Analysis of Land Cover Changes Based on Classification Result Using PlanetScope Satellite Imagery

Abstract

Compared to the imagery produced by traditional satellites, PlanetScope satellite imagery has made it possible to easily capture remotely-sensed imagery every day through dozens or even hundreds of satellites on a relatively small budget. This study aimed to detect changed areas and update a land cover map using a PlanetScope image. To generate a classification map, pixel-based Random Forest (RF) classification was performed by using additional features, such as the Normalized Difference Water Index (NDWI) and the Normalized Difference Vegetation Index (NDVI). The classification result was converted to vector data and compared with the existing land cover map to estimate the changed area. To estimate the accuracy and trends of the changed area, the quantitative quality of the supervised classification result using the PlanetScope image was evaluated first. In addition, the patterns of the changed area that corresponded to the classification result were analyzed using the PlanetScope satellite image. Experimental results found that the PlanetScope image can be used to effectively to detect changed areas on large-scale land cover maps, and supervised classification results can update the changed areas.

1. Introduction

A land cover map is a topographic map that is classified per certain scientific criteria (Ministry of Environment, 2017). Although land cover mapping is used in various fields, such as environmental monitoring, disaster analysis and urban planing, its effective utilization requires the latest information to reflect current land cover. The production of land cover maps requires current aerial photographs or satellite images of the corresponding area. However, it is difficult to update a land cover map using data that reflect the latest land cover due to various factors, such as an image’s acquisition time and adequate spatial resolution. Although Earth observation satellites, which are mainly large, were developed in the 2000s, startup companies became interested in developing microsatellites, and cube satellites (CubeSats) were developed in 2013. Because the development costs of CubeSats are lower than those of traditional satellites, the biggest advantage is the ability to produce and launch multiple satellites of the same standard at once. CubeSats also minimize the timeliness and periodicity problems of conventional satellite images.

Various methods have been studied to create or update land cover maps using satellite images. Lee et al. (2013) analyzed the processes of land cover mapping and suggested improvement plans to the Ministry of Environment. In addition, Lee et al. (2015) proposed a technique to automatically generate the supervised classification results by Landsat satellite images and to automatically update the land cover map via an existing map. Kim and Yeom (2012) studied the possibilities and limitations of the object-based image analysis method for producing land cover maps in rural areas using multi-temporal RapidEye satellite images. Walter (2004) automatically extracted reference datasets based on a land cover map and used it to perform object-based image classification. Marcal et al. (2005) conducted supervised classification using ASTER satellite imagery and conducted a study to utilize it as data for renewal of a land cover map. In addition, Rawat and Kumar (2015) used time-series Landsat data to generate and analyze time series classification images for India. The existing researches are aimed to develop a classification technique for generating or updating the land cover map. In addition, various parameters and empirical thresholds are needed to update the land cover map.

In this study, we propose an effective method for updating the land cover map of middle category using image classification techniques for satellite images that were acquired using PlanetScope imagery by Cubesats. For this purpose, we analyzed the classification results that were obtained from the Random Forest (RF) classification method, which is a representative machine learning classification technique. In addition, by analyzing the limitations of the image classification, we evaluated which class characteristics can and cannot update land cover.

2. Study Area and Data

PlanetScope, also known as DOVE, is the representative CubeSats for remote sensing that is produced by Planet Labs in the United States (Planet, 2018). Planet Labs made first experimental launch of CubeSats in 2013, and PlanetScope had more than 175 satellites in operation by 2017. The actual size of the satellite is 10 × 10 × 30 cm (see shape and specifications in Fig. 1 and Table 1, respectively).

Fig. 1.  PlanetScope (DOVE) satellite sensor.

Table 1.  PlanetScope satellite system characteristics

In this study, an area near Sejong City, where urban areas have undergone many changes in land cover over many years, was used for experiment. A total of three PlanetScope satellite images taken on September 24, 2016 were used, as shown in Fig. 2. The land cover map of middle category that was produced in 2013 was also used, as shown in Fig. 3. After image mosaic of three satellite images were performed, and regions with the same coordinates as the land cover map were extracted for the experiment. Considering the spatial resolution of the satellite image, the land cover map was reclassified into five classes, including urban area, forest, cropland, paddy fields, and water.

Fig. 2. Study site.

Fig. 3. Land cover map of the experimental site.

3. Methodology

1) Atmospheric Correction

To perform a supervised classification, the Digital Number of the PlanetScope imagery had to be converted to the surface reflectance value. The radiance data of the PlanetScope imagery were converted to the top of atmospheric (TOA) reflectance using Eq. (1):

\(R E F(i)=R A D(i) \frac{\pi \times \text { SunDist }^{2}}{\operatorname{EAI}(i) \times \cos (\text { SolarZenith })}\)       (1)

where, REF(i) is the TOA reflectance of the i^th band, RAD(i) means the radiance value of the i^th band, SunDist is the distance between the earth and the sun, EAI(i) is the exo-atmospheric irradiance of the i^th band, and SolarZenith is the solar zenith angle.

The TOA reflectance value is then converted to the surface reflectance value using the atmospheric correction algorithm. Atmospheric correction is a method of calibration that uses the atmospheric model equation, which includes the atmospheric conditions at the time of satellite image acquisition. It is based on the radiative transfer model. A Quick Atmospheric Correction (QUAC), Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes (FLAASH), and Atmospheric Correction (ATCOR) are typical algorithms. QUAC performs atmospheric correction using only the characteristics of pixels in the image without auxiliary information related to field data and image acquisition. It is also applicable to multi-spectral and hyper-spectral data (Bernstein et al., 2012). Therefore we used QUAC algorithm using ENVI commercial remote sensing software. Fig. 4 is the result of applying the QUAC method to the PlanetScope image. As shown in Fig. 4(c), vegetation pixel of images was converted to reflectance value, compared to TOA reflectance value of Fig. 4(b).

Fig. 4.  Example of the atmospheric correction using QUAC.

2) RF classifier

Artificial intelligence techniques, particularly machine learning algorithms, have recently utilized in satellite image processing (Li et al., 2016). There are decision tree methods for classification via machine learning, which consist of leaves that correspond to the class, nodes that correspond to the attributes of the data to be classified, and arcs that correspond to the optional values for the attributes (Belgiu and Draguţ, 2016). RF classification techniques are the preferred model for constructing a number of decision trees through training datasets and utilizing them to extract optimal classification results, with the highest probability of seeing the results from a number of decision trees. Each decision tree is created by the following rules (Belgiu and Draguţ, 2016):

1. Random reconstruction of N samples from all training data to generate individual training data for generation of each decision tree model.

2. X bands among M bands are randomly extracted, and optimal node variables for a decision tree are determined.

3. Each decision tree does not perform pruning but is generated up to the maximum possible size.

Because the RF classification method uses a large number of decision trees in the classification process, it is possible to generate classification results with high classification accuracy in comparison with the decision tree technique, and it is known to be effective in supervised classification of large capacity remote sensing data (Belgiu and Draguţ, 2016).

3) Generation of additional features

To improve the accuracy of the pixel-based classification based on RF, the Normalized Difference Vegetation Index (NDVI) and Normalized Difference Water Index (NDWI) values for classifying water regions were generated and used as additional data for classification. The NDVI is an index that can analyze changes and trends in vegetation growth and conditions. It is defined as Eq. (2) using the spectral reflectance of the image data that were obtained in the red and near-infrared bands (Rouse et al., 1974).

\(N D V I=\frac{N I R-R E D}{N I R+R E D}\)       (2)

NDWI is a remote sensing index related to moisture. When the focus is on the water content of the vegetation, the index is calculated using the spectral reflectance of the near-infrared and short-wave infrared bands. However, if the target is placed in a pure water system, the index can be calculated using the green and near-infrared bands (Eq. 3) (Xu, 2006). As with NDVI, NDWI has a range of -1 to 1, and the closer to 1 that it is, the higher the moisture content.

\(N D W I=\frac{G R E E N-N I R}{G R E E N+N I R}\)       (3)

Fig. 5 represents subsets of NDVI and NDWI images of the PlanetScope imagery.

Fig. 5.  Examples of NDVI and NDWI images.

4. Experimental Results and Discussion

1) Evaluation of pixel-based classification results by RF

For pixel-based classification by RF, six classes of forest, water, cropland, urban area, grassland, and paddy field were selected. Table 2 shows the number of pixels of the training and reference datasets that were selected. Based on the training data selected above, decision trees of one hundred were generated. The classification result is shown in Fig. 6. The results confirm that the overall classification of each class is good, although some error was found in the class with significant vegetation, such as forests and grasslands.

Fig. 6. Pixel-based classification result by RF.

Table 2. Training and reference data of the study area

To evaluate the accuracy of the classification results, the accuracy and the kappa coefficient were calculated using the error matrix for both the training and reference data. The error matrix is shown in Table 3. After performing the RF classification using the training data, the overall accuracy was 95.54%, and the kappa coefficient was 0.94. Specifically, the user’s and producer’s accuracies for cropland were 99% and 97%, respectively. Errors in the classification results occurred mostly in areas related to grassland, forest and cropland. However, from the perspective of overall accuracy, it is estimated that changes in the existing land cover map can be checked effectively.

Table 3.  Error matrix of the classification result

2) Analysis of the updated results in relation to the changed area

To analyze the results that were obtained through RF classification, which is a pixel-based technique, the classification result was transformed into the objects by calculating majority pixels of classification result in each object of land cover map, as shown in Fig. 7(b). And then, the forest and grassland classes were integrated into the vegetation class due to misclassification among vegetation regions, such as forest and grassland. To confirm the feasibility of updating existing land cover maps, it was assumed that, in each class, the spatial characteristics or boundary of final object of updated land cover maps maintained the characteristics of the existing land cover maps. All changed objects were then compared with the existing land cover map classes. We analyzed the classes that could update the land cover by comparing the existing land cover maps with the image classification results, as shown in Fig. 7(a-b).To evaluate the accuracy of the updated land cover map, the differences between the existing land cover map and the classification result image are shown in Fig. 7(c). Per the object-based land cover map, it was estimated that the property values of 1,527 objects among 7,162 objects of the existing land cover map had changed. Among 1,527 objects, 80 points from the region where the change of land cover was judged to occur were manually extracted as a random sampling, as shown in Fig. 7(d), in order to evaluate the accuracy. As a result, a quantitative evaluation was performed on the classification results that were generated by the RF classification method and the existing land cover map. The error matrix of the existing land cover map and reference data are shown in Table 4. The accuracy of the existing land cover map based on the reference data is 10%. This indicates that 10% of the objects that were estimated to have changed land cover are false positives; hence, 90% of the changed areas that were extracted in this study were correctly detected.

Fig. 7. Updating result of land cover map using the classification result.

Table 4. Error matrix of the existing land cover map based on the reference data

The error matrix of the reference data and the classification result that were generated by the RF classification technique are shown in Table 5. Overall accuracy was about 78.7%. User’s and producer’s accuracy of the vegetation class were 88.4% and 74.1%, respectively. Based on the reference data, it was confirmed that changes in both the urbanization area and vegetation were effectively detected. By contrast, some areas in the paddy field class were detected as various other classes; however, most were misclassified as an urban region because some paddy fields and complex urban area have high spectral reflectance values. In conclusion, it was confirmed that the land cover map can be effectively updated through the classification results of PlanetScope images.

Table 5. Error matrix of the updated land cover map using the classification result based on the reference data

Fig. 8 shows the results of comparing the land cover map with the detailed area of the classification result. As shown in Fig. 8(a), in the existing land cover maps, it can be confirmed that the area that was previously cropland and paddy field had changed into an urban area through development. It is also confirmed that many apartments were built in the area using PlanetScope imagery. The regions were detected as changed areas and confirmed as effectively updated. Fig. 8(b) shows that the cropland in the existing land cover mp changed to a paddy field due to urban development, and the regions were effectively extracted from the updated land cover map. In addition, some greenhouses in the cropland were also detected as urban areas. Conversely, as seen in Fig. 8(c), some area that was classified as a paddy field in the existing land cover map was updated to an urban area, correctly. However, because the existing land cover map’s objects do not reflect the spatial characteristics of the current changed area, the change of the objects’ shape should be considered when updating the land cover map.

Fig. 8. Detailed updating results of the land cover map using the classification result.

5. Conclusion

In this study, we tried to analyze the possibility of land cover map updating using PlanetScope satellite imagery. For this purpose, supervised classification was performed using a PlanetScope satellite image. Image classification results based on the RF showed 95.5%accuracy in six classes. Although some errors occurred in the cropland and vegetation areas, it was confirmed that the results could be used as basic data for updating existing land cover maps. The quantitative analysis of the areas that were judged to have changed based on the existing land cover map found that the accuracy of the updated land cover map represented 78.7% when updating five classes. In this study, the update was performed based on the existing land cover map’s objects. Therefore, when the land cover was mixed in some objects, it was not reflected. Nonetheless, this study confirmed that the proposed method can effectively detect the presence and/or absence of changed areas.