Introduction
Land serves as the foundation of urban development [1]. With escalating urbanization and pervasive human activities, the land use and cover patterns within the Yellow River Delta wetland region have undergone profound changes. Land use and cover changes (LUCC) wield significant influence over biodiversity and ecosystems [2-5]. In recent years, vast swaths of agricultural land in the Yellow River Delta wetland have been replaced by urban lands and industrial areas, disrupting the delicate balance of the ecosystem. Through the examination of land use changes in the Yellow River Delta wetland, valuable insights emerge, offering crucial references and lessons for future resource management, environmental conservation, and urban-rural development endeavors.
Studies on wetland mapping in the Yellow River Delta region have predominantly relied on remote sensing techniques thus far. For instance, research investigating land use types and their dynamic changes in the region encompasses a range of aspects, including land cover types [6], wetland classification [7], information on saline and alkaline lands [8], spatial and temporal land use change patterns and drivers [9], as well as analysis of cropland degradation and its underlying causes [10]. These studies utilize various remote sensing techniques and methodologies, such as supervised classification [6], decision tree classification, visual interpretation [7], object-oriented classification, support vector machines [11], and tasseled cap transformation [12]. However, these studies exhibit certain limitations: some are confined to single counties or cities [8, 9, 11] and lack comprehensive coverage of the entire region; others focus solely on specific features like saline soils [8] or arable land [10] without conducting a comprehensive classification. Even when a comprehensive classification is undertaken, the division of land use types may be coarse, the study duration short, and the number of image periods insufficient [6] for long-term dynamic monitoring. Moreover, machine learning methods have been underutilized in these studies. V.F. Rodriguez-Galiano et al. demonstrated that the random forest algorithm can enhance the accuracy of land cover classification and outperform a single decision tree [13]. Therefore, adopting the random forest algorithm for mapping in the Yellow River Delta wetland is deemed feasible.
Based on the above discussion, this paper combines feature extraction with the employment of the random forest algorithm to classify land use in the Yellow River Delta using five-phase Landsat-5/7/8 images from 2002 to 2021, evaluating the accuracy of the classification results. Finally, it conducts analyses on land use change and landscape pattern change with the aim of promoting sustainable development and ecological environmental protection in the Yellow River Delta region.
Material and Methods
Study Area and Data Sources
Study Area
The Yellow River Delta (36º55′N to 38º10′N, 118º07′E to 119º10′E) predominantly encompasses Dongying, Kenli, Hekou, Lijin, and Guangrao counties in Shandong Province [14], as depicted in Fig. 1. Adjacent to the Bohai Sea, the Yellow River Delta boasts the most comprehensive wetland ecosystem within the Chinese warm temperate zone, thus exhibiting a diverse array of vegetation types [15]. Vegetation in the Yellow River Delta is categorized into artificial and natural vegetation. Artificial vegetation comprises crops and man-made forests. Crops such as wheat, corn, and rice are cultivated, while artificial forests primarily consist of acacia and dry willow trees. Crop cultivation follows a biennial cycle, employing crop rotation between wheat and corn [16]. Natural vegetation includes winged alkali flora, tamarisk scrub, reed scrub, and others [17]. The Yellow River Delta, characterized by its flat terrain and abundant mineral, hydrological, and land resources, offers ample space for development.
Data Source and Processing
The remote sensing data utilized in this study encompasses Landsat-5, Landsat-7, and Landsat-8, spanning the years 2002, 2008, 2016, 2018, and 2021, with a spatial resolution of 30 meters. For Landsat-5 TM and Landsat-7 ETM+ data, bands {1, 2, 3, 4, 5, 7} were selected. For Landsat-8 OLI data, bands {1, 2, 3, 4, 5, 6, 7} were utilized. The Landsat data employed are geometrically fine-corrected Level-1 Terrain (L1T) products, necessitating pre-processing steps such as radiometric calibration and atmospheric corrections before subsequent analysis.
Research Methods
The technical process of land use information extraction in this study is shown in Fig. 2, and the main steps are (1) image downloading and preprocessing;
(2) feature extraction; (3) classification scheme development and sample selection; (4) random forest model training; (5) land use information extraction and accuracy evaluation.
Feature Extraction
The features utilized in this study encompass surface reflectance features, spectral index features (NDVI, SAVI, and MNDWI), tasseled cap transform features (brightness, greenness, and humidity), and second-order texture features.
Spectral indices can be used to assist in distinguishing between different classes of features [18]. In this study, the spectral indices utilized include NDVI (Normalized Difference Vegetation Index) [19], SAVI (Soil Adjusted Vegetation Index) [18, 20], and MNDWI (Modified Normalized Difference Water Index) [21], applied as follows:
where ρ(Green), ρ(Red), ρ(NIR), and ρ(MIR) respectively represent the surface reflectance of the green, red, near-infrared, and mid-infrared, with L being an empirical value set to 0.5.
The tasseled cap transformation is a multispectral transformation method where the transformed axes point to directions associated with vegetation and soil [22, 23]: Brightness, Greenness, and Wetness [24]. In this study, the brightness, greenness, and wetness features obtained through tasseled cap transformation are utilized to enhance the separability between spectral features of similar land covers, thereby aiding in the identification of different types of vegetation.
Texture features play a crucial role in improving the accuracy of feature extraction [25, 26]. In this study, six minimally correlated texture features based on the Gray-level co-occurrence matrix (GLCM) are utilized: mean (MEA), standard deviation (STD), entropy (ENT), angular second-order moments (ASM), dissimilarity (DIS), and homogeneity (HOM) [27, 28], as shown below:
where N represents the number of gray levels, P denotes the normalized gray-level co-occurrence matrix of size N×N, and P(i, j) represents the standardized gray-level value in the ith row and jth column of the matrix. This study only conducted second-order texture feature extraction for the green band [26], using a window size of 5 pixels by 5 pixels [18].
Classification Programming
The classification system developed in this study comprehensively considers multiple factors:
(1) understanding the historical development of land use, geographic conditions, vegetation types, and farming methods in the study area; (2) analyzing the actual interpretability of the images used; (3) referring to the national standard “Classification of Land Use Status GB/T 21010-2007.” The developed classification scheme is shown in Table 1.
Sample Selection
In this study, a stratified random sampling method was employed for sample selection, with the sampling units being small polygonal blocks. Each small polygonal block contains 2 to 9 pixels, and it is assumed that all pixels within each small polygonal block belong to the same land cover type.
Random Forest Algorithm
Due to its high accuracy and effectiveness in handling large datasets, the random forest algorithm has been widely applied in the field of remote sensing [13, 29]. The principles of the random forest classification [13, 30] are illustrated in Fig. 3.
When using the random forest method for classification, it is necessary to set the number of decision trees forming the random forest (ntree) and the number of predictor variables randomly selected for decision tree splitting (mtry). In this study, ntree was set to an empirical value of 500, and mtry was set to the square root of the total number of input features [31].
Accuracy Evaluation
After training the model to classify remote sensing images, the accuracy of the classification results is evaluated using evaluation metrics generated from the confusion matrix: Overall Accuracy (OA) and Kappa coefficient (Kappa) [32]. The calculation formulas are as follows:
where c represents the number of categories (same below), Gii represents the number of test samples of category i correctly categorized into category i (i = 1, 2, ..., c) (same below), and N represents the total number of samples in the test set (same below).
Table 1.
Yellow River Delta classification scheme.
where P0 represents the observed agreement rate, Pe represents the expected agreement rate, Gij represents the number of test samples in category i that are categorized as category j, and Gji represents the number of test samples in category j that are categorized as category i.
Results and Discussion
Classification Results and Accuracy Validation
The land use classification results of the Yellow River Delta wetland obtained from applying the model to the images of 2002, 2008, 2016, 2018, and 2021 are shown in Fig. 4. The accuracy evaluation metric calculation results are presented in Table 2.
From the calculation results, it can be observed that the maximum OA was 91.4% in 2016, while the minimum was 89.0% in 2002, with an average of 90.5%. The maximum Kappa was 0.902 in 2016, while the minimum was 0.877 in 2002, with an average of 0.891. Hence, the land use classification results are quite satisfactory. This is mainly attributed to the multi-temporal data and selected features, which enhance the separability between different land cover types. Additionally, the random forest algorithm is capable of handling high-dimensional feature input data with excellent accuracy.
According to the classification results combined with actual field research, from 2002 to 2008, the built-up area was primarily concentrated in the central part of the Yellow River Delta, including Xicheng in Dongying District, Kenli Town in Kenli District, and the urban area of Hekou District. Subsequently, it was expanded to Dongcheng in Dongying District, Shikou Town, the urban area of Guangrao County, Dawang Town, Daizhuang Town, Shengtuo Town in Kenli County, and Chenzhuang Town and Yanwo Town in Lijin County. The expansion of Dongcheng stemmed from economic development driven by the Shengli Oilfield and the relocation of municipal government, while the expansion of townships was due to economic prosperity and increased infrastructure investment.
Wheat fields are predominantly cultivated in Guangrao County, where the soil is fertile and rivers provide water. Paddy fields are located in Xicheng, Dongying District, the urban area of Hekou District, and the high-standard farmland project area, benefiting from water sources from reservoirs and ponds. Tamarisk shrubs are found in the northeast, woodlands in the southwest of Gudao Town in Hekou District and the Yellow River Estuary National Nature Reserve, reed meadows in the northern part of the urban area of Hekou District and the reserve, and Suaeda meadows are near coastal mudflats. Unused land from 2016 to 2021 has been situated in the eastern part of Dongying District near the Guangnan and Guangbei reservoirs, the eastern part of Kenli District near inland areas, the northeastern part of Lijin County, and the western and southwestern parts of Hekou District. Tidal flats, breeding aquatics, and salterns are located in the northeast and east, adjacent to the Bohai Sea, and suitable for marine aquaculture. The fresh waters are found in the eastern part of Dongying District and Hekou District.
Table 2.
Accuracy of land use mapping in the Yellow River Delta.
| Year | OA | Kappa |
|---|---|---|
| 2002 | 89.0% | 0.877 |
| 2008 | 90.5% | 0.888 |
| 2016 | 91.4% | 0.902 |
| 2018 | 90.4% | 0.890 |
| 2021 | 91.2% | 0.900 |
The article utilizes stacked histograms to illustrate the proportion of different land use types, as shown in Fig. 5. From 2002 to 2016, the proportion of cropland (wheat fields, dry farmland, and paddy fields) was the largest, indicating the significant importance of crop cultivation. However, the proportion of wheat fields and paddy field was relatively small, suggesting that grain crops did not dominate the planting structure. This is mainly due to the high degree of soil salinization and alkalization in the study area, as well as the scarcity of available water resources, which restrict the cultivation of wheat and rice [33].
Analysis of Land Use Changes
The Yellow River Delta wetland exhibits significant changes in land use types, as illustrated in Table 3. Studies indicate that from 2002 to 2021, there has been a substantial increase in cropland, impervious surfaces, and waters, while woodland, grassland, tidal flats, and unused land have significantly decreased. The reduction in woodland is primarily attributed to its conversion to cropland, whereas unused land has mainly transformed into water, impervious surfaces, and cropland. Research suggests that the changes in land use types are influenced by multiple factors: location, hydrology, and soil conditions determine the overall pattern and trends of land use; population growth leads to the conversion of large areas of woodland and unused land into cropland; economic development results in the conversion of cropland around cities into built-up areas and the transformation of tidal flats and coastal saline-alkali land into breeding areas and salterns.
Analysis of Landscape Pattern Changes
Landscape patterns significantly impact the structure and function of ecosystems within a region [34, 35]. In this study, six landscape pattern indices were utilized, including the number of patches (NP) [34], patch density (PD), mean patch area (AREA_MN), landscape shape index (LSI), patch cohesion index (COHESION) [36], and aggregation index (AI) [35]. The calculation results are presented in Table 4.
From the calculation results, it can be observed that NP and PD in the Yellow River Delta decreased from 2002 to 2008, followed by a continuous increase from 2008 to 2021. Meanwhile, AREA_MN increased before 2008 and then steadily declined thereafter. These indices indicate that landscape fragmentation in the study area decreased before 2008 but has since been increasing. LSI exhibited a decrease followed by an increase around 2008, indicating that the shape of patches in the study area became simpler before becoming more complex. COHESION showed slight growth from 2002 to 2008, followed by a gradual decrease, with a slight Unit: acre increase from 2018 to 2021, suggesting that landscape connectivity increased initially, then decreased, and increased slightly again from 2018 to 2021. AI increased from 2002 to 2018 and then steadily decreased after 2018, indicating that different landscape patches were relatively dispersed from 2002 to 2008 and gradually aggregated thereafter.
Table 3.
Markov Transfer Matrix for the area of each land use type in the Yellow River Delta.
Table 4.
Landscape pattern index of the Yellow River Delta from 2002 to 2021.
In summary, the fragmentation, shape complexity, and aggregation of landscape patches in the Yellow River Delta wetland generally experienced a turning point around 2008, exhibiting a trend of decreasing followed by increasing. This is primarily attributed to the continuous expansion of cropland, impervious surfaces, and waters in the Yellow River Delta from 2002 to 2021, initially forming relatively large patches and maintaining a relatively continuous landscape with lower fragmentation. However, as activities continued, these patches were subdivided, leading to an increase in landscape fragmentation, shape complexity, and aggregation.
Conclusions
This study utilized Landsat-5/7/8 data and integrated surface reflectance features, spectral index features, tasseled cap transform features, and second-order texture features. Utilizing the random forest method, the images from 2002, 2008, 2016, 2018, and 2021 were classified into land cover types. The accuracy of the classification results was evaluated, and analyses of land use change and landscape pattern change were conducted. The main conclusions are as follows:
(1) The random forest algorithm achieved good results in land cover classification, with overall accuracy ranging from 89.0% to 91.4% and the Kappa coefficient ranging from 0.877 to 0.902.
(2) The built-up area in Dongying District and some townships expanded after 2008. Wheat fields are predominantly cultivated in Guangrao County, while paddy fields are found in Xicheng, Dongying District, the built-up area of Hekou District, and the high-standard farmland project area. Tamarisk shrubs are located in the northeast, woodlands in the southwest of Gudao Town in Hekou District and at the Yellow River Estuary National Nature Reserve, reed meadows in the northern part of the built-up area of Hekou District and the reserve, and suaeda meadows near coastal tidal flats. Unused land from 2016 to 2021 was distributed in the eastern part of Dongying District near Guangnan and Guangbei reservoirs, in the eastern part of Kenli District near the inland, in the northeastern part of Lijin County, and in the western and southwestern parts of Hekou District. Tidal flats, breeding areas, and salterns are mainly located in the north and east. Croplands dominated the area from 2002 to 2016, but the proportion of food cropland was relatively small.
(3) The land use change in the Yellow River Delta wetland was drastic from 2002 to 2021, with increases in cropland, impervious surfaces, and waters and decreases in woodlands, grasslands, tidal flats, and unused land. Reduced woodlands have mainly been converted to cropland, while unused land has been converted to water, impervious surfaces, and cropland.
(4) Landscape pattern indices generally experienced a turning point around 2008. NP, PD, and COHESION first increased and then decreased, while AI, AREA_ MN, and LSI first decreased and then increased. Thus, landscape fragmentation, shape complexity, and aggregation decreased from 2002 to 2008 and increased from 2008 to 2021.
Conducting mapping and landscape pattern evolution analysis in the Yellow River Delta wetland is of great significance for promoting regional sustainable development. This study will provide scientific support for regional development and help establish a more harmonious development model.
