Land cover is defined as the observed biophysical state of the earth’s surface, and is largely described by the presence or absence of various vegetation types [1
]. In contrast, land use normally refers to the arrangements, activities and inputs people engage in a certain land cover type to produce, change or maintain it [2
]. As previous studies reported, land cover information is a fundamental variable for many hydrological and climate studies. Land cover characteristics have close links to the human and physical environments, also govern and affect many environmental variables [3
], including surface roughness, albedo, moisture availability, mechanisms for runoff generation [4
], and water quality [5
]. Therefore, accurate land-cover mapping becomes essential for modeling and understanding these biogeophysical properties of the land surfaces.
Remote sensing provides an effective way to depict land cover as it produces a map-like representation of the Earth’s surface that is spatially continuous and highly consistent, as well as available at a range of spatial and temporal scales [1
]. The Landsat satellites have monitored the Earth's terrestrial surfaces for about 40 years [6
], from which the long, consistent and free record allows scientists to study the current and also the past land surface patterns. Because of that, Landsat data are widely applied in land cover classification and monitoring on a regional or global scale. Numerous studies have proved the usefulness of Landsat imagery in agricultural land cover classification [7
], forest dynamics monitoring [8
], urban land use classification [9
], other land cover dynamics or land use land cover (LULC) change detection [6
]. Other satellite products such as the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) sensor imagery have also been widely used for regional scale land cover classification [12
] or land cover change detection [15
In most cases, the LULC classification is based on the multispectral characteristics and/or the multi-temporal biological properties of the Earth’s surface. In previous studies, numerous efforts have been made to improve the classification accuracy by constructing different spectral features, developing new methods, or integrating multi-source data for the single or time series of images. Lunetta and Balogh [17
], for example, evaluated the identification of wetlands with the bands 2 to 5 of the single-date and multi-temporal Landsat 5 images. The overall accuracy (OA) was 69% of the single-date image compared to 88% from the two-date images with a significant increase in the Kappa test statistics. Murai and Omatu [18
] proposed a pattern classification method which integrates the advantages of both the neural network and knowledge-based system. The single Landsat 5 TM image with the bands 3 to 5 was used and they found that the misclassification can be revised more easily because of introducing the geographical knowledge into the system. Maxwell et al.
] introduced an automated approach to classify four land cover types using only the bands 2 and 4 from Landsat MSS with 92.2% OA.
Langley et al.
] compared the single-date imagery and multi-temporal images for land cover classification with the bands 3 to 5 of TM image. They concluded that the multi-temporal images have improved the accuracies of some landscapes but the single-date image may provide a reliable vegetation cover map in semi-arid environments. Saadat et al.
] utilized two single-date Landsat ETM+ image without the thermal bands for LULC classification in Iran with OA of 95% and 82% respectively for the late summer image and the spring image. He recommended that when the satellite image is limited the late summer image would be most suitable for the LULC classification. Guerschman et al.
] also suggested that, if possible, three images (spring, early summer, late summer) be used in the identification of agricultural types. Yuan et al.
] used multi-temporal TM images from 1986 to 2002 to monitor the LULC dynamic with the average OA of 94% and proved the potential of multi-temporal Landsat data for accurate and economic land cover change analysis. Besides the different band combinations, the normalized difference vegetation index is also commonly used for the LULC change detection with the multi-temporal images [6
]. For both single-date image and multi-temporal images, several studies focus on the algorithm development such as nearest neighbor (NN) [28
], modified NN [7
], random forest classifier [30
], and rule-based classification [33
], which all provide accurate land cover classification maps.
However, surprisingly, the thermal information provided by many of the satellite platforms has rarely been used for land cover classification [35
]. Thermal remote sensing allows for the continuous representation of land surface temperature [36
], which is widely used for the monitoring of urban climate [37
], the modeling of the hydrological cycle [38
], vegetation monitoring [39
] and mapping land surface energy and water vapor fluxes [40
]. Although the spatial resolution of the thermal band is coarser when compared to the visible bands of the same satellite, the thermal information may contain valuable information related to the spatial variations of land surface and therefore vegetation properties [41
], which has so far not been explored to its full extent.
The objective of this study therefore is to investigate the value and effectiveness of the thermal remote sensing data for improving land cover classification. The test region is the Attert catchment in Luxemburg/Belgium providing a landscape with a variety of land cover types, mainly including forest, agriculture land, pasture and residential area. Based on the land cover change maps from 1990 to 2006 provided by CORINE and the change maps from 1972 to 1990 and 2006 to 2011 from CAOS project (not present in this paper), the land cover changes in quite small extent in 5 to 8 years especially in the early 1990s. In this study, the variation among different land covers is ignored and the land cover types of Level 1 and Level 2 are assumed to be constant during the periods from 1984 to 1990 and 2006 to 2011. Two of the most often and successfully applied standard methods, the k-
], as well as the Random Forest method [44
], will be applied to Landsat 4/5 and Landsat 8 images. Three groups of the single-date Landsat 8 images with different visible and thermal bands combinations will be classified into three levels of land use land cover categories, in order to evaluate the effectiveness of the thermal band in single image classification. The combination of band 3 and band 4, principal components, 6 bands combination without the thermal band, the thermal band and a 7 bands combination including the thermal band from time series Landsat 4/5 images listed in two groups will be classified into two levels for comparison and performance analysis. Ten-fold cross validation will be applied for the accuracy assessment with the overall accuracy.
The effectiveness of the thermal information/bands with regard to land cover classification using a single Landsat 8 image (including two thermal bands) and time series of Landsat 4/5 images (including one thermal band) was investigated for the Attert Catchment in Luxemburg. The single image was classified into three levels with 4, 7 and 14 LULC classes, respectively, and the time series of images were classified into the first two levels (Level 3 could not be analyzed due to the lack of ground truth data during the time frame of available images). The k-NN and the Random Forest algorithm were applied and assessed within a 10-fold cross-validation framework.
Firstly, the accuracy results from three variants of the single-date Landsat 8 image indicate that adding the thermal bands has clearly improved the accuracy of the Level 2 and Level 3 classification. The three variants achieved similar high OA of 98% ± 0.4% for the Level 1 classification. For the Level 2 and Level 3, Bands10T performed well with the best accuracy data, followed by the Bands6T and Bands4, which is 6% and 12% higher for the Level 3 classification. The OA from Bands6T including the two thermal bands are 3% and 6% higher respectively for the Level 2 and Level 3 category than the data of Bands4 without thermal bands. The results indicate that for the single Landsat 8 image classification, adding the thermal band to the VIR/NIR bands could improve the accuracy by 3% to 6% for Level 2 and Level 3 classification. As thermal bands are routinely available from different sensor platforms, their incorporation as input into the classification should also be done on a routine base, thereby significantly increasing classification accuracy.
Secondly, the results from time series of thermal images also demonstrate that the inclusion of thermal band significantly improves the LULC classification, compared to using standard VIS/NIR bands. The classification based on time series of thermal images provided comparably high OA when compared to the B3B4, 3PC, 6Bands and 7Bands images. TS1 thermal images obtained the best OA of 99.1% for Level 1 classification and 96.3% for the Level 2 classification. The time series of TS2 thermal images achieved the OA of 98.2% for the Level 1 and 93.9% for the Level 2. It is interesting to observe that the time series of thermal images could provide OA that are as good or even better than using the visible and near-infrared bands in the land cover classification, especially when the combination number of images used is higher than five.
Time series of TS2 thermal images also achieved comparatively high accuracy at the image number of 6, although the value is not higher than other images. Based on our results, a time-series of at least 5 or 6 thermal images is recommended as being almost optimal for situations that are similar to our study area. If the images from different years are obtained in the area with varying land cover and land use, the classification catalog and the selection of training set should be paid more attention with more land cover catalogs or taking the land cover change as the new class to ensure the consistence of the images in different years. In this study, the cloud free Landsat images were received mainly in the spring and summer time. They demonstrate the temperature discrepancies between various types of land cover especially for the agriculture areas, which is very effective for the Level 2 classification with the high OA from 93.9% to 96.3%. For the classification with the time series of images in the same year, at least two images from spring or summer time are recommended as the complementary sources.
Our study is not aimed at replacing the existing profound classification methods, but trying to add the thermal bands to improve the land cover classification based on the single image or the time series of images. The incorporation of thermal information improved the land cover classification indicated by better OA and Kappa statistics. However, in addition, thermal information alone provides similar or even better results when compared to the other time series of visible and near-infrared bands combination and/or principal components. Therefore, in case of failures or non-availability of VIS/NIR band data (as has been the case e.g. for the ASTER NIR bands), the thermal information could serve as a good substitute input in land cover classification experiments.
So far, the study area is limited to the Attert catchment in Luxembourg and the detailed land cover catalog is only classified at Level 2 (seven classes) when using time series of images, due to the lack of images in the same year for the agricultural area and therefore missing ground truth information. Because of the complicated atmosphere conditions, the preprocessing of the time series images probably could further benefit from other novel correction procedures, such as relative radiometric normalization [50
]. Further investigation of the time series of thermal remote sensing will be extended to the more specific classification for higher level with more specific land covers (such as Level 3 CORINE classes and/or application in agricultural and hydrological land cover types). The thermal bands in Landsat satellites have the limitation of a coarser spatial resolution when compared to the VIS/NIR bands, but the developed data fusion methods (such as e.g. the wavelet fusion method [64
] or the Spatial and Temporal Adaptive Reflectance Fusion Model (STARFM) [65
]). Other thermal sensors with wider spectral range such as ASTER or the sensor installed on the drone with finer spatial resolution and hyperspectral data [67
] should also be explored to aggregate the information for the regional land cover classification, but this is subject of ongoing and future research.