 100m spatial resolution (Roy et al., 2014). In this study the Sentinel-2 data was used specifically Band 2 -Band 12 excluding Band 8a – Band 10 (Table 1), in addition to the thermal bands of Landsat 8. All the data were radiometrically and geometrically corrected and projected into Universal Transverse Mercator (UTM) and World Geodetic System 1984 (WGS 84) with datum-zone 36 N. The S2 data were available in Geographic Markup Language JPEG2000 (GMLJP2) format while the Landsat 8 data were available in Geotiff format. The S2 data were imported into QGIS 2.18.0 software which supports the GMLJP2 format and exported as Geotiff into ENVI software for further processing, while the Landsat 8 data were directly processed in ENVI. Both S2 and Landsat 8 bands which were not at 10m spatial resolution were resampled into 10m using Band 8. The selected bands of each image were stacked into one multiband image resulting two images for each satellite. Each two images were mosaicked together to create a new scene and a subset of this scene covering the geographical extent of the study area was extracted.

Data Pre-processing Utilizing the ENVI software the unsupervised

classification was applied on the S2 image to collect the required information about the study area and to choose the training sets which represent the different land use/land cover classes. Based on literature and initial field work and knowledge of the study area six classes were used for the unsupervised classification including the following: vegetation, aquatic vegetation, water, fishponds, fallow soils and urban area (Table2).

Table 1: demonstrates the characteristics of the S2 bands used in this study

Band No. Abbreviations* Band centre (nm) Spatial resolution(m)

Band 2 B 490 10

Band 3 G 560 10

Band 4 R 665 10

Band 8 IR 842 10

Band 11 SWIR1 1610 20

Band 12 SWIR2 2190 20 *B:Blue, G:Green, R:Red, IR: Infra Red, SWIT1: Short Wave Infra Red band1, SWIT2: Short Wave Infra Red band2



Table 2: Characteristics of the Land use\ Land cover classes

Class No. Class Name Definition

Class 1 Water El-Manzala lake with relatively deep water

Class 2 Fishponds Areas used to grow fish and usually at the initial stages of

land reclamation.

Class 3 Vegetation Agricultural area with various crops at various growth

stages

Class 4 Aquatic vegetation It occupies parts of El-Manzala Lake where the water is

shallow

Class 5 Fallow soils Prepared for cultivation or in the seedling phase

Class 6 Urban area Villages and towns

Accordingly, samples representing each class

were observed from the study area and a total of 55 polygons representing these samples were delineated on the S2 image. Spectral signatures and signature separability

Region of interests (ROI) representing each sample was extracted from the above mentioned polygons. Spectral signatures were created for each of the six land cover\ land use classes based on values extracted from S2 image. The spectral separability of the training sets was evaluated to examine the spectral separability between each two classes. The separability was determined by using the “Jeffries-Matusita and Transformed Divergence separability” measures, and a separability index was computed between each pair of training samples (Richards and Jia, 2006) and Sunil et al., (2010). The separability between each two classes should be more than 1.8 to indicate that each class is properly separated (Zhu et al., 2007). Taking into account that the more separable training data is, the better classification is expected.

Data processing Using ENVI, the data in the form of vegetation

indices (VIs) were produced. Furthermore, the principal component analysis (PCA) of the selected bands was used in this study. PCA is a methodology that can reduce the redundancy in multispectral data (Pemberty, 2002). The bands representing the PCA data are non-correlated and independent, and ordered in terms of the amount of variance (Santiesteban, 2003).

Integration between DT and NN In this study the transformation approach was

used to integrate the DT and NN. Using this approach we extracted the data from the DT and used it to design a feed-forward neural network. When designing the neural network 70% of the

samples were used for training, 15% for testing and 15% for validation. The appropriate neural network was assessed based on prediction accuracy. The structure of the neural network was manually alternated using a trial-and-error process until the highest accuracy achieved. RESULTS AND DISCUSSION

The separability between classes was more than 1.9 according to both Jeffries-Matusita and Transformed Divergence measures for each two classes indicating a good separability between classes. Furthermore, the feature space plot was applied to further analysis of the separability of the classes. However, the feature plot demonstrated overlaps between some of the classes. These overlaps were observed between the vegetation and aquatic vegetation classes (figure 2), the water & fish ponds classes and the urban area & fallow soils classes. These overlaps could be explained when examining the spectral profile of the studied classes (figure 3). In this graph the spectral similarities of each two classes could be observed. The vegetation and aquatic vegetation classes had almost identical spectral profile. Similar properties were observed between the water & fish ponds classes and the urban area & fallow soils classes. Such similarity makes these classes the hardest to differentiate.

Decision trees Based on the above, a decision tree was

manually constructed examining the spectral characteristics of the studied classes. When examining the spectral curve, it was concluded that a simple ratio between blue and shortwave infrared bands could differentiate the water and fish ponds from the other classes(Eq1). This ratio exceeded 2 for the fish ponds and was higher for the water class exceeding 4 (figure 4A).

Ref

lect

ance



Figure 2: Feature space plot of the R and IR bands for the vegetation and aquatic vegetation classes

Figure 3: Spectral profiles of the land use/ land cover classes in the study area



Figure 4: Value maps A: B/SWIR1, B:SAVI, C:Thermal PCA and D:Short Wave PCA

It was less than 1 for the remaining four classes. B/SWIR1=RB/RSWIR1 Eq1

Where R refers to reflectance, and the subscripts refer to a specific spectral band as shown in Table 1. Thereafter, the vegetation and aquatic vegetaion were differntiated from other classes using the soil adjusted vegetaion index (SAVI) developed by Huete (1988) (Eq2). In sparse vegetated areas, the reflection of the soils are much higher than reflection of vegetation and therefore, the detection of vegetation cover is difficult. Therefore, the soil-adjusted vegetation

index (SAVI) was developed to minimize soil influences on canopy spectra by incorporating a soil adjustment factor L into the denominator of the normalized difference vegetation index (NDVI) equation (Qi et al., 1994). We used L value of 0.5 as recommended by Huete (1988) as this could reduce soil noise considerably throughout a wide range of vegetation amounts.

SAVI=((BIR–BR)/(BIR+BR+L))*(1+L) Eq2 Where B refers to reflectance, and the subscripts refer to a specific spectral band as shown in Table 1 and L is the soil adjustment factor.

A B

C D



SAVI values were negative for the water and fish ponds classes while it excedded 0.5 for both the vegetaion and aquatic vegetaion classes.It ranged from 0 – 0.4 for the urban area & fallow soils classes. Nevertheless, it was difficult to differentiate between the the vegetaion and aquatic vegetaion classes using SAVI because their values interacted (figure 4B). The SAVI values ranged from 0.56 to 0.84 in case of the aquatic vegetation, while it ranged from 0.25 to 0.98 in case of the vegetation. Therefore, to differntiate between these two classes the thermal bands of landsat 8 were used. Furthermore, instead of using the two thermal bands we used the first PCA of these two bands. Using this PCA, the water, fish ponds and aquatic vegetion classes had positave values while all other classes had a negative values (figure 4C).

Thereafter, another spesfic PCA was applied on the shortwave infrared bands and this PCA could be used to differntiate the urban area & fallow soils classes. The urban area values ranged from 1000-1700, while the fallow soils was more than 2000. The remaining classess had values less than 900 (figure 4D). The designing of the proposed decision tree using the above mentioned paramters is shown in Figure 5. The ENVI software decision tree tool was used to apply the previous classification tree. The overall accuracy was 70.2 % and the resulting classified map is shown in figure 6. These results were unsatisfactory as these were lower than the accuracy produced by methods such as Maximum Likelihood classifier (MLC) and Support Vector Machine (SVM) (84 % and 79 % respectively). Nevertheless, they were slightly higher the the ANN which had accuracy 65% when applied to a similar area (Morgan et al., 2015).

Hybrid method The selected parameters of the design tree

were used in building the topology of the neural nettwork. These parametrs included the B/SWIR1, SAVI and the first PCA of both the thermal and shortwave infrared bands. Therafter, a four band image was created with the previously mentaioned parameters representing each band. For each observation, the values representing each parameter and the cooresponding classses exceperesed as numirical value were imported into Matlab. The data was transformed into a formate accepted by Neural Network pattern recognition tool using the Matlab script. Therafter the netwrok was designed as shown in figure 7.

The designed network had one input layer with 4 nodes representing the four parameters and a hidden layer with 10 nodes and one output layer with six nodes representing the land use\ land cover classes. The overall accuracy reached 97.2 %.The confusion matrix produced by the neural network is shown in table 3. The rows represented the predicted class and the columns represented to the actual class. The diagonal cells correspond to observations that are correctly classified and the remaining cells correspond to incorrectly classified observations.

Examining the confusion matrix, both the urban area and fallow soils had an actual and prediction accuracy of 100%. On the other hand, 25.0% of the water class were misclassified as fish ponds reducing the prediction classification accuracy of the fish ponds class to 90.0 %. That could be rendered to the similarity in depth and clarity of both classes in the selected samples. Likewise, only 83.3% of the vegetation class were correctly classified and the rest was misclassified as aquatic vegetation and thus reducing the prediction accuracy of the aquatic vegetation to 95.2%.

Thereafter, the four band image was converted into tiff format and imported into Matlab and processed using the designed neural network. Nevertheless, it is worth mentioning that due to memory problems the image was partionated into smaler parts for Matlab to process. After the processing, the processed tiff images were imported into QGIS and glued into one image and imported into ILWIS for further processing. The classified land use\ land cover map is shown in figure 8 and the area covered by each class is shown in table 4.

The lake area covered only 875.4 Km2 whereas In 1970 the area of the lake as citied by El-Wakeel and Wahby (1970) was 1275 km2. Taking into account that only 31.1% of this area is covered by open water and the rest is occupied by invasive aquatic vegetation species such as Phragmites australis (common reed) and Typha domigensis (reed maca) which has increased over the years due to the degredation of the lake water quality due to the discharge of drainage water into the lake. Furthermore, the expansion of the fish ponds in the north and north western parts of the lake could be noticed which is further reducing the lake area.



Figure 5: The proposed decision tree with six classes

Figure 6: Land use\ land cover classification according to the DT



Figure 7: The architect of the proposed neural network

Table 3: Confusion Matrix for produced by the Matlab

pre

dic

ted

Class

Actual

Water Fishponds Vegetation Aquatic

vegetation Fallow soils

Urban area

Water 75.0 0.0 0.0 0.0 0.0 0.0

Fishponds 25.0 100 0.0 0.0 0.0 0.0

Vegetation 0.0 0.0 83.3 0.0 0.0 0.0

Aquatic vegetation 0.0 0.0 16.7 100.0 0.0 0.0

Fallow soils 0.0 0.0 0.0 0.0 100.0 0.0

Urban area 0.0 0.0 0.0 0.0 0.0 100.0

Total 100.0 100 100 100.0 100.0 100.0

Figure 8: Land use\ land cover classification map according to the hypird method



Table 4: The area covered by the land use\ land cover classes

Additionally, the fish ponds to the south of the lake are mostly depending on the drainage water of Bahr El-Bakar drain to to water shortage which could have serious impacts on the quality of the fish produced from these ponds and consequently on the health of the humman consuming these fishes. Moreover, the expansion in the urban area especially to the south of the lake is also observed.

CONCLUSION Using a hybrid approach of decision tress and

neural network could improve the accuracy of the Land Use/ Land Cover classification to reach 97.3%. The designing of the decision tree relied on the use of vegetation index, band ratio and specific PCAs. Thus facilitating the designing of the neural network and increasing its performance. Furthermore, the reflection and emission data of both S2 and Landsat8, respectively, took part in this study. Nevertheless, it was also observed that the vegetation index used was ineffective in the differentiation between the two vegetation classes as their values interacted. Therefore, it is recommended that a more water sensitive index could be tested to facilitate the differentiation between the two classes. On the other hand, the differentiation between water and fish ponds classes is extremely difficult because the characteristics of the water or the lake near the shore are extremely identical to the fish ponds. Based on the results, the study area is subjected to numerous threats such as the conversion of the shoreline of the lake into fish ponds and the invasion of the aquatic vegetation which is covering vast areas of the lake especially to the south where the drainage water is discharged. The uncontrolled conversion of agriculture soils especially to the southern of the

lake into urban area is also considered a threat that affects the area of the fertile agriculture soils. CONFLICT OF INTEREST The authors declared that present study was performed in absence of any conflict of interest. AUTHOR CONTRIBUTIONS RS developed the research idea and prepared remote sensing data and analyzed the data and collected field data M helped with all the above.

Copyrights: © 2017 @ author (s). This is an open access article distributed under the terms of the Creative Commons Attribution License (CC BY 4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original author(s) and source are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

REFERENCES Agatonovic-Kustrin S and Beresford R, 2000.

Basic concepts of artificial neural network (ANN) modelling and its application in pharmaceutical research. Journal of pharmaceutical and biomedical analysis 22(5): 717-727.

Basu JK, Bhattacharyya D and Kim T, 2010. Use of Artificial Neural Network in Pattern Recognition. International Journal of Software Engineering and Its Applications 4(2): 23-34.

Beucher A, Moller AB, Greve MH, in press. Artificial neural networks and decision tree

Class Area (km2) %

Water 273.8 9.9

Fishponds 601.6 21.7

Vegetation 342.9 12.4

Aquatic

vegetation

677.4 24.5

Fallow soils 373.9 13.5

Urban area 500.4 18.1

Total 2769.9 100.0

https://creativecommons.org/licenses/by/4.0/

https://creativecommons.org/licenses/by/4.0/



classification for predicting soil drainage classes in Denmark. Geoderma in press.

Chasmer L, Hopkinson C, Veness T, Quinton W, Baltzer J, 2014. A decision-tree classification for low-lying complex land cover types within the zone of discontinuous permafrost. Remote Sens of Environ 143: 73–84.

Clevers JGPW and Gitelson AA, 2013. Remote estimation of crop and grass chlorophyll and nitrogen content using red-edge bands on Sentinel-2 and -3. International Journal of Applied Earth Observation and Geoinformation 23: 344–351.

Elnaggar AA, 2013. Spatialand Temporal Changesin Agricultural Lands Eastern Nile-Delta, Egypt.J. Soil Sci. and Agric. Eng., Mansoura Univ 4 (3): 187 – 201.

El-Wakeel SK, Wahby SD, 1970. Hydrography and chemistry of Lake Manzalah, Egypt. Archiv fur Hydrobiologie 67(2): 173-200.

Ezzeldin MM, El-Alfy KS, Abdel-Gawad HA, Abd-Elmaboud ME, 2016. Land Use Changes in the Eastern Nile Delta Region; Egypt Using Multi-Temporal Remote Sensing Techniques. International Journal of Scientific & Engineering Research 7(12): 78-98.

Friedl MA and Brodleyf CE, 1997.Decision Tree Classification of Land Cover from Remotely Sensed Data. Remote Sens of Environ 61: 399–401.

Ghulam A, Porton I, Freeman K, 2014. Detecting sub canopy invasive plant species in tropical rainforest by integrating optical and microwave (In SAR/ Pol In SAR) remote sensing data, and a decision tree algorithm. ISPRS Journal of Photogrammetry and Remote Sensing 88: 174–192.

Huete AR, 1988. Asoil-adjusted vegetation index (SAVI),Remote Sens Environ 25:295-309.

Kandrika S, Roy PS, 2008. Land use land cover classification of Orissa using multi-temporal IRS-P6 awifs data: A decision tree approach. International Journal of Applied Earth Observation and Geoinformation 10: 186–193

Kim KA, 2016. Hybrid classification algorithm by subspace partitioning through semi-supervised decision tree. Pattern Recognition 60:157–163

Morgan RS, Rahim IS, Abd El-Hady M, 2015. A Comparison of Classification Techniques for the Land Use/ Land Cover Classification Global Advanced Research Journal of Agricultural Science 4(11): 810-818.

Otukei JR, Blaschke T, 2010. Land cover change

assessment using decision trees, support vector machines and maximum likelihood classification algorithms. International Journal of Applied Earth Observation and Geoinformation 12: 27–31.

Pal M., Mather PM, 2003. An assessment of the effectiveness of decision tree methods for land cover classification. Remote Sens of Environ 86: 554–565.

Pemberty JFP, 2002. Assessment of Morph dynamic Processes and Soil Pollution as Indicators of Land Degradation. A Case Study in the Tabernas Area, South-eastern Spain. M.Sc. Thesis, International Institute for Aerospace Survey and Earth Sciences (ITC), Enschede, The Netherlands.

Punia M, Joshi PK, Porwal MC, 2011. Decision tree classification of land use land cover for Delhi, India using IRS-P6 AWiFS data. Expert Systems with Applications 38: 5577–5583.

Qi J, A. Chehbouni AR, Huete YH, Kerr S, Sorooshian ,1994. A Modified Soil Adjusted Vegetation Index . Remote Sens Environ 48:119-126.

Richards JA, Jia X, 2006. Remote Sensing Digital Image Analysis: An Introduction, 3rd ed.; Springer: New York, NY, USA, pp. 47–54.

Roy DP, Wulder MA, Loveland TR, Woodcock CE, Allen RG, Anderson MC, Helder D, Irons JR, Johnson DM, Kennedy R, Scambos TA, Schaaf CB, Schott JR, Sheng Y, Vermote EF, Belward AS, Bindschadler R, Cohen WB, Zhu Z, 2014. Landsat-8: Science and product vision for terrestrial global change research. Remote Sens of Environ 145: 154-172

Santiesteban IA, 2003. Multi-sensor image fusion applied to Morph dynamic dynamics of coastal environments in Belize. M.Sc. Thesis, International Institute for Aerospace Survey and Earth Sciences (ITC), Enschede, The Netherlands.

Sunil, B, Shanka, P, Maria, R, 2010. Per-pixel and object-oriented classification methods for mapping urban features using Ikonos satellite data. Appl Geogr 4:650–665.

Vajsová B, Aastrand PJ, 2015. New sensors benchmark report on Sentinel -2A. EUR 27674 EN. Publications Office of the European Union.doi: 10.2788/544302.

Zhang X, Treitzc PM, Chen D, Quanb, Shib L, Lia X, 2017. Mapping mangrove forests using multi-tidal remotely-sensed data and a decision-tree-based procedure. Int J Appl Earth Obs Geoinformation 62:201–214.

https://www.sciencedirect.com/science/article/pii/S003442571400042X#!

https://www.sciencedirect.com/science/journal/00344257

https://www.sciencedirect.com/science/journal/00344257/145/supp/C



Zhao Q, 2004. Learning and understanding Basedon Neural Network Trees. Computing 3(1):84-92.

Zhou ZH, Chen ZQ, 2002. Hybrid Decision Tree. Knowledge-Based Systems 15(8): 515-528.

Zhu XF, Pan Y, Zang JS, Wang S , Gu, XH, Xu C, 2007. The effects of training samples on the wheat planting area measure accuracy in TM scale (I): The accuracy response of different classifiers to training samples. J. Remote Sens 6: 826–836.

Documents

Improving land use/ land cover classification utilizing a