Recently I published a new article, which uses the method of sample point migration to quickly realize global long-term rapid land classification. This article released an application APP, and users can experience the effect of using rapid classification online. Original link:Land | Free Full-Text | Rapid Land Cover Classification Using a 36-Year Time Series of Multi-Source Remote Sensing Data
Research objectives
In this study, we implemented a radio frequency classifier in GEE to perform time series land classification at different spatial scales during the 2022 vegetation growing season using Landsat-8 and Sentinel-2 datasets. Our primary goal is to use different land classification models constructed from multi-source remote sensing variables to establish an efficient, accurate and universal land classification model for time series data sets, and to determine the land based on the difference in image values of sample points where land classification changes have not occurred. Classification sample points and migration thresholds. Our objectives were to (1) determine the threshold for sample point migration based on no change in land classification; (2) analyze the accuracy of a land classification model based on the threshold based on a 36-year time series of Landsat remote sensing images and high-precision Sentinel images; ( 3) Determine the optimal radio frequency land classification model based on different combinations of multi-source remote sensing variables and compare the impact of image resolution on classification accuracy.
Summary
Long-term series of land cover classification information are the basis for scientific research on urban expansion, vegetation change and carbon cycle. The rapid development of cloud computing platforms, such as Google Earth Engine (GEE), and access to multi-source satellite imagery from Landsat and Sentinel-2 have enabled the application of machine learning algorithms in image classification. Here, we use the random forest algorithm to quickly implement time series land cover classification at different scales based on selecting fixed land classification sample points from images obtained in 2022, and the annual spectral differences of the sample points. The classification accuracy is improved by utilizing multi-source remote sensing data such as synthetic aperture radar (SAR) and digital elevation model (DEM). The results show: (i) By calculating the sampling points in each band of the Landsat time series from 1986 to 2022, the maximum difference (threshold) of the sampling points without land level changes is determined to be 0.25; (ii) The same sensor in Landsat 8 The kappa coefficient and observation accuracy are both higher than the results of TM and ETM+ sensor data
Article flow chart
data set
| Name |
Earth Engine Snippet |
Date |
Resolution |
| Landsat 5 |
LANDSAT/LT05/C02/T1_L2 |
“1984-03-16”- “2012-05-05” |
30 m |
| Landsat 7 |
LANDSAT/LE07/C02/T1_L2 |
"1999-05-28" - |
30 m |
| Landsat 8 |
LANDSAT/LC08/C02/T1_L2 |
"2013-03-18"- |
30 m |
| Sentinel 1 |
Copernicus/S1_GRD |
"2014-10-03"- |
10 m |
| Sentinel 2 |
COPERNICUS/S2 |
"2015-06-23" - |
10 m |
| DEM |
NASA/NASADEM_HGT/001 |
"2000-02-11" |
30 m |
Data preprocessing process
Preprocessing of optical remote sensing images includes image stitching, cloud removal, mosaic and cropping. Among them, image cloud removal methods all remove clouds and cloud shadow elements by calling the QA quality band of Landsat and Sentinel-2 data and operating the mask bit by bit. The mosaic processing of the images was fused using the median method, and then the Landsat series of images from 1986 to 2022 and the Sentinel-2 vegetation growing season remote sensing images from 2019 to 2022 were obtained.
The Sentinel-1 polarization data GEE has officially gone through the processes of Ground Distance Detection (GRD) boundary noise removal, thermal noise removal, radiometric calibration and radiometric correction. The VV and VH polarization bands in the interferometry wide scan (IW) mode suitable for land surface remote sensing studies were selected for this study. The DEM data were reprojected and resampled to extract variables such as elevation, slope, and elevation as terrain factors to participate in the construction of land classification models.
Multi-source remote sensing variable combination
| Multi-source remote sensing image |
Variable factors |
| Spectral Band |
Blue, Green, Red, Nir, Swir1, Swir2 |
| Spectral Index |
NDVI, NDBI, NDWI, RVI, DVI |
| Terrain |
Elevation, Slope, Aspect |
| SAR |
HH, HV |
Sample point migration results
Landsat long time series analysis
Landsat remote sensing images from 1986 to 2022 were used for land cover classification. The sample point migration threshold was 0.25. The accuracy was evaluated using OA and kappa coefficients, and the number of migrated sample points was calculated. The results show that the classification accuracy of images is highest in years close to the fixed land classification of 2022, and the difference between the kappa coefficient and OA becomes larger as the number of years from the initial land classification sample point in 2022 increases. However, overall land classification accuracy remains high, with 1999 having the lowest kappa coefficient of 0.60 and OA the lowest at 0.75. The number of classified sample points decreased with the increase in 2022, and the migrated sample point data stabilized at 900, accounting for approximately 60% of the original number of sample points. Notably, differences between Landsat TM/ETM and OIL sensor technology may explain the lower accuracy of results between 1986, when the study began, and 2012.
Analysis of differences between Landsat and sentinel series images
In order to verify the universality of this article in different remote sensing images and the repeatability under complex terrain conditions, we selected the Huodong National Planned Mining Area in Shanxi Province with complex terrain conditions as the research area, and conducted the Sentinel-2 high altitude survey in 2019-2022. The mining category has been added to the land classification system of high-resolution remote sensing images. By counting each band of land types in different years, the land cover classification accuracy in different threshold ranges (0.1 - 0.4) was evaluated. The results show that when the threshold for training sample point migration is set in the range of 0.20 - 0.30, the land classification accuracy is higher, and the number of land classification sample points after threshold screening is maintained at about 70% of the original sample points, which can be achieved to a greater extent. The number of sample points required to satisfy the land classification. At the same time, the kappa coefficient from 2019 to 2021 is stable at around 0.90, and OA is also around 0.91.
| Threshold |
Method |
2019 |
2020 |
2021 |
|||
| Accuracy |
Number of Samples |
Accuracy |
Number of Samples |
Accuracy |
Number of Samples |
||
| 0.1 |
Kappa |
0.333 |
19 |
0.639 |
56 |
0.582 |
11 |
| OH |
0.500 |
0.923 |
0.684 |
||||
| 0.15 |
Kappa |
0.707 |
108 |
0.644 |
160 |
0.867 |
70 |
| OH |
0.818 |
0.792 |
0.896 |
||||
| 0.20 |
Kappa |
0.829 |
560 |
0.910 |
681 |
0.935 |
556 |
| OH |
0.874 |
0.949 |
0.941 |
||||
| 0.25 |
Kappa |
0.884 |
863 |
0.886 |
956 |
0.914 |
901 |
| OH |
0.907 |
0.908 |
0.931 |
||||
| 0.30 |
Kappa |
0.901 |
1028 |
0.914 |
1094 |
0.870 |
1055 |
| OH |
0.919 |
0.931 |
0.910 |
||||
| 0.35 |
Kappa |
0.882 |
1112 |
0.921 |
1157 |
0.889 |
1132 |
| OH |
0.903 |
0.904 |
0.876 |
||||
| 0.40 |
Kappa |
0.846 |
1173 |
0.891 |
1193 |
0.926 |
1176 |
| OH |
0.875 |
0.905 |
0.893 |
||||
Accuracy comparison of different multi-source remote sensing combinations of Landsat-8
The combination of multi-source remote sensing variables improves the model accuracy of land classification. The model accuracy increases with the increase of different variables. In particular, the combined model of spectral band + index + SAR + surface variables and terrain has the best effect. Taking 2019 as an example, the kappa coefficient of a single spectral band finally increased from 0.863 to 0.910 of spectral band + index + terrain + SAR, while the OA of the sample variable combination also increased from 0.888 to 0.927. In addition, compared with the participating land classification accuracy in 2022, sample points filtered by the threshold can eliminate misclassification of sample points during the selection process, so the land classification accuracy in 2019-2021 is better than the land classification accuracy in 2022.
| Variable combinations |
2019 |
2020 |
2021 |
2022 |
||||
| Kappa |
OH |
Kappa |
OH |
Kappa |
OH |
Kappa |
OH |
|
| Spectral band |
0.863 |
0.888 |
0.877 |
0.900 |
0.867 |
0.893 |
0.860 |
0.887 |
| Spectral Band + Index |
0.874 |
0.907 |
0.878 |
0.900 |
0.867 |
0.892 |
0.883 |
0.905 |
| Spectral band + SAR |
0.866 |
0.890 |
0.878 |
0.901 |
0.907 |
0.924 |
0.875 |
0.896 |
| Spectral band + Index + SAR |
0.903 |
0.915 |
0.913 |
0.929 |
0.896 |
0.916 |
0.900 |
0.915 |
| Spectral band + index + Terrain + SAR |
0.910 |
0.927 |
0.880 |
0.903 |
0.921 |
0.936 |
0.889 |
0.919 |
Accuracy comparison of different multi-source remote sensing combinations of Sentinel-2
The land classification accuracy of Landsat-8 under different variable combinations (Table 7; Figure 5) is lower than the multi-source remote sensing land classification accuracy based on Sentinel-2 images. Taking 2022 as an example, the land classification accuracy of the spectral band + index + synthetic aperture radar combination is the highest, and the model combination of spectral band + synthetic aperture radar is better than the spectral band + index. In 2019 and 2020, the full variable combination had the highest accuracy, while in 2021 and 2022 the best variable combination was spectral band + index + SAR.
| Variable combinations |
2019 |
2020 |
2021 |
2022 |
||||
| Kappa |
OH |
Kappa |
OH |
Kappa |
OH |
Kappa |
OH |
|
| Spectral band |
0.833 |
0.864 |
0.828 |
0.864 |
0.836 |
0.869 |
0.881 |
0.903 |
| Spectral band + Index |
0.837 |
0.868 |
0.835 |
0.866 |
0.851 |
0.879 |
0.828 |
0.861 |
| Spectral band + SAR |
0.848 |
0.877 |
0.870 |
0.896 |
0.846 |
0.876 |
0.882 |
0.903 |
| Spectral band + Index + SAR |
0.831 |
0.864 |
0.866 |
0.894 |
0.871 |
0.894 |
0.917 |
0.933 |
| Spectral band + Index + Terrain + SAR |
0.872 |
0.897 |
0.892 |
0.913 |
0.848 |
0.878 |
0.900 |
0.919 |
in conclusion
Rapid land cover classification was performed using the GEE remote sensing cloud platform using Landsat 5, 7, 8 and Sentinel-2 remote sensing images with a time series spanning 36 years. A single sampling point migration method was used to generate time series land cover classification maps at provincial-regional scale and mining operation scale. The final sample point migration threshold is 0.25, corresponding to no change in classification. The optimal combination of multi-source remote sensing variables for parameterization of RF machine learning algorithms is: spectral bands of Landsat 8 and Sentinel-2 generated data + index + terrain + synthetic aperture radar. The RF model produced the highest accuracy classification map for 2022 using Landsat 8 data, with an OA value of 0.90 and a Kappa coefficient of 0.919. Our analysis shows that higher accuracy can be achieved using images with higher spatial and temporal resolution. Further work on collation of low-resolution remote sensing images and machine learning techniques will enable the evaluation of long-term series of land cover classification maps at the global scale. As sensor technology develops, we expect that the accuracy of land cover classification will continue to improve, enabling the identification of land cover classes that have not yet been considered in the future.
Quick Land Classification APP
APP link
Land classification of Landsat imagery
Program Introduction This program mainly uses fixed sample points for land classification in one year, selects images of any year from 1984 to the present (landsat 5 / 7 / 8 / 9 SR), and performs land classification in different years according to the random forest algorithm.
1. Use the drawing tools (rectangle and polygon available) in the upper left corner of the map toolbar to select the specified study area. 2. Select the image time of the sampling point in the current year, load the image, and select points on the true color image. 3. According to the geometry import on the left side of the map tool, select the designated land classification sample point classification label, delete the origin through the tool, select the specified research area, and restart the sample points of different land classifications in the specified area. If there is no such land type in the study area, it can be ignored. 4. Select the time range of images to be classified. 5.Select the image threshold based on the Threshold checkbox. 6. Select different image sets based on time and load the land classification results within the specified time range.
Select sampling time
Here we first need to collect the land classification of sample points. First, select the image at a specific time, and then select the sample point of the required classification on the image. When selecting the sample point, select the sample point layer we need to specify. To proceed, that is, the different sample point layers in the upper left corner of the GEEAPP interface.
Select images and quick sort
Here we need to specify the year and image. Note that the time must be consistent with the Landsat series of images you selected.
The final result after classification
2023
2022
Migrated sample points
Article citation
MDPI and ACS Style
Yan, X.; Li, J.; Smith, A.R.; Yang, D.; Ma, T.; Su, Y. Rapid Land Cover Classification Using a 36-Year Time Series of Multi-Source Remote Sensing Data. Land 2023, 12, 2149. https://doi.org/10.3390/land12122149
AMA Style
Yan X, Li J, Smith AR, Yang D, Ma T, Su Y. Rapid Land Cover Classification Using a 36-Year Time Series of Multi-Source Remote Sensing Data. Land. 2023; 12(12):2149. https://doi.org/10.3390/land12122149
Chicago/Turabian Style
Yan, Xingguang, Jing Li, Andrew R. Smith, Di Yang, Tianyue Ma, and Yiting Su. 2023. "Rapid Land Cover Classification Using a 36-Year Time Series of Multi-Source Remote Sensing Data" Land 12, no. 12: 2149. https://doi.org/10.3390/land12122149
