A Study on the Difference of LULC Classification Results Based on Landsat 8 and Landsat 9 Data

Abstract

Landsat 9 enhances the radiation resolution of the operational land imager from the 12 bits of Landsat 8 to 14 bits. The higher radiation resolution improves the sensitivity of the sensor to detect many subtler differences, especially in the case of dense forests or water. However, it remains unclear whether the difference in radiation resolution between Landsat 8 and Landsat 9 actually affects the classification results of water and tree species. Accordingly, the spectral reflectance and vegetation indices were extracted in this study, based on Landsat 8 and Landsat 9 images. Then, the classification models of land use and land cover (LULC) and tree species were developed by using a gradient tree boosting algorithm. Subsequently, the results were analyzed to further investigate how the differences in radiation resolution affect the classification results of LULC and tree species. It is shown that the LULC classification results of Landsat 8 and Landsat 9 are relatively favorable in most cases. However, the LULC classification results are relatively poor in test areas with a lower classification accuracy of water. Further analysis, in the case of test areas with poor classification results, indicates that there are significant differences in the water classification results between the two datasets. In other words, Landsat 9 produces better water classification results than Landsat 8 in most test areas. However, a temperature close to zero may lead to inverse water classification results. In addition, it indicates that the difference in forest classification results between the two datasets is small, but the results of forest tree species classification based on Landsat 9 are superior to those based on Landsat 8, with an improvement in overall accuracy of 6.01%. The results demonstrate that the difference in radiation resolution between Landsat 8 and Landsat 9 has little impact on the results of LULC classification in most cases. Nevertheless, in the case of some test areas, Landsat 9 is better suited for enhancing the classification accuracy of water and tree species.

1. Introduction

Regional sustainable development is a complex system that encompasses the sustainability of regional resources, the environment, society, the economy, and other factors [1]. Regional sustainable development relies heavily on the carrying capacity of resources and the natural environment. Land resources, as one of the most important natural resources, are a significant medium for the interaction of humans, the biosphere, the atmosphere, and other systems. In addition, land resources can provide material, energy, ecological, and other services for human survival and well-being [2]. The change in LULC (land use and land cover) is the most obvious manifestation of the human impact on the environment [3]. Due to its close relationship with regional natural factors and socio-economic factors, LULC serves as the basis of regional sustainable development. At the same time, regional sustainable development could, in turn, enhance the LULC change [4,5].

With the development of population and society, the extent and magnitude of land use are constantly increasing, resulting in a series of environmental issues, such as greenhouse gases emissions, biodiversity loss, soil erosion, desertification, and so on [6,7]. These environmental issues severely affect regional sustainable development. To better realize the regional sustainable development, the interaction mechanism among LULC, as well as the driving forces and environmental changes, are analyzed. The effects of LULC changes on the environment are evaluated to determine the primary driving factors affecting LULC and environmental changes, and the targeted prevention and improvement measures are implemented to achieve regional sustainable development. Therefore, how to realize the precise classification of LULC and long-term dynamic change monitoring is the central issue to be solved, in order to realize regional sustainable development.

As an essential source of data, satellite remote sensing, has been widely utilized, in order to produce LULC [8,9,10]. Landsat images have been commonly utilized, in order to produce high-quality multispectral images with medium spatial resolution and time resolution for scientific research since 1972. Since Landsat images are available for long-term series image storage and free access, they still continue to be one of the most commonly used essential sources of data, even today. Landsat 8, the fourth generation Landsat sensor is widely used in a variety of LULC classification studies. For example, Huan et al. [11] conducted a LULC classification study based on Landsat 8 images via the object-oriented method. The object-oriented support vector machine, based on texture, was found to be the most accurate, with an accuracy of 85.67%. Jiang et al. [12] also conducted LULC classification research based on Landsat 8 images by using the mixed decomposition of multi-seasonal spectrum and a decision tree classification method, and the accuracy was 90.94%. Further, Ren et al. [13] conducted LULC classification research based on Landsat 8 images. The results indicate that the highest obtainable classification accuracy, based on the combination of pixel and object-oriented methods, is 92.99%. Sharma et al. [14] conducted LULC classification research based on Landsat 8 images by employing the wavelet transform. According to the results, the highest obtainable accuracy of the five-level wavelet transform is 95.03%. Gupta and Shukla [15] studied the influence of terrain correction on the LULC classification based on Landsat 8 images, and the classification results before and after the terrain correction process were 65.60% and 82.40%, respectively. According to the aforementioned studies, it is evident that Landsat 8 images are suitable for LULC classification studies.

Landsat 9, the sister satellite of Landsat 8, was successfully launched on 27 September 2021 [16]. The two satellites share many similarities, such as being equipped with an operational land imager (OLI), a thermal infrared sensor (TIRS), 11 spectral bands, a maximum resolution of 15 m, a revisit period of 16 days, etc. It is possible to cover the LULC of the Earth repeatedly over an 8-day period by combining Landsat 8 and Landsat 9 [17]. Although the two satellites have many similar characteristics, Landsat 9 is capable of enhancing the radiation resolution of OLI from the 12 bits obtained through Landsat 8 to 14 bits. The resulting higher radiation resolution improves the sensitivity of the sensor, regarding brightness and color. Consequently, the sensor is capable of detecting many more subtle differences, especially in the case of darker areas with dense forests or water [16]. However, there have been only a limited number of studies conducted on the LULC classification based on Landsat 9 data, and even fewer studies have examined the impact of the differences in the radiation resolution between Landsat 8 and Landsat 9 on the LULC classification results. Accordingly, it is necessary to study whether the differences in radiation resolution between the two datasets have any significant impact on the LULC classification results, in order to facilitate the future collaborative application of the two datasets.

Taking the aforementioned considerations into account, this study examines China as the study area. The LULC and tree species classifications are conducted by utilizing a gradient tree boosting algorithm, based on the spectral reflectance and vegetation indices extracted from Landsat 8 and Landsat 9 images. Subsequently, the classification results are compared and analyzed to investigate the classification differences between water and tree species in the two datasets. The results are expected to provide adequate technical support for future long-term dynamic change monitoring studies on large-area LULCs, as well as accurate basic data for regional sustainable development.

2. Materials and Methods

2.1. Study Area

In order to better understand the difference between Landsat 8 and Landsat 9 images, in terms of LULC classification capability, this study investigates 40 test areas located throughout China in the time span between November 2021 and June 2022. The specific distribution of test areas is presented in Figure 1.

2.2. Data Introduction

2.2.1. Landsat 8 Data

Landsat is the main system for medium-resolution remote sensing in the earth observation system of the United States. Four generations of Landsat have been developed since 1972, among which, Landsat 8 is the most commonly used. It has a revisit period of 16 days, with a 15 m panchromatic resolution and a 30 m multi-spectral resolution [18]. The satellite was successfully launched on 11 February 2013. It was equipped with an OLI to measure the visible, near-infrared, and short-wave infrared parts of the spectrum, as well as a TIRS, which employs a new type of technology to measure the surface temperature in two thermal bands [19]. The Landsat 8 data can be downloaded for free through the Earth Resources Observation and Science (EROS) Center. The wavelength range and radiation resolution of Landsat 8 and Landsat 9 are depicted in

Table 1. The wavelength range and radiation resolution of Landsat 8 and Landsat 9.

Sensor	Band	Wavelength Range/μm	Radiation Resolution/Bit
OLI/ OLI-2	Band 1	0.435–0.451/0.435–0.450	12/14
	Band 2	0.452–0.512/0.452–0.512	12/14
	Band 3	0.533–0.590/0.532–0.589	12/14
	Band 4	0.636–0.673/0.636–0.672	12/14
	Band 5	0.851–0.879/0.850–0.879	12/14
	Band 6	1.566–1.651/1.565–1.651	12/14
	Band 7	2.107–2.294/2.105–2.294	12/14
	Band 8	0.504–0.676/0.503–0.675	12/14
	Band 9	1.363–1.384/1.363–1.384	12/14
TIRS/ TIRS-2	Band 10	10.60–11.18/10.45/11.20	12/14
	Band 11	11.51–12.50/11.58–12.50	12/14

.

2.2.2. Landsat 9 Data

Landsat 9 was successfully launched on 27 September 2021, equipped with OLI-2 and TIRS-2 [16]. In comparison to Landsat 8, Landsat 9 enhances the radiation resolution of OLI-2 (from the 12 bits obtained through Landsat 8 to 14 bits), increases the sensitivity of the sensor to brightness and color, and enables the sensor to detect many more subtle differences, especially in the case of darker areas with dense forests or water [16]. In addition to the improvement of OLI-2, TIRS-2 also significantly reduces stray light, resulting in an improved atmospheric correction process, as well as a more accurate measurement of the surface temperature. The Landsat 9 data can also be downloaded for free through the EROS Center.

As a result of the time difference between Landsat 8 and Landsat 9 images, as well as the impact of clouds and rain on image quality, the median value extracted from all available Landsat 8 and Landsat 9 images within a one-month period was utilized in the subsequent parameter extraction and LULC classification, respectively.

2.2.3. Sample Data

Sample Data Used for LULC Classification

Forty test areas were selected in different regions of China between November 2021 and June 2022; the corresponding months, when the images were obtained from various test areas, are provided in

Table 2. Schedule of corresponding months of the images obtained from the 40 test areas.

Test Areas	Month Image
1, 2, 3, 4	November 2021
5, 6, 7, 8, 9, 10, 11	December 2021
12, 13	January 2022
14, 15, 16	February 2022
17, 18, 19, 20, 21, 22, 23, 24	March 2022
25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36	April 2022
37	May 2022
38, 39, 40	June 2022

, and the spatial distribution of the 40 test areas is depicted in Figure 1.

Similarly, a European Space Agency (ESA) 10 m 2020 land-use dataset and an Environmental Systems Research Institute (ESRI) 10 m 2020 land-use dataset were employed as the basic data, in order to maintain the quality and quantity of the samples utilized in LULC classification. First, the two datasets were resampled to 30 m, and then the intersection of the two datasets with the same land cover type was determined. The specific area beyond the boundary line of the intersection area, buffered by 30 m inward, was utilized as the primary data for sampling. Furthermore, stratified random sampling was carried out using each individual type of land cover as the layer, with 150 samples per layer. The ESA 10 m 2020 land-use dataset is a global land cover product with a resolution of 10 m in 2020, generated by the European Space Agency, based on Sentinel-1 and Sentinel-2 datasets [20]. The ESRI 10 m 2020 land-use dataset is a global land cover product with a resolution of 10 m in 2020, constructed based on Sentinel-2 image data [20], as presented in Figure 2.

Sample Data Used for Forest Tree Species Classification

To understand the difference in forest tree species classification between Landsat 8 and Landsat 9 images, a total of 1481 sample points were selected by field survey and high-resolution images in Liuzhou, Guangxi Zhuang Autonomous Region, China. The specific description of the study area and sampling data can be found in [21]. The spatial distribution and detailed description of sample data are depicted in Figure 3 and

Table 3. Category and quantity of sample data.

Category of Sample Data	Quantity of Sample Data
Eucalyptus	148
Bamboo trees	164
Pine trees	122
Cedar	407
Orange trees	51
Tea bushes	47
Shrub	107
Mixed broad-leaved forest	85
Water	80
Crops	139
Built area	74
Grass	57

, respectively.

2.3. Method

In this study, the Landsat 8 and Landsat 9 images were processed to extract different parameters. On the basis of the parameters extracted from the Landsat 8 and Landsat 9 images, models of LULC and forest tree species were developed using gradient tree boosting method. On the basis of sampling and data, the accuracy of models was evaluated with a 10-fold cross-validation. The classification results based on Landsat 8 and Landsat 9 images were then compared and analyzed to investigate the differences between the two datasets in the classification of LULC and forest tree species. The flowchart used in this study is depicted in Figure 4.

2.3.1. Image Data Processing

The Landsat 8 and Landsat 9 images utilized in the study are Level-2 productions from the google earth engine (GEE) platform, which have been processed by geometric correction, radiometric correction, and atmospheric correction. During data processing, the Level-2 band reflectivity data will be scaled and offset by a scaling of 0.0000275 and an offset of 0.2. Therefore, the band reflectivity data must be conversed by multiplying 0.0000275 and subtracting 0.2 before the data are used to classify LULC and forest tree species. The GEE platform was then used to conduct a series of processes, such as cloud removal, clipping, splicing, and median calculation.

2.3.2. Parameter Extraction

On the basis of Landsat 8 and Landsat 9 images data processing, two kinds of parameters were extracted. The extracted specific parameters are shown in

Table 4. The extracted specific parameters.

Kinds	Number	Parameters
Spectral reflectance	7	B1, B2, B3, B4, B5, B6, B7
Spectral indices	9	Normalized difference vegetation index (NDVI), transformed chlorophyll absorption in reflectance index (TCARI), normalized difference water index (NDWI), modified chlorophyll absorption in reflectance index (MCARI), ratio difference vegetation index (RDVI), triangular vegetation index (TVI), soil adjusted vegetation index (SAVI), moisture stress index (MSI), land surface water index (LSWI)

, and the calculation formulas for the spectral indices are shown in

Table 5. Calculation formula of spectral indices.

Spectral Indices	Formula	Reference
NDVI	(B5 − B4)/(B5 + B4)	Broge and Mortensen [22]
TCARI	3 × ((B5 − B4) − 0.2 × (B5 − B3)) × (B5/B4)	Haboudane et al. [23]
NDWI	(B3 − B5)/(B5 + B3)	McFeeters [24]
MCARI	(B5 − B4) − 0.2 × (B5 − B3)) × (B5/B4)	Daughtry et al. [25]
RDVI	(B5 − B4)/pow(B5 − B4, 0.5)	Huete et al. [26]
TVI	0.5 × (120 × (B5 − B3)/200 × (B4 − B3))	Broge and Leblanc [27]
SAVI	(1 + 0.2) × float(B5 − B4)/(B5 + B4 + 0.2)	Bolyn et al. [28]
MSI	B5/B3	Bolyn et al. [28]
LSWI	(B5 − B6)/(B5 + B6)	Bridhikitti and Overcamp [29]

.

2.3.3. Classification Algorithm

Gradient tree boosting (GTB) is an ensemble learning method that is commonly used in data mining [30]. Its base model is a tree model, but it differs significantly from existing tree ensemble methods. In particular, the GTB algorithm is capable of reducing the variance of the overall model by a random sampling of features. In addition, it can also flexibly process various types of data, including continuous values and discrete values. Compared to other machine learning methods, the GTB algorithm is regarded as the most robust and does not require too much time for tuning parameters, and the results of the model are also satisfactory [31,32,33]. Accordingly, the GTB algorithm is commonly utilized, in order to develop models for LULC classification and tree species classification.

2.3.4. Accuracy Evaluation

In this study, a 10-fold cross-validation was implemented, in order to avoid the influence of accidental factors on the classification results. In addition, a confusion matrix was implemented, in order to evaluate the accuracy of classification results. The evaluation indicators included user’s accuracy (UA), producer’s accuracy (PA), overall accuracy (OA), and kappa coefficient. The specific calculation formulas are as follows:

$$U{A}_i=\frac{p_{ii}}{p_{i+}}$$

(1)

$$P{A}_i=\frac{p_{ii}}{p_{i+}}$$

(2)

$$OA=\frac{{\displaystyle {\sum }_{i=1}^k{p}_{ii}}}{p}$$

(3)

$$Kappa=\frac{p{\displaystyle \sum {}_{i=1}^k{p}_{ii}-{\displaystyle \sum {}_{i=1}^k{p}_{i+}{p}_{+i}}}}{p^2-{\displaystyle \sum {}_{i=1}^k{p}_{i+}{p}_{+i}}}$$

(4)

where p is the total number of samples, k is the total number of categories, $p_{ii}$ is the number of samples correctly classified, $p_{+i}$ is the number of samples of class i, and $p_{i+}$ is the number of samples predicted to be in class i.

3. Results

3.1. LULC Classification Results Based on Landsat 8 Images

Figure 5 presents the LULC classification results obtained from the 40 test areas, based on Landsat 8 images. According to the results presented in Figure 5, it is evident that the classification results were relatively favorable in the majority of the test areas, with an overall accuracy higher than 80%. The highest accuracy in these areas was 96.25%, and the kappa was 0.94. However, several test areas show relatively poor classification results, with an overall approximate accuracy of 70%. The lowest of these areas was 63.78%, and the kappa was 0.55. These include the following test areas: 1, 3, 8, 10, 11, 12, 13, 15, 27, 32, 33, 37, 38, and 39.

In order to determine whether the difference in radiation resolution between Landsat 8 and Landsat 9 data will lead to different classification results of water and forest. A detailed analysis of the classification results for water and forest from the 40 test areas is provided in Figure 6 and Figure 7, respectively.

Upon analyzing the classification results of water presented in Figure 6, it is evident that the classification results of water were similar to the LULC classification results presented in Figure 5. In other words, the classification results of water in the majority of the test areas were reasonably good, with the UA and PA higher than 80%. However, the classification results of water were poor in some test areas. The test areas with an accuracy of less than 40% include: 1, 3, 8, 10, 11, 12, 13, 15, 27, 32, 33, 37, 38, and 39. The UA and PA of test areas 3, 8, 10, 11, 12, 13, 15, 27, 37, 38, and 39 were all 0.00%.

Further, upon the analysis of the classification results of forest, presented in Figure 7, it is evident that the PA of forest classification results in all test areas was higher than 80.00%. In contrast, the UA of test areas 8, 10, 11, and 15 was relatively low, with respective values of 71.00%, 45.10%, 48.62%, and 64.50%. The UA of the forest classification results was still far higher than the UA of water classification results.

The correlation between the LULC classification results and the classification results of water and forest was analyzed based on the results presented in Figure 5, Figure 6 and Figure 7. The results of the analysis are provided in Figure 8.

The results presented in Figure 8 demonstrate that the correlation between the LULC classification results and water PA was relatively high, R² = 0.91, whereas the correlation between LULC classification results and forest PA was weak, R² = 0.04. Additionally, as presented in Figure 8a, when the OA of the LULC classification results was lower than 80.00%, the PA of the water classification result was generally less than 40%. In conjunction with the increase of the water classification result, the LULC classification result also gradually increased, until it reached a stable level. Consequently, the results of LULC classification based on Landsat 8 images were consistent with those of water classification.

3.2. LULC Classification Results Based on Landsat 9 Images

The LULC classification results of the 40 test areas based on Landsat 9 images are presented in Figure 9. According to Figure 9, the classification results for the majority of the test areas were relatively favorable, with an overall accuracy of more than 80%. The highest accuracy in these areas was 97.12%, and the kappa is 0.95. However, the classification results of some test areas were relatively poor, with an overall accuracy of about 70%, of which the lowest was 66.80%, and the kappa was 0.59. In particular, these include the following test areas: 1, 3, 5, 8, 10, 11, 15, 32, 33, 37, and 38. The classification results of the two images were roughly the same as those of the LULC classification results of Landsat 8 images. However, in test areas 12, 13, 27, and 39, the OA of the Landsat 9 image classification result was greater than 80.00%, whereas the OA of the Landsat 8 image classification result was less than 80.00%. In the case of test area 5, the OA of the Landsat 8 image classification result was greater than 80.00%, although the OA of the Landsat 9 image classification result was less than 80.00%.

The radiation resolution of Landsat 9 was better than that of Landsat 8. According to the water classification results presented in Figure 10, it is evident that the classification results of water in most test areas are relatively favorable, with a UA and PA higher than 80%. However, the classification results of water were poor in some test areas, and the test areas with an accuracy of less than 40% include: 1, 3, 4, 5, 7, 8, 10, 11, 15, 24, 32, 33, 38, and 39. This is roughly consistent with the test areas of the LULC classification results, with an OA of less than 80.00%, based on Landsat 9 images.

The forest classification results based on Landsat 9 images are presented in Figure 11. The results presented in Figure 11 demonstrate that the PA of the forest classification results in all test areas was higher than 80.00%, and the UA of forest classification results in the majority of the test areas was higher than 80%, whereas the UA of test areas 5, 10, and 11 were relatively low, 70.24%, 45.10%, and 53.38%, respectively. Consequently, the result of forest classification based on Landsat 9 images was similar to the classification results of the Landsat 8 images.

The correlation between the LULC classification results and the classification results of water and forest was analyzed based on the results presented in Figure 9, Figure 10 and Figure 11. The results are presented in Figure 12.

The results presented in Figure 12 indicate that the correlation between LULC classification results and water PA was relatively high (R² = 0.87), but there was no obvious correlation with forest PA (R² = 0.00). In a similar vein to the results provided in Figure 8, in cases where the OA of the LULC classification results was lower than 80.00%, the PA of water classification results was generally less than 40%, besides test area 37. In conjunction with the increasing water classification result, the LULC classification result also gradually increased, until it reached a stable level. Therefore, it is evident that the results of LULC classification based on Landsat 8 images are consistent with those of water classification results.

3.3. Comparative Study of Classification Results Based on Landsat 8 and Landsat 9 Images

The aforementioned analysis indicates that, despite the similarities between the classification results based on Landsat 8 and Landsat 9, they still differ significantly in certain test areas. In order to investigate the similarities and differences between the classification results of the two datasets, the correlation between the OA and kappa coefficients was analyzed based on the classification results of the two datasets. The results of the analysis are presented in Figure 13.

According to the results presented in Figure 13, the classification results of Landsat 8 and Landsat 9 were relatively poor in all 40 test areas. The R² of OA was only 0.43, and the R² of kappa coefficient was 0.45, which was mainly affected by test areas 5, 12, 13, 21, 27, 37, and 39. After excluding these seven test areas, the results are presented in Figure 13c,d. It is obvious that the correlation of the two data classification results was greatly improved, and the R² of OA and kappa coefficient increased to 0.92. Thus, the classification results of Landsat 8 and Landsat 9 were similar in most cases; however, they differred significantly in some regions.

The classification results of the above-mentioned test areas, i.e., 5, 12, 13, 21, 27, 37, and 39, are presented in Figure 14.

Based on the results presented in Figure 14, in the case of the seven test areas, the water classification results of Landsat 8 and Landsat 9 produced the most significant difference. For example, in the six test areas of 12, 13, 21, 27, 37, and 39, the water classification results of Landsat 9 were superior to those of Landsat 8. The water classification results of the Landsat 8 images in other test areas were completely inaccurate, except for test area 21. In test area 5, the water classification results of the Landsat 8 images were more favorable than those of Landsat 9, although the Landsat 9 water classification was inaccurate, as presented in Figure 14a,b.

According to the above analysis, it is evident that Landsat 8 and Landsat 9 have great differences, in terms of water classification results. Therefore, the correlation between the water classification results of the two datasets was analyzed, as presented in Figure 15.

Figure 15 demonstrates that the correlation of the water classification results of Landsat 8 and Landsat 9 was relatively poor, in which the correlation R² of UA was only 0.40, whereas the correlation R² of PA was 0.38. Upon analyzing the results presented in Figure 15, it is evident that the poor correlation between the two datasets of water classification results is largely attributable to the inaccurate classification of the water in some test areas. For example, the UA of the Landsat 8 water classification results in test areas 12, 13, 27, 37, and 39 was 0, whereas the UA of the Landsat 9 water classification results was 96.15%, 90.68%, 100.00%, 75.14%, and 90.77%, respectively. For test area 5, the UA of the Landsat 9 water classification results was 0. However, the UA of the Landsat 8 water classification results was 93.33%. The classification results of Landsat 8 and Landsat 9 include significant differences in the classification of water in some test areas, which may also be attributed to the difference in the radiation resolution between the two datasets.

The correlation between the forest classification results of the two data was analyzed, in order to demonstrate the differences between Landsat 8 and Landsat 9 in forest classification. The results of the analysis are presented in Figure 16.

Figure 16 demonstrates that the correlation between the forest classification results of the two datasets was relatively favorable, the R² of UA was 0.58, and the R² of PA was 0.71. It is evident that the forest classification results of the LULC classification are not capable of highlighting the differences between Landsat 8 and Landsat 9 in forest classification.

In order to further investigate the differences in forest classification between Landsat 8 and Landsat 9, the two datasets were employed to classify the forest tree species. The results are provided in

Table 6. Results of forest tree species classification based on Landsat 8 and Landsat 9.

Data	OA/%	Kappa Coefficient
Landsat 8	56.31	0.49
Landsat 9	62.32	0.56

and Figure 17.

Upon analyzing the results provided in Table 6, it is evident that the classification results of the forest tree species based on Landsat 9 are superior to those of Landsat 8, with an increase in the overall accuracy of 6.01%, as well as an increase in the kappa coefficient by 0.07. The results presented in Figure 17 indicate that the UA of all tree species based on Landsat 9 was higher than that of Landsat 8. As for PA, the results of Landsat 8 of crops, eucalyptus, pine trees, and mixed broad-leaved forest were superior to that of Landsat 9. In contrast, the results of Landsat 9 of built area, shrub, orange trees, cedar, water, bamboo trees, tea bushes, and grass were superior to those of Landsat 8. The result of water classification was significantly different. In particular, the PA of water classification results based on Landsat 8 was 0, and the PA of water classification results based on Landsat 9 was 92.50%, which is consistent with the water classification results of some of the aforementioned test areas. The classification results of forest tree species, based on Landsat 8 and Landsat 9, are presented in Figure 18.

According to Figure 18, the classification results of forest tree species in the Landsat 8 and Landsat 9 datasets are consistent in most areas. However, there are great differences in some areas. For example, in Figure 18, the classification result of the Landsat 8 image in the identified area was shrub, while the classification results of the Landsat 9 image were shrub and bamboo trees.

4. Discussion

4.1. Analysis on the Difference of Water Classification Results

Based on the aforementioned analysis, the correlation between the water classification results of Landsat 8 and Landsat 9 was relatively poor, with the correlation R² of UA being only 0.40, and the correlation R² of PA being 0.38. In order to investigate the reasons further, the correlation of the spectral reflectance of water was analyzed in bands 2, 3, 4, 5, 6, and 7 of Landsat 8 and Landsat 9 images in the seven test areas: 5, 12, 13, 21, 27, 37, and 39. Based on the water results extracted from the better water classification results in the two images, the Landsat 8 and Landsat 9 images were masked. In addition, the reflectivity of the different bands in the masked area was extracted separately, in order to perform a correlation analysis. Since it is unnecessary to analyze the reflectance correlation of different bands for all the seven test areas, only three of them were selected for the analysis. In particular, test areas 5, 21, and 27 were taken as examples, and the results are presented in Figure 17, Figure 18 and Figure 19, respectively. Test area 5 included inaccurate water classification results based on Landsat 9 data, test area 21 included inaccurate water classification results based on Landsat 8 data, and test area 27 included partially accurate classification results based on both Landsat 8 and Landsat 9 data.

The LULC classification results of Landsat 8 and Landsat 9 in test area 5 are presented in Figure 14a,b. It indicates exactly what part of the water in the Landsat 8 image was classified accurately and what part was classified incorrectly. In contrast, the entire water part in the Landsat 9 image was classified inaccurately. The reflectance correlation of the different bands between the two datasets is presented in Figure 19. Figure 19 demonstrates that the correlation between different bands of Landsat 8 and Landsat 9 images was relatively weak, with the exception of the blue band, and that the band reflectivity of Landsat 8 was greater than that of Landsat 9. Specifically, the correlation of the blue band was the highest, with an R² of 0.71. While the correlation of the other bands was relatively low, with an R² less than 0.50; the correlation of the SWIR2 band was the lowest, with an R² of 0.21. Since Landsat 8 has greater reflectivity than Landsat 9, it leads to a large difference between the two datasets in the water classification results. In other words, the classification results of water in Landsat 8 images were better than that of Landsat 9. Similarly, the water classification results were analyzed in conjunction with the temperature, in order to determine what exactly is causing the large differences.

It was discovered that the Landsat 8 images utilized in test area 5 were from 4 and 20 December 2021, and the lowest temperature reached −1 °C on 20 December 2021, whereas the Landsat 9 images utilized were from 12 and 28 December 2021, and the lowest temperature also reached −1 °C on 28 December. The lower temperature can freeze the water or slow down the flow rate, which may lead to an error in the water classification process. In addition, since Landsat 9 enhances the radiation resolution more than Landsat 8 and refines the difference between water surface and ice surface, to a certain extent, the classification result of the Landsat 9 image was weaker than that of the Landsat 8 image in cases where the temperature was low.

The LULC classification results of Landsat 8 and Landsat 9 in test area 21 are presented in Figure 14i,j. According to Figure 14, one part of the water in the Landsat 8 image was correctly classified, and one part was classified inaccurately. However, most of the water classification results were accurate in the Landsat 9 image. The reflectance correlation of different bands between the two datasets is also presented in Figure 20. The results presented in Figure 20 indicate that, with the exception of the blue band, the correlation between the different bands of the Landsat 8 and Landsat 9 images was relatively favorable. Specifically, the correlation of the blue band was lowest, with an R² equal to 0.14. While the correlation of other bands was higher, particularly in the case of the SWIR and SWIR2 bands, R² was greater than 0.80. The high correlation of band reflectance led to the correct classification of water, in the case of both datasets. However, the low correlation of some bands also led to the large difference in the classification of water between the two datasets. Upon further analysis, it was discovered that the Landsat 8 image utilized in test area 21 was from 2 March 2022, and the Landsat 9 image was from 10 March 2022, with a temperature greater than 9 °C.

Based on the rice planting schedule of the locals, the paddy fields are likely to be heavily irrigated in March, resulting in a similar spectral reflectance between the paddy field and the water. The radiation resolution of Landsat 8 was relatively lower than that of Landsat 9. In addition, the process of distinguishing the spectral reflectance differences between the paddy fields and the water was not as effective as Landsat 9. Consequently, some bodies of water were inaccurately classified as cropland. In contrast, Landsat 9 has a greater radiation resolution of 14 bits and is capable of better distinguishing the differences between paddy fields and water. As a result, most of the water bodies in test area 21 were correctly classified in the Landsat 9 image.

The LULC classification results of Landsat 8 and Landsat 9 in test area 27 are presented in Figure 14k,l. According to Figure 14, all the water in the Landsat 8 image was inaccurately classified. However, most of the water parts were classified accurately in the Landsat 9 image. The reflectance correlation of the different bands between the two datasets is presented in Figure 21. The results presented in Figure 21 demonstrate that all reflectance correlations of different bands between Landsat 8 and Landsat 9 were weak, and that the band reflectivity of Landsat 9 was higher than that of Landsat 8. In particular, the correlation of the SWIR band was the highest, with an R² of 0.41, and the correlation of the blue band was the lowest, with an R² of 0.09. As a result, there were great differences in the classification of water between the two datasets. In other words, the results of water classification in the Landsat 9 image were superior to those of Landsat 8. The Landsat 8 images utilized in test area 27 were from 3 April, 19 April, and 28 April 2022, with the lowest temperature being 5 °C on 28 April. However, the Landsat 9 images were from 4 April, 20 April, and 27 April 2022, with the lowest temperature being 3 °C on 27 April. Accordingly, freezing water had no effect on the difference between the water classification results of the two datasets. However, it did affect the difference in radiation resolution between the two datasets.

By analyzing the results of water classification, it was found that the water classification results of the Landsat 9 images were superior to those of Landsat 8 in the majority of the test areas. The primary reason for this result is that the Landsat 9 is capable of enhancing the radiation resolution of OLI-2 from 12 bits on Landsat 8 to 14 bits. The increased radiation resolution improves the sensitivity of the sensor to brightness and color; consequently, the sensor is capable of detecting many more subtle differences, particularly in the case of darker areas with water. The finding has been acknowledged by previous studies, as well. For example, Niroumand-Jadidi et al. [34] believed that the enhanced radiometric resolution of Landsat 9 (14-bit data) could provide greater sensitivity over water bodies and achieve better retrieval accuracy of water quality than the Sentinel-2 data (12-bit data). Niroumand-Jadidi et al. [35] retrieved the river depth using Landsat 9 and Sentinel-2 data and found that the retrieval accuracy of Landsat 9 was superior to that of Sentinel-2. By providing the enhanced radiometric resolution of Landsat 9, we could provide new opportunities for monitoring inland and coastal waters by providing higher sensitivity to water-leaving radiance.

4.2. Analysis on the Difference of Forest Tree Species Classification Results

According to the results presented in Figure 17, it is evident that the PA of eucalyptus, pine trees, and mixed broad-leaved forest classification in Landsat 8 images was better than that of Landsat 9. However, the PA of shrub, orange trees, cedar, bamboo trees, and tea bushes in Landsat 9 was better than that of Landsat 8. In order to investigate the cause behind the difference in forest tree species classification results between the Landsat 8 and Landsat 9 images, the reflectance correlation of different bands of eucalyptus and shrub were analyzed. The results are presented in Figure 22 and Figure 23.

Based on the results presented in Figure 22, it is evident that the correlation of different bands of Landsat 8 and Landsat 9 images is relatively favorable, with the SWIR2 band having the highest correlation, an R² of 0.57, and the blue band having the lowest correlation, with an R² of 0.27. This is consistent with the prior classification results of eucalyptus based on Landsat 8 and Landsat 9 images. Specifically, the results presented in Figure 17b demonstrate that the PA of eucalyptus classification based on the Landsat 8 image was 76.26%, while the PA of eucalyptus classification based on the Landsat 9 image was 56.08%, with a difference of 20.18%.

The reflectance correlation of the different bands, between Landsat 8 and Landsat 9, of shrubs is presented in Figure 23. According to the results presented in Figure 23, it is evident that the correlation of each band of Landsat 8 and Landsat 9 images was relatively low, with the SWIR band having the highest correlation, an R² of 0.48, and the blue band having the lowest correlation, an R² of 0.10. This is consistent with the previous classification results of shrub species based on Landsat 8 and Landsat 9 images. The results presented in Figure 17b demonstrate that the PA of shrub classification based on the Landsat 8 image was 27.66%, whereas the PA of shrub classification based on the Landsat 9 image was 64.49%, with a difference of 36.83%. As compared to the results of eucalyptus classification between the two datasets, the lower correlation of band reflectance leads to a greater difference in the shrub classification results between the two datasets.

Overall, the large band reflectance deviation between Landsat 8 and Landsat 9 images leads to significant differences in the classification of water and forest tree species between the two datasets. Therefore, if the two datasets are to be employed together for collaborative purposes in the future, it will be critical to effectively eliminate or weaken the reflectance deviation between the two datasets.

5. Conclusions

In this paper, the spectral reflectance and vegetation indices extracted from Landsat 8 and Landsat 9 datasets were utilized to develop the LULC and tree species classification models by employing the GTB algorithm. On this basis, the effects of differences in radiation resolutions between Landsat 8 and Landsat 9 images on LULC and tree species classification were investigated. The main conclusions are as follows:

(1) In most test areas, the LULC classification results of Landsat 8 and Landsat 9 images were relatively favorable, with an overall accuracy of more than 80%. However, in test areas with low classification accuracy of water, the LULC classification results were relatively poor, with an approximate accuracy of 70%.

(2) There were considerable differences in the water classification results between the Landsat 8 and Landsat 9 images, where the correlation R² of UA was only 0.40, and the correlation R² of PA was 0.38. In most test areas, the water classification results of the Landsat 9 images were better than those of Landsat 8. In cases where the temperature was close to zero and the water was mixed with ice, the results of the water classification became inverse. In other words, the results of water classification based on Landsat 8 images, in these cases, were better than those based on Landsat 9 images.

(3) The classification results of the forest in the LULC classification of Landsat 8 and Landsat 9 images were roughly the same. However, the results of tree species classification based on Landsat 9 images were better than those based on Landsat 8, which enhances the overall accuracy by 6.01%. In regard to the UA, the classification results of all tree species based on Landsat 9 were higher than those of Landsat 8. In regard to the PA, the classification results of eucalyptus, pine trees, and mixed broad-leaved forest based on Landsat 8 were better than those of Landsat 9, whereas the PA of shrub, orange trees, cedar, bamboo trees, and tea bushes based on Landsat 9 were better than those of Landsat 8.

This study demonstrates that the difference in radiation resolutions between Landsat 8 and Landsat 9 images had no significant impact on the results of LULC classification, although, in some test areas, Landsat 9 was capable of improving the accuracy of the classification results of water and forest tree species. However, only 40 test areas and only images, taken between November 2021 and June 2022, were used in the study. In the future, with the continuous increase in available monthly image data, more test areas and longer periods of time must be included, in order to further investigate the differences between the two datasets and to provide technical support for their collaborative application.