Environ Monit Assess (2015) 187:291 DOI 10.1007/s10661-015-4489-3

Land cover mapping based on random forest classification of multitemporal spectral and thermal images Vahid Eisavi & Saeid Homayouni & Ahmad Maleknezhad Yazdi & Abbas Alimohammadi

Received: 9 December 2014 / Accepted: 30 March 2015 # Springer International Publishing Switzerland 2015

Abstract Thematic mapping of complex landscapes, with various phenological patterns from satellite imagery, is a particularly challenging task. However, supplementary information, such as multitemporal data and/or land surface temperature (LST), has the potential to improve the land cover classification accuracy and efficiency. In this paper, in order to map land covers, we evaluated the potential of multitemporal Landsat 8’s spectral and thermal imageries using a random forest (RF) classifier. We used a grid search approach based on the out-of-bag (OOB) estimate of error to optimize the RF parameters. Four different scenarios were considered in this research: (1) RF classification of multitemporal spectral images, (2) RF classification of multitemporal LST images, (3) RF classification of all multitemporal LST and spectral images, and (4) RF classification of selected important or optimum features. The study area V. Eisavi (*) : A. M. Yazdi Remote Sensing and GIS Department, Tarbiat Modares University, Tehran, Iran e-mail: [email protected] A. M. Yazdi e-mail: [email protected] S. Homayouni Department of Geography, University of Ottawa, Ottawa, ON, Canada e-mail: [email protected] A. Alimohammadi GIS Engineering Department, Faculty of Geodesy and Geomatic Engineering, Khajeh Nasir Toosi University, Tehran, Iran e-mail: [email protected]

in this research was Naghadeh city and its surrounding region, located in West Azerbaijan Province, northwest of Iran. The overall accuracies of first, second, third, and fourth scenarios were equal to 86.48, 82.26, 90.63, and 91.82 %, respectively. The quantitative assessments of the results demonstrated that the most important or optimum features increase the class separability, while the spectral and thermal features produced a more moderate increase in the land cover mapping accuracy. In addition, the contribution of the multitemporal thermal information led to a considerable increase in the user and producer accuracies of classes with a rapid temporal change behavior, such as crops and vegetation. Keyword Land cover mapping . Multispectral . Temporal . Thermal remote sensing data . Random forest classifier

Introduction In the last decades, our knowledge of the relationship between land cover types and socio-ecological and environmental systems has increased significantly (Gillanders et al. 2008). This knowledge increases the demand for timely and reliable mapping of land cover for sustainable exploitation and management of natural resources (Elhadi et al. 2014). Remotely sensed observations are becoming an increasingly attractive alternative source of information for ground-based surveying and mapping methods. This is due mainly to the potential of repetitive data acquisition systems and their

291

Page 2 of 14

advantages of cost and time saving for larger area mapping (Lu et al. 2004; Ghosh et al. 2014). Thematic mapping of complex landscapes with various phenological patterns from satellite imagery is a particularly challenging task. Similar spectral characteristics of land cover types complicate the classification process. This is more evident for agricultural land, especially when we use monotemporal remotely sensed imagery. For example, the summer crops (e.g., cucumber, corn, tomato, etc.) can be confused with orchards and alfalfa in the image, while the autumn and spring imageries can help significantly to discriminate between these agricultural crops (Rodriguez-Galiano et al. 2012a). Urban areas can also be easily confused with the bare soils which have a high amount of surface reflectance. The inclusion of spring images allows for discrimination between urban areas and soils, which are usually covered by grass in the spring (Rodriguez-Galiano et al. 2012a). For this reason, the multitemporal observations have been employed and demonstrate their benefits for reducing the classes’ confusion. For the vegetative areas with the phenological behavior of different plants or crops in intricate landscapes, these multitemporal data are of particular importance (Langley et al. 2001; Yuan et al. 2006; Joshi et al. 2006; Rodriguez-Galiano et al. 2012a). In addition to the temporal and spectral merits of satellite observations, the thermal bands provide an important source of information about the various land cover. Thermal information is complementary to the visible and reflected infrared bands for identification of different land covers (Ehsani and Quiel 2010). For instance, Dator et al. (1998) demonstrate that using the thermal band of Landsat TM in the classification is the best way to differentiate between gypsiferous and saline soils, and Alavipanah et al. (2007) conclude that the demeanor of Landsat TM thermal and reflective bands strongly depends on the type of land cover. Accordingly, this auxiliary information may improve the land cover classification accuracy. Meanwhile, the inclusion of temporal and thermal earth observations in the classification process may increase the dimensionality of data. On the one hand, the available software may not be able to deal with this large volume of data, while on the other hand, the increased number of input variables may introduce additional complexity regarding the increase of computational time (Bellman 2003). Feature selection (FS) is an important preprocessing step in many machine learning applications. It is used

Environ Monit Assess (2015) 187:291

for selecting the smallest subset of relevant features that builds robust learning models (Blum and Langley 1997; Guyon and Elisseeff 2003; Saeys et al. 2007). The interest in using a feature selection method is that the data contains redundant information. Thus, FS may be able to cope with the problem of data redundancy and correlation. FS can help to increase classification models by alleviating the effect of dimensionality, speed up the training process, and increase the model interpretability (Rodriguez-Galiano et al. 2012a). Consequently, it is necessary not only to obtain more accurate maps but also to understand which spectral, thermal, or temporal variables are most relevant in the classification process. Among the different methods, in this paper, we consider the application of the random forest (RF) algorithm for feature selection due to its interesting properties, such as high accuracy and robustness against overfitting the training data (Diaz-Uriarte et al. 2006). RF has recently been proposed and used for improving land cover mapping from remotely sensed images (Pal 2005; Ham et al. 2005; Gislason et al. 2006; Lippitt et al. 2008; Rogan et al. 2008; Ghimire et al. 2010; Rodriguez-Galiano et al. 2012b; Elhadi et al. 2014; Ghosh et al. 2014; Puissant et al. 2014; Van Beijma et al. 2014). It provides an approach for assessing the importance of features or predictors, which can be seen as a useful parameter for studying the role of each temporal, spectral, or thermal feature for land cover discrimination analysis (Breiman 2001). For land cover classification, researchers have employed various remotely sensed data (French et al. 2008; Wulder et al. 2008; Stefanov and Netzband 2005), multitemporal (Rodriguez-Galiano et al. 2012a; Nitze et al. 2014), and thermal observations (RodríguezGaliano et al. 2012). Nevertheless, all of this temporal, spectral, and thermal information has rarely been used and analyzed together in the classification process. As a result, the main objective of this study is to evaluate the potential of multitemporal, multispectral, and thermal Landsat 8 observations for mapping a heterogeneous landscape using the reliable and robust RF classifier.

Materials and methods Study area and remote sensing data The study area is Naghadeh City and its surrounding area, located in the West Azerbaijan Province, northwest

Environ Monit Assess (2015) 187:291

of Iran (see Fig. 1). The city is located in the Gadar River valley, 23 km south of Lake Urmia, at an elevation of 1300 m above sea level. The climate of the West Azarbaijan province is semi-arid and is characterized by hot and dry summers, and wet and cold winters. The region’s economy is mainly based on agricultural activities, particularly the production of fruit, grain, and timber. The dominant land cover types in the region include the built-up areas, bare lands, farmlands, grasslands, orchards, and bodies of water. In this research, the spring, summer, and autumn images are used to cope with confusion regarding complex phenological behavior. Seven Landsat 8 scenes of the same area are acquired. The images are from April 28th, June 15th, July 17th, August 11th, September 12th, October 14th, and November 15th in 2013. These multitemporal data are used for classification of the land cover in the study area. Extensive anthropogenic influences and phenological patterns help to identify 13 different land cover classes within the study area (see Table 1). For the objective of training and testing the classifier efficiency, the reference data are collected from

Page 3 of 14 291

agricultural crops by field observations using both cadastral maps and global positioning system (GPS). The reference data (e.g., range, urban, road, water, and wetland vegetation) are obtained by field surveying during the growing season. The data is then randomly divided into the training (40 %) and testing (60 %) datasets (see Table 1). In order to obtain the real surface reflectance and reduce the atmospheric and environmental effects from the images in this study, we have used FLAASH® (Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes) in ENVI software. Image analysis methodology Figure 2 illustrates the methodology used in this research. This section consists of data preparation, image preprocessing, LST calculation, and RF classification. Spectral processing is comprised of principal component analysis (PCA) which is applied to each spectral dataset (namely, t1,…, t7) to segregate the noise components and to reduce the data dimensionality. Thermal bands contribute to calculation of LST. The first and

Fig. 1 The natural color 3D representation of the study area: Naghadeh City and its surrounding region

291

Environ Monit Assess (2015) 187:291

Page 4 of 14

Table 1 Land cover classes, training, and test samples Class ID

Class name

Training pixel

Test pixel

1

Apple

661

992

2

Vineyards

611

916

3

Grassland

583

875

10

4

Urban

360

541

11

Summer crops

259

388

5

Wetland vegetation

337

506

12

Corn

487

730

6

Water

596

842

13

Bare land

240

360

7

Wheat

607

910

second principal components (PCs) from all spectral datasets (namely, PC1t1, PC2t1,…, PC1t7, PC2t7) and the multitemporal LST features are used as the inputs in decision tree induction. In order to evaluate the potential of multitemporal, multispectral, and thermal images of Landsat 8 observations for land cover mapping, we define four different scenarios as follows: (1) RF classification of the spectral images, (2) RF classification of LST images, (3) RF classification of stacked LST and spectral images, and (4) RF classification of selected important or optimum features.

Fig. 2 Overview of analysis framework. Note that FLAASH® was applied for atmospheric correction of spectral bands, and the equation of atmospheric correction (Eq. 2) was used for thermal bands (http://www. atmcorr.gsfc.nasa.gov)

Class ID

Class name

Training pixel

Test pixel

8

Salt area

428

643

9

Road

206

328

Fallow

437

656

Land surface temperature (LST) Radiometric conversion of digital numbers (DN) to the physical variables, such as temperature, requires radiometric calibration (Wukelic et al. 1989) and atmospheric correction (Cooper and Asrar 1989). DN values are converted to the at-sensor radiance using the gain and bias coefficients of the Landsat sensor (Lamaro et al. 2013): Lλ ¼ gain*DN þ bias

ð1Þ

Environ Monit Assess (2015) 187:291

Page 5 of 14 291

where Lλ is the uncorrected spectral radiance at λ wavelength, DN is the digital number, and gain and bias are the calibration parameters of the sensor. Gains and biases of Landsat 8 thermal bands (i.e., both 11th and 12th bands) are 3.3420E−04 and 0.1, respectively. The equation for atmospheric correction is (Srivastava et al. 2009): LλðT S Þ ¼

Lλ −Lλup 1−ελ − *Lλdown t*ελ ελ

ð2Þ

where LλðT S Þ is the corrected land surface radiance and Lλ is the uncorrected spectral radiance calculated in Eq. (1), Lλup is the upwelling radiance, t is the atmospheric transmissivity, ελ is the water emissivity, and Lλdown is the downwelling radiance. Lλdown, Lλup, and t are atmospheric parameters, obtained from Atmospheric Correction Parameter Calculator (http://www.atmcorr.gsfc.nasa. gov), which is based on the MODTRAN simulator (Barsi et al. 2003). In addition, Normalized Difference Vegetation Index (NDVI) threshold method is applied to determine the surface emissivity (ελ) This parameter is defined based on three classes of NDVI values (Sobrino et al. 2004): 1. NDVI0.5: represents areas with dense vegetation cover. Emissivity of vegetation cover is assumed as: εVeg =0.99 3. 0.2≤NDVI≤0.5: the pixels in this case are mixture of soil and vegetation cover. Emissivity εmix is calculated according to the following equation: ε ¼ εveg *Pv þ εsoil *ð1−Pv Þ

ð3Þ

where Pv is proportion of vegetation cover and is computed as:  Pv ¼

NDVI−NDVImin NDVImax −NDVImin

2 ð4Þ

 ln

K2 K1 LλðT S Þ þ 1



Random forest classifier Ensemble learning algorithms (e.g., RF or Boosting) demonstrate their superiority and robustness to noise when compared to the single classifiers (Breiman 1996; Dietterich 2000). A random forest classifier (RFC) is an ensemble of classification trees, where each tree contributes with a single vote for the assignment of the most frequent class to the input data (Breiman 2001; Guo et al. 2011; Rodríguez-Galiano et al. 2012). A RFC can handle thousands of input features or input bands. It uses a random vector, sampled independently from the input vector in the division of every node, instead of using the best variables, which reduces the generalization error (Breiman 1999). RF builds many binary classification trees (ntree) using several bootstrap samples with replacements drawn from the original observations (Elhadi et al. 2014). The samples that are left out from the bootstrap sample are called out-of-bag (OOB) samples. The OOB samples help to evaluate the misclassification error and the variable importance estimation. At each node, a given number of input variables (mtry) are randomly chosen from a random subset of the features. Then, the best split is calculated by using only this subset of features. No pruning is performed to ensure lower diversity between individual trees. As a result, all trees are grown maximally and low bias is achieved (Genuer et al. 2010; Lin et al. 2010; Breiman 2001). The RFC uses the Gini index in the variable selection process for decision tree induction, whereby it measures the impurity of a variable (Breiman et al. 1984). For a given training set T, selecting one case (pixel) at random and saying that it belongs to some class Ci, the Gini index can be modeled as: XX

In this equation, NDVImin =0.2  and  NDVImax =0.5 Corrected surface radiance LλðT S Þ is converted into the surface temperature (TS) using Eq. 5 TS ¼

where K1 is the calibration constant (774.89 and 480.89 for bands 10 and 11, respectively), LλðT S Þ is the corrected surface radiance, and K2 is another calibration constant (1321.08 and 1201.14 for bands 10 and 11, respectively).

ð5Þ

    ð f ðC i ; T Þ=jT jÞ f C j ; T =ðjT jÞ

ð6Þ

j≠i

where f(Ci,T)/|T| is the probability that the selected case belongs to class Ci (Pal 2005). The default number of trees (ntree) is 100, while the default value for the number of variables (mtry) is the square root of the total number of bands in the image pffiffiffi data, i.e., p (Breiman 2001). In this paper, we use the

291

Page 6 of 14

grid search approach based on the OOB error values to optimize the ntree and mtry parameters.

Results and discussions In this section, the classification results of four proposed scenarios, by means of RFC, are presented, evaluated, and discussed. First scenario: multitemporal spectral data The normalized spectral band importance estimated for the whole training data in the first scenario and the result of grid search procedure are shown in Fig. 3. As can be seen, the mtry value of 2 combined with an ntree of 500 resulted in the lowest OOB error rate (i.e., 0.1). Consequently, these optimum values were selected to train the RF algorithm to classify the land cover classes. Regarding the spectral information in Fig. 3, PC1 of June 15th, PC1 of April 28th, and PC1 of November Fig. 3 a Optimization of mtry and ntree parameters using grid search procedure. The OOB sample was used to determine the error rate for the different combinations of mtry and ntree. b Normalized spectral feature importance based on Gini index (scenario 1)

Environ Monit Assess (2015) 187:291

15th images have a larger relative importance than the rest of the spectral bands. This may be due to the phenological sequences of vegetative lands which lead to better class separability. In June and April, the majority of the study area is under agricultural cultivation (e.g., wheat and corn) since the summer crops are not cultivated until then. Hence, these months can contribute to separate wheat and corn classes from the other the summer crops. Vineyards and apple orchards are in full greenness and production during the growing season. In November, the harvest has ended, thus orchards can be easily distinguished from other agricultural lands. To evaluate the performance of RFC results of this scenario, the overall accuracy, producer and user accuracies, F1 accuracy (Puissant et al. 2014), and the conditional kappa coefficient (Congalton and Green 2008) are reported. The accuracy assessment of first scenario, i.e., spectral bands, is presented in Table 2. In this scenario, the low accuracy for some crops, such as corn and vineyards, may be due to the spectral and phonological similarity of plantations. Moreover,

Environ Monit Assess (2015) 187:291 Table 2 Per-class accuracies of first scenario classification results based on independent test data set. (i.e. multitemporal spectral features). The accuracies include: overall accuracy, kappa coefficient, conditional kappa, F1 accuracy, user’s and producer’s accuracies

Class

1. Apple

Page 7 of 14 291 Overall accuracy=86.5 %

Kappa coefficient= 0.8

User’s

Producer’s

F1

Conditional kappa

86.18

78.10

81.94

0.85

2. Vineyards

61.44

78.15

68.80

0.60

3. Range

78.83

69.00

73.58

0.78

4. Urban

81.19

74.03

77.44

0.80

5. Wetland vegetation

87.40

64.80

74.42

0.86

6. Water

98.11

98.02

98.06

0.97

7. Wheat

83.95

80.15

82.00

0.82

8. Salt area

87.30

83.02

85.10

0.86

9. Road

41.20

65.40

50.55

0.40

10. Fallow

81.56

79.50

80.73

0.80

11. Summer crops

78.18

82.13

80.10

0.77

12. Corn

77.21

78.00

77.60

0.75

13. Bare land

60.39

77.00

67.69

0.59

there is confusion between bare land, salt area, and rangeland. Kappa accuracy (KA) of apple, wetland vegetation, water, and salt area were more than 85 %, while very low KA was observed in road, vineyards, and bare classes. Despite using multitemporal Landsat images, many pixels were still misclassified due to their spectral similarity. The results of this scenario suggest that there is some deficiency with regard to the use of only spectral bands for land cover classification. Second scenario: multitemporal thermal data The thermal features include far fewer variables than the spectral dataset, which simplifies the classification trees in RF. The optimum RF parameters and the normalized LST variable importance in the second scenario are shown in Fig. 4. The highest and lowest OOB errors in different mtry and ntree are indicated in this graph. LST of the June 15th data has the most outstanding value of the contribution of thermal feature to the classification. LST of the November 15th data has the lowest importance in classification results. According to Figs. 3b and 4b, we can conclude that June is the most informative month in regard to LST features and spectral bands, which can be the most influential time in the classification process. Table 3 shows the confusion matrices and accuracy assessment of LST variables classification,

namely the second scenario. Although multitemporal LST variables show lower overall accuracy than spectral variables, they show higher accuracies in some specific classes and promising land cover classification capabilities. There are spectral confusions, however, between vineyards and range, road and fallow, and bare land and salt area. The LST features result in a reduction of the commission and omission errors and improved bare land, fallow, road, rangeland, and vineyards classification accuracies. The remaining classes show the highest values in both user’s and producer’s accuracies in the first scenario. Third scenario: all multitemporal spectral and thermal features Figure 5 shows the variable importance graph of the third scenario. The combination of spectral and thermal features changes the order of variable importance. The combination of mtry value of 4 and an ntree value of 2500 produce the lowest OOB error in this scenario. The first components of June 15th and April 28th are the first and second significant variables in this scenario, respectively. An analysis of RF’s variables importance indicates that the spectral data are more valuable for land cover mapping than the LST data. Nevertheless, the land cover information provided by both multitemporal

291

Environ Monit Assess (2015) 187:291

Page 8 of 14

Fig. 4 a Optimization of mtry and ntree parameters using grid search procedure. The OOB sample was used to determine the error rate for the different combinations of mtry and ntree. b Normalized LST feature importance based on Gini index (scenario 2)

thermal and spectral criteria can be complementary rather than redundant. In this scenario, the RFC is applied to 21 features in total and has achieved the high accuracies. Table 4 illustrates the accuracy analysis of classification results of multispectral bands, and LST features.

Table 3 Per-class accuracies of second scenario classification results based on independence test data set (i.e. multitemporal thermal features). The accuracies include: overall accuracy, kappa coefficient, conditional kappa, F1 accuracy, user’s and producer’s accuracies

Class

The overall accuracy estimated for this scenario is 90.63 %. The inclusion of LST information has increased the classification accuracy by almost 4 %, in comparison to the first and 8 % to the second scenarios, respectively. Similar improvements in land cover classification accuracy, using thermal information, have

Overall accuracy=82.3 %

Kappa accuracy=0.7

User’s

Producer’s

F1

Conditional kappa

1. Apple

83.12

56.50

67.31

0.80

2. Vineyards

74.53

63.20

68.98

0.75

3. Range

88.64

73.50

80.14

0.87

4. Urban

74.93

72.80

73.64

0.73

5. Wetland vegetation

41.07

59.70

48.88

0.49

6. Water

98.34

98.49

98.94

0.97

7. Wheat

80.49

79.20

79.79

0.79

8. Salt area

77.43

81.04

79.88

0.76

9. Road

48.24

50.52

49.68

0.49

10. Fallow

84.93

74.20

79.24

0.83

11. Summer crops

49.14

91.51

63.14

0.48

12. Corn

64.77

72.92

68.38

0.63

13. Bare land

65.18

44.26

52.52

0.64

Environ Monit Assess (2015) 187:291

Page 9 of 14 291

Fig. 5 a Optimization of mtry and ntree parameters using grid search procedure. The OOB sample is used to determine the error rate for the different combinations of mtry and ntree. b Normalized spectral and LST features importance based on Gini index (scenario 3)

been achieved in other studies. For example, RodríguezGaliano et al. (2012) reported a 10 % increase in accuracy for land cover classification when combining thermal information with multispectral bands. In this scenario, the fallow, bare land, and salt land covers seem to be misclassified. Based on the comparison of user and producer accuracies, the omission and commission errors corresponding to similar spectral and phenological behavior in agricultural categories, such as vineyards and apple, are reduced using multitemporal thermal features. This indicates that the LST features are helpful for improving the land cover classification. The classes such as road, vineyards, and range show the greatest improvement in KA and are effectively classified.

Fourth scenario: optimum features In the third scenario, we apply the RFC to 21 features (i.e., all multitemporal spectral and thermal data), despite the fact that this large a number of features may be highly correlated and generate the so-called curse of dimensionality problem in classification. Meanwhile, the knowledge of which spectral and thermal features are more relevant in the classification process can be used to decrease the dimensionality and increase the mapping accuracy. To evaluate the effects of these important or optimum features on the classification efficiency, we use these variables as the inputs of classification, in a consecutive way. In other words, at first, we use only the most important variable (i.e., PC1 15th

291

Environ Monit Assess (2015) 187:291

Page 10 of 14

Table 4 Per-class accuracies of third scenario classification results based on independence test data set (i.e. multitemporal spectral and thermal features). The accuracies parameters include: overall accuracy, kappa coefficient, conditional kappa, F1 accuracy, user’s and producer’s accuracies

Class

1. Apple

Overall accuracy=90.63 %

Kappa accuracy=0.86

User’s

Producer’s

F1

Conditional kappa

86.18

85.17

85.67

0.85

2. Vineyards

67.29

85.40

75.27

0.66

3. Range

86.36

83.88

85.10

0.85

4. Urban

83.93

80.15

81.99

0.82

5. Wetland vegetation

87.48

63.87

73.83

0.86

6. Water

98.05

99.41

98.72

0.97

7. Wheat

82.25

80.67

81.45

0.81

8. Salt area

86.48

82.07

84.22

0.85

9. Road

63.23

75.40

68.78

0.62

10. Fallow

76.30

84.43

80.16

0.75

11. Summer crops

74.60

81.71

77.99

0.74

12. Corn

76.14

76.58

76.36

0.75

13. Bare land

56.33

62.86

59.41

0.55

June). Then, in the second round of classification, we use both first and second important variables (i.e., PC1 15th June and PC1 28th April), and continued to the last classification that use all the available multitemporal spectral and thermal features. Figure 6 shows the overall accuracy of all these classification steps. According to Fig. 6, in the first step, the most important band (i.e., PC1 15th June) has an overall accuracy of 65 % for all land cover classes in this study area. In the next step, by adding the second important band (i.e., PC1 28th April), the overall classification accuracy has significantly increased. These accuracy improvements continue until the combination of seven important features. Adding more features beyond the seventh will result in slight fluctuations and even decrease the classification accuracy. Accordingly, a combination of the most important seven features yields the most satisfactory results with respect to other combinations. Fig. 6 The overall accuracies of different combinations of features classification

The first seven numbers of the selected features, in order of importance, are PC1 (15th June), PC1 (28th April), LST (17th July), LST (15th June), LST (11th August), PC2 (15th June), and PC1 (17th July). These important variables are appropriately spread over 4 months, which can efficiently represent phenological orders and the growing seasons of the study area. As a result, in trained RFs, more weights are given to temporal variables since the temporal and phenological variables are able to yield higher accuracies of vegetation classes. The grid search method’s results of optimum seven features are indicated in Fig. 7. We choose mtry value of 2000 and ntree value of 3 for training RF classifier. Table 5 depicts the accuracy assessment for selected features combination. The overall accuracy of this scenario, i.e., 91.82 %, is slightly higher than third scenario. Low differences between the accuracies of these two

Environ Monit Assess (2015) 187:291

Page 11 of 14 291

Fig. 7 Optimization of mtry and ntree parameters using grid search procedure. The OOB sample was used to determine the error rate for the different combinations of mtry and ntree (scenario 4)

scenarios suggest that the selected important variables are not able to significantly improve the land cover classification. The urban and wheat classes have the high KA of 84.12 and 82.6 %, respectively, so the differences between user and producer accuracies have increased. Nonetheless, this scenario demonstrates that RF can be effectively applied to feature selection and dimensionality reduction in any image analysis application. Figure 8 illustrates the land cover maps obtained by four designed and implemented scenarios in this paper. These maps represent 13 land cover classes. As can be

Table 5 Per-class accuracies of fourth scenario classification results (selected features) based on independence test data set. The accuracies parameters include: overall accuracy, kappa coefficient, conditional kappa, F1 accuracy, user’s and producer’s accuracies Class

Overall accuracy= 91.8 % User’s

Kappa accuracy=0.90

1. Apple

85.25

80.1

2. Vineyards

67.12

84.4

74.77 0.66

3. Range

86.46

83.61

85.01 0.85

4. Urban

85.93

80.42

83.08 0.84

5. Wetland vegetation 6. Water

86.51

60.87

71.45 0.86

98.11

95.41

96.74 0.97

Producer’s F1

Conditional kappa

82.59 0.84

7. Wheat

83.51

81.47

82.47 0.83

8. Salt area

86.12

83.27

84.67 0.85

9. Road

63.87

73.2

68.21 0.62

10. Fallow

76.3

84.43

80.15 0.75

11. Summer crops 12. Corn

75.65

80.05

77.78 0.74

75.92

73.74

74.81 0.75

13. Bare land

56.01

60.41

58.12 0.55

seen, there are evident differences between some classes such as water, summer crops, wetland vegetation, and vineyards.

Conclusion This paper evaluates the potential of multitemporal spectral and thermal remote sensing observations for land cover classification using a relatively novel technique, the RFC. The study area, Naghadeh city and its surrounding agricultural and natural lands, has specific characterizations and various natural and manmade surfaces types. This study provides new insights on the performance of multitemporal Landsat 8 thermal and spectral imagery using RFC in mapping land cover. RFC offers a wide range of decision-making alternatives regarding the variable importance, which guides users toward the important variables or features. Using the RF importance measures reduced the dimensionality of input features. The RF importance variables showed how the multitemporal spectral and thermal features provide the greatest influence on class separability in the study area. LST features and spectral bands were compared in different scenarios for land cover classification. A quantitative assessment of the results demonstrated that the important features present the largest increase in the overall class separability, while the spectral and LST features produce a more moderate increase in land cover mapping accuracy. In other words, selecting important multitemporal spectral bands and thermal features using RFC to produce land cover maps helps to overcome the difficulty of discriminating between classes which have close spectral characteristics or exhibit similar phenology. Incorporation of the multitemporal LST information leads to a considerable decrease in the omission and commission errors. In this study, the LST information

291

Page 12 of 14

Environ Monit Assess (2015) 187:291

Fig. 8 Land cover maps resulting from RF classifications of a spectral features, b LST features, c all spectral and thermal features, and d optimum/important features

shows important discriminating capability between different types of vegetation. This was mainly because most of the agricultural lands have relatively different surface temperature than the other vegetation and land covers, such as salt area and bare lands. This research clearly shows that LST features are important in the classification of certain classes. These results demonstrate that the RF model is able to efficiently map the land cover of highly fragmented areas. In addition, the variable importance, as an output of RFC, can be used to determine the optimum date during the image acquisition period. This study also highlights the significance of multitemporal thermal images in addition to multispectral images for land cover classification purposes. This research confirms that RF has some interesting results, such as decreased variable input dimensionality and higher quality land cover classification. Feature importance measure is another advantage of random forests which was successfully applied in this study. Further, RF is an easy classifier which only requires two parameters to be set. Moreover, RFC may offer insights to assess other parameters, such as spatial information in land cover mapping.

A key outcome of our study is that integration of auxiliary information (e.g., spectral, temporal, and thermal data) may lead to redundancy. However, FS can reduce this problem and consequently increase model performance. For future study, we plan to apply this approach to different landscape types in order to test its robustness for mapping agricultural and urban areas. Acknowledgments The authors would like to acknowledge the USGS for providing Landsat 8 imagery, and the R development team and the EnMAP-Box team for making these software packages publicly available.

References AlaviPanah, S. K., Komaki, C. B., Goorabi, A., & Matinfar, H. R. (2007). Characterizing land cover types and surface condition of Yardang Region in Lut desert (Iran) based upon Landsat satellite images. World Applied Sciences Journal of Arid Environments, 2, 212–228. Barsi, J. A., Barker, J. L., Schott, J. R. (2003). An atmospheric correction parameter calculator for a single thermal band

Environ Monit Assess (2015) 187:291 earth-sensing instrument. Paper presented at the IGARSS03, Toulouse, France. Bellman, R. (2003). Dynamic programming (2nd ed.). Mineola: Dover Publications. Blum, A. L., & Langley, P. (1997). Selection of relevant features and examples in machine learning. Artificial Intelligence, 97, 245–271. Breiman, L. (1996). Bagging predictors. Machine Learning, 24, 123–140. Breiman, L., Friedman, J. H., Olshen, R. A., & Stone, C. J. (1984). Classification and regression trees. Monterey: Wadsworth. Breiman, L. (1999). Random forests—random features. Technical Report 567, Statistics Department, University of California, Berkeley, ftp://ftp.stat.berkeley.edu/pub/users/breiman. Breiman, L. (2001). Random forests. Machine Learning, 45, 5–32. Congalton, R. C., & Green, K. (2008). Assessing the accuracy of remotely sensed data principles and practices (2nd ed.). New York: CRC press Taylor & Francis Group. Cooper, D. I., & Asrar, G. (1989). Evaluating atmospheric correction models for retrieving surface temperatures from the AVHRR over a tall grass prairie. Remote Sensing of Environment, 27, 93–102. Dator, J., Goossens, R., & Van Ranst, E. (1998). The use of remote sensing to map gypsiferous soils in the Ismailia Province (Egypt). Geoderma, 87, 47–56. Diaz-Uriarte, R., Alvarez de Andres, S. (2006). Gene selection and classification of microarray data using random forest. BMC Bioinformatics, 7. Dietterich, T. G. (2000). An experimental comparison of three methods for constructing ensembles of decision trees: bagging, boosting, and randomization. Machine Learning, 40, 139–157. Ehsani, A. H., & Quiel, F. (2010). Efficiency of Landsat ETM+ thermal band for land cover classification of the biosphere reserve BEastern Carpathians^ (Central Europe) using SMAP and ML algorithms. International Journal of Environmental Research and Public Health, 4, 741–750. Elhadi, A., Onisimo, M., John, O., & Elfatih, M. A. R. (2014). Land-use/cover classification in a heterogeneous coastal landscape using RapidEye imagery: evaluating the performance of random forest and support vector machines classifiers. International Journal of Remote Sensing, 35, 3440– 3458. French, A., Schmugge, T., Ritchie, J., Hsu, A., Jacob, F., & Ogawa, K. (2008). Detecting land cover change at the Jornada experimental range, New Mexico with ASTER emissivities. Remote Sensing of Environment, 112, 1730– 1748. Genuer, R., Poggi, J. M., & Tuleau-Malot, C. (2010). Variable selection using random forests. Pattern Recognition Letters, 31, 2225–2236. Gillanders, S. N., Coops, N. C., Wulder, M. A., Gergel, S. E., & Nelson, T. (2008). Multitemporal remote sensing of landscape dynamics and pattern change: describing natural and anthropogenic trends. Progress in Physical Geography, 32, 503–528. Gislason, P. O., Benediktsson, J. A., & Sveinsson, J. R. (2006). Random forests for land cover classification. Pattern Recognition Letters, 27, 294–300. Ghimire, B., Rogan, J., & Miller, J. (2010). Contextual land-cover classification: incorporating spatial dependence in land-cover

Page 13 of 14 291 classification models using random forests and the Getis statistic. Remote Sensing Letters, 1, 45–54. Ghosh, A., Sharma, R., & Joshi, P. K. (2014). Random forest classification of urban landscape using Landsat archive and ancillary data: combining seasonal maps with decision level fusion. Applied Geography, 48, 31–41. Guo, L., Chehata, N., Mallet, C., & Boukir, S. (2011). Relevance of airborne lidar and multispectral image data for urban scene classification using random forests. ISPRS Journal of Photogrammetry and Remote Sensing, 66, 56–66. Guyon, I., & Elisseeff, A. (2003). An introduction to variable and feature selection. Journal of Machine Learning Research, 3, 1157–1182. Ham, J., Chen, Y., Crawford, M. M., & Ghosh, J. (2005). Investigation of the random forest framework for classification of hyperspectral data. IEEE Transactions on Geoscience and Remote Sensing, 43, 492–501. Joshi, P. K., Roy, P. S., Singh, S., Agrawal, S., & Yadav, D. (2006). Vegetation cover mapping in India using multi-temporal IRS wide field sensor (WiFS) data. Remote Sensing of Environment, 103, 190–202. Lamaro, A. A., Marinelarena, A., Torrusio, & Sala, S. (2013). Water surface temperature estimation from Landsat7 ETM+ thermal infrared data using the generalized single-channel method: case study of Embalse del RıoTercero (Cordoba, Argentina). Advances in Space Research, 51, 492–500. Langley, S. K., Cheshire, H. M., & Humes, K. S. (2001). A comparison of single date and multitemporal satellite image classifications in a semi-arid grassland. Journal of Arid Environments, 49, 401–411. Lin, X., Sun, L., Li, Y., Guo, Z., Li, Y., Zhong, K., Wang, Q., Lu, X., Yang, Y., & Xu, G. (2010). A random forest of combined features in the classification of cut tobacco based on gas chromatography fingerprinting. Talanta, 82, 1571–1575. Lippitt, C. D., Rogan, J., Li, Z., Eastman, J. R., & Jones, T. G. (2008). Mapping selective logging in mixed deciduous forest: a comparison of machine learning algorithms. Photogrammetric Engineering and Remote Sensing, 74, 1201–1211. Lu, D., Mausel, P., Brondízio, E., & Moran, E. (2004). Change detection techniques. International Journal of Remote Sensing, 25, 2365–2401. Nitze, I., Brian Barrett, B., & Cawkwell, F. (2014). Temporal optimization of image acquisition for land cover classification with random forest and MODIS time-series. International Journal Applied Earth Observation and Geoinformation, 34, 136–146. Pal, M. (2005). Random forest classifier for remote sensing classification. International Journal of Remote Sensing, 26, 217–222. Puissant, A., Rougiera, S., & André, S. (2014). Object-oriented mapping of urban trees using random forest classifiers. International Journal of Applied Earth Observation and Geoinformation, 26, 235–245. Rodriguez-Galiano, V. F., Chica-Olmo, V., Abarca-Hernandez, F., Atkinson, P. M., & Jeganathan, C. (2012a). Random forest classification of Mediterranean land covers using multiseasonal imagery and multi-seasonal texture. Remote Sensing of Environment, 121, 93–107. Rodriguez-Galiano, V. F., Ghimire, B., Rogan, J., Chica-Olmo, M., & Rigol-Sanchez, J. P. (2012b). An assessment of the effectiveness of a random forest classifier for landcover

291

Page 14 of 14

classification. ISPRS Journal of Photogrammetry and Remote Sensing, 67, 93–104. Rodríguez-Galiano, V. F., Ghimire, B., Pardo-Igúzquiza, E., Chica-Olmo, M., & Congalton, R. G. (2012). Incorporating the downscaled Landsat TM thermal band in land-cover classification using random forest. Photogrammetric Engineering and Remote Sensing, 78, 129–137. Rogan, J., Franklin, J., Stow, D., Miller, J., Woodcock, C., & Roberts, D. (2008). Mapping land-cover modifications over large areas: a comparison of machine learning algorithms. Remote Sensing of Environment, 112, 2272–2283. Saeys, Y., Inza, I., & Larrañaga, P. (2007). A review of feature selection techniques in bioinformatics. Bioinformatics, 23, 2507–2517. Sobrino, J. A., Jiménez, M., & Paolinib, C. J. (2004). Surface temperature retrieval from LANDSAT TM 5. Remote Sensing of Environment, 90, 434–440. Srivastava, P. K., Majumdar, T. J., & Bhattacharya, A. K. (2009). Surface temperature estimation in Singhbhum Shear Zone of India using Landsat-7 ETM+ thermal infrared data. Advances in Space Research, 43, 1563–1574.

Environ Monit Assess (2015) 187:291 Stefanov, W. L., & Netzband, M. (2005). Assessment of ASTER land cover and MODIS NDVIData at multiple scales for ecological characterization of an arid urban center. Remote Sensing of Environment, 99, 31–43. Van Beijma, S., Comber, A., & Lamb, A. (2014). Random forest classification of salt marsh vegetation habitats using quadpolarimetric airborne SAR, elevation and optical RS data. Remote Sensing of Environment, 149, 118–129. Wukelic, G. E., Gibbons, D. E., Martucci, I. M., & Foote, H. P. (1989). Radiometric calibration of Landsat thematic mapper thermal band. Remote Sensing of Environment, 28, 339–347. Wulder, M. A., White, J. C., Goward, S. N., Masek, J. G., Irons, J. R., Herold, M., Cohen, W. B., Loveland, T. R., & Woodcock, C. E. (2008). Landsat continuity: issues and opportunities for landcover monitoring. Remote Sensing of Environment, 112, 955–969. Yuan, F., Sawaya, K. E., Loeffelholz, B. C., & Bauer, M. E. (2006). Land cover classification and change analysis of the Twin Cities (Minnesota) Metropolitan area by multitemporal Landsat remote sensing. Remote Sensing of Environment, 98, 317–328.

Land cover mapping based on random forest classification of multitemporal spectral and thermal images.

Thematic mapping of complex landscapes, with various phenological patterns from satellite imagery, is a particularly challenging task. However, supple...
3MB Sizes 2 Downloads 5 Views