High-precision remote sensing mapping of aeolian sand landforms based on deep learning algorithms

2.1 Overview of the study area

The western part of Horqin Sandy Land (118°31′40″E–120°45′10″E, 42°39′29″N–43°24′44″N) is located in Wengniute Banner, Inner Mongolia Autonomous Region, with a total area of 5788.63 km² (Figure 1). The terrain is high in the west and low in the east, and the elevation is between 284 and 909 m. The study area is located in a temperate semi-arid continental climate zone with an annual average temperature is 5.2–6.4°C with an annual average precipitation of 300–400 mm. The precipitation is mainly in summer. The average annual wind speed is 3.5 m/s, and the maximum wind speed is 21.7 m/s. The frost-free period is from 140 to 150 days, and the water heat and strong wind conditions have become the main factors for the formation and development of sandy land. The area lies between the Xar Moron River and the Shaolang River. The soil type is mainly sandy soil. The main vegetation is the large scale of grassland and man-made forest including Ostryopsis davidiana decaisne and Pinus tabuliformis carriere. Nowadays, thanks to the national policies on the protection of grass, sand, tourism, and other related industries, the desertification condition has been effectively improved [18].

Figure 1

Location of study area.

2.2 Data source and processing

Since the successful launch of the Sentinel binary star system in 2014, its higher spectral and spatial resolution image data have been widely used in ground object recognition research in lake and wetland and mountain vegetation [19,20]. Compared with Landsat series data, its resolution has been significantly improved. In this study, it was merged with DEM data to significantly improve the accuracy of remote sensing classification of sandy landforms [21,22].

Sentinel-1 radar images were downloaded from the official website of Google Earth Engine (https://earthengine.google.com), and the VV and VH backscattering coefficient raster images of the western Horqin Sandy Land in 2015 and 2020 were obtained, respectively [23]. Sentinel-1 is an all-weather, all-weather radar imaging system, including two satellites A and B. Sentinel-1 A was successfully launched on April 3, 2014, and then, Sentinel-1B was successfully launched on April 25, 2016. The Sentinel-1C-band-based imaging system uses four imaging modes (with a resolution of up to 5 m and a width of 400 km) for observation. Sentinel-1 operates in a near-polar sun-synchronous orbit. The revisit period of one satellite is 12 days, and observations of two satellites complement each other, and the revisit period is 6 days. The Sentinel-2 multispectral high-resolution remote sensing image (Table 1) was downloaded from the ESA data website (https://glovis.usgs.gov). The Sentinel-2 high-resolution multispectral imaging satellite is divided into two satellites, 2A and 2B. 2A was launched by the “Vega” carrier rocket on June 23, 2015, and 2B was launched by the “Vega” carrier rocket on March 7, 2017. The Sentinel-2 satellite carries a multispectral imager (MSI) with 13 spectral bands. The spatial resolution of bands 2, 3, 4, and 8 is 10 m, and the spatial resolution of bands 5, 6, 7, 8A, 11, and 12 is 20 m. The band 1, 9, and 10 is 60 m, and the revisit period of one satellite is 10 days. The observations of two satellites complement each other, which can form a revisit period of less than 5 days for terrestrial monitoring. The Sentinel image was calibrated based on atmospheric, terrain, and super-resolution fusion by using official processing software SNAP (V8.0) of ESA, and the data cropping, mosaic, and convolutional neural network deep learning classification were completed with ENVI (V5.6) [24]. GDEM (V3) data were downloaded from Geospatial Data Cloud (http://www.gscloud.cn/search) [25]. Based on Sentinel and GDEM (V3) remote sensing images, Sentinel-1VV, VH polarization backscattering coefficients, DEM and Sentinel-2 spectra, vegetation, water bodies, and textures were extracted (Table 2). The study acquired 2,786 survey samples through Google Earth observations and field surveys in October 2020, including information about types of sand dune landform, latitude and longitude (error < 5 m), vegetation and corresponding photos, 464 longitudinal dunes, 342 flat sand lands, 440 nest-shaped dunes, 466 mild undulating sand lands, 230 parabolic dunes, and 724 barchan dune and dune chains, which are randomly allocated as training samples and verification points in a 6:4 ratio.

3.1 Technical process for classification of sand dune landform types

Based on high-resolution Sentinel-1 and Sentinel-2, GDEM (V3) remote sensing images and verification point data of 2,768 aeolian landform fields collected from 2015 to 2020 (Figure 2), this article calibrated the atmosphere and terrain images and obtained waveband characteristics including radar, DEM, visible light, red edge, vegetation index, water index, and texture [26]; by using convolution neural network deep learning algorithm, this article mapped remote sensing classification of aeolian sand landform in the west of Horqin Sandy Land; the classification results are calibrated by wind direction, vegetation, terrain, and other factors, and the verification points were used for accuracy verification, and the evolution map of the aeolian landforms in the west of the Horqin Sandy Land from 2015 to 2020 is obtained, and the deep learning and high-resolution remote sensing of the aeolian sandy landforms in the west of the Horqin Sandy Land are discussed. The advantages of images combined with information recognition and the evolution process of sandy land.

Figure 2

Flow chart.

3.2 Convolutional neural network algorithm

CNN is a deep learning neural network designed for processing high-dimensional grid data [27]. It can start directly from the original pixels and recognize the laws of vision with little preprocessing. The most significant difference between CNN and other deep learning models is that it has the characteristics of shared weights and sparse connectivity. Shared weight can significantly reduce the number of parameters. Sparse connectivity structure can simplify the network calculation and effectively extract primary visual features and spatial information such as directional line segments, inflection points, and endpoints. Compared with classic neural networks, CNN has multiple convolutional layers and pooling layers for calculation [28].

The convolutional layer is composed of multiple convolution kernels, which are used to calculate different feature maps. Each neuron of the feature map is connected to the area of the adjacent neuron in the previous layer. Such adjacent neurons are called receptive fields of neurons in the previous layer. A new feature map will be obtained by convolving and learning the feature map of the first layer and by applying the nonlinear activation function to the convolution result. In convolution processing, to generate each feature map, the convolution kernel is shared by all spatial positions of the input. The complete feature map is calculated by all the convolution kernels. The formula is expressed as follows:

where Y is the output image pixel matrix, X (…×…) is the nonlinear function, a is the weight matrix, b is the input image pixel matrix, n is the number of bands, and c is the offset matrix.

The pooling layer, also known as the downsampling layer or subsampling layer, is another core structural layer of CNN. By sampling the space of each dimension of the input data, the data scale can be further reduced, the local linear transformation is invariable with the input data, and the generalization processing of the network is enhanced. Each feature map of the pooling layer is connected to the corresponding feature map at the previous convolutional layer. After multiple convolutions and pooling, the dimensionality of the input features has been reduced to the point where the feed-forward network can be used directly for processing. At the fully connected layer, the feature maps of all two-dimensional images are spliced into one dimension as fully connected input. Each neuron in the fully connected layer is connected to the neuron at the previous layer, but the neurons at the same layer are not connected.

The U-net deep learning model has the advantages of the fast training speed, small training data demand, and maximum fidelity and restoration of original image information. The U-net model under the tensor flow framework is used for deep learning of image features of multiscale object-oriented segmentation. The model follows a typical convolution neural network structure, which forms a symmetrical structure of “shrinking” and “expanding” through multiple convolutions, downsampling, and upsampling can construct high-level feature learningand feature training library. The image is convolved with a 3 × 3 convolution kernel and a modified linear unit activation function, and the downsampling and upsampling are completed, respectively, by the 2 × 2 maximum pooling and matrix cascade methods. At the same time, the convergence speed of the model is controlled by setting the learning rate and dropout to prevent overfitting.

3.3 Accuracy verification

Considering the complexity of the land type in the study area, combined with the verification sample data to make a 64 × 64 pixel size image to form a labeled data set, including three parts: training data set, verification data set and test data set. Among them, the training data set is used to train the U-net deep learning network. The verification data set is used to verify and debug the U-net deep learning network. After the debugging of the U-net deep learning network is completed, the U-net deep learning model is generated, and then, the test data set is used to test the model and evaluate its running results.

By means of confusion matrix, the overall classification accuracy, Kappa coefficient, mapping accuracy, and user accuracy of the classification results are obtained. By comparing extracted category attributes of the corresponding position in the classification with each test pixel’s position and category attributes, the accuracy of classification results is obtained and evaluated [29]. The formula for overall accuracy (OA) and Kappa coefficient is given as follows:

where T p is a positive sample that the model correctly classifies, T n is a negative sample correctly classified by the model, F p is a negative sample that the model incorrectly classifies, and F n is a positive sample incorrectly classified by the model.

where P o is the overall classification accuracy, a is the number of actual samples for each category, b is the number of samples predicted by each category, x is the total number of categories, and n is the total number of samples.

4.1 Spatial distribution of aeolian sand landform types

The aeolian sandy landform in the west of Horqin Sandy Land mainly includes longitudinal dune, flat sand land, mild undulating sand land, nest-shaped dune, parabolic dune, crescent dune, barchan dune, and dune chain. There are six types of sand dune landform distributed alternately and relative to each other (Figures 3 and 4). The longitudinal dunes are mainly distributed in the middle of the study area, with the largest area of 1735.62 km² among these six types of sandy dune landform (Table 3); the nest-shaped dunes and barchans dune and dune chains are mainly distributed around the longitudinal dunes, with an area of 902.07 and 110.63 km², respectively; flat sand land and parabolic dunes are mainly distributed in the west and south of the study area, with an area of 51.32 and 49.57 km², respectively; mild undulating sand land is mainly distributed in the northeast of the study area, with an area of 251.38 km².

Figure 3

Spatial distribution of each aeolian landform type in western Horqin Sandy Land in 2015.

Figure 4

Spatial distribution of each aeolian landform type in western Horqin Sandy Land in 2020.

4.2 Dynamic evolution of aeolian sand landform types

The aeolian sandy landforms in Horqin Sandy Land are still in the process of continuous evolution. From 2015 to 2020, the areas of nest-shaped dunes, parabolic dunes, and flat sand land increased by 233.34, 18.26, and 16.75 km², respectively (Figure 5). The increased areas are mainly concentrated in the central, southern, and southwestern parts of the study area. In the same period, the area of mild undulating sand land, longitudinal dune, and barchans dune decreased by 8.97 200.14, and 148.51 km², respectively. The reduced areas are mainly distributed in the northeast, middle, and west of the study area.

Figure 5

Sankey diagram of the conversation for each class type in the study area.

4.3 Accuracy evaluation of classification of aeolian sand landform

With the help of the confusion matrix, the overall classification accuracy, Kappa coefficient, mapping accuracy, and user accuracy of the classification results are obtained, and the classification accuracy is calculated by checking the position of each test pixel and comparing the classification with the classification of the corresponding position in the classification image. The accuracy of the classification results is evaluated [30].

Based on the sample data of 1,114 ground survey verification points, a confusion matrix analysis was performed on the deep learning remote sensing classification results of the aeolian sand landform types in the western Horqin Sandy Land (Table 4). The results show that the effect of CNN deep learning on aeolian landforms has a significant effect on remote sensing classification. The accuracy and Kappa coefficient are as high as 95.51% and 0.8961, respectively. The mapping accuracy and user accuracy of each dune landform type are above 90%, including barchan dune and dune chain, parabolic dune, and longitudinal dune. The mapping accuracy of sand dunes and sand ridges are 97.58, 96.29, and 95.33%. The user accuracy of barchan dune and dune chain, mild undulating sand land, and longitudinal dune are 96.79, 96.24, and 95.49%, respectively. Therefore, the CNN deep learning algorithm combined with Sentinel high-resolution remote sensing images can significantly improve the remote sensing classification accuracy of sandy landforms.

5.1 Discussion

The evolution process of sandy land is manifested as the change of land surface landscape pattern, which can be characterized by landscape index [31]. Li et al. [29] believed that landscape fragmentation and spatial heterogeneity have a positive effect on the desertification process. Li et al. [29] also proposed that the area of mobile dune patches is positively correlated with the degree of desertification. The Horqin Sandy Land is characterized by a decrease in the total area of aeolian landforms and an increase in fixed and semi-fixed dunes such as shrub-grass dunes, showing the development of deserts in the direction of oasis, which is consistent with the report by Yue et al. [32]. The Horqin sandy landscape is in a dynamic balance under the interactive erosion of wind and water and maintains its development stability. Moisture is the first dominant factor restricting the growth and survival of plants, and it has an important impact on the structure of the regional sand flow field and sand transport [33]. In addition, the construction of reservoirs in the upper reaches of the river, such as the Hongshan Reservoir in the upper reaches of the Laoha River, reaches in the Shanwanzi Reservoir, Gangouzi Reservoir, and Ulan Besu Reservoir built in the middle and lower reaches of the river will inevitably affect the downstream runoff and change the water and sand conditions of the river. If the flow is reduced, the river bed is dry, and the broad floodplain is exposed, the river bank sand dunes can develop. Surface water decreases, groundwater level drops, and the groundwater level of yards and sand marshes will also drop. Wet-loving plants will decrease and drought-tolerant plants will increase. In particular, trees will die, vegetation coverage will be less, sand dunes will be activated, and desertification will be repeated and will continue to expand [34,35]. As a unique geographical unit, the Horqin Sandy Land has an obvious trend of warming and drying. Zhao et al. [36] also confirmed this point in the study of climate change in the Horqin Sandy Land. Increasing temperature and increased evaporation will lead to a decrease in water volume, degradation of vegetation, and increased desert activities [37], which plays an important role in inhibiting the evolution of desertification. At the same time, it should be pointed out that short-term and high-intensity human activities can accelerate or delay the evolution of sand dunes on the surface. Wang et al. [38] surveyed the Horqin area and showed that due to human adjustment of land use structure, there was a need for the implementation of measures such as returning farmland to forests and grasses, restraining overgrazing, and so on. The desertification of the land in this area has the reversed trend.

Based on Sentinel high-resolution remote sensing images, this research applies the CNN deep learning algorithm to the remote sensing classification of aeolian sand landforms and also conducts verification studies in the western part of Horqin Sandy Land. The mapping is accurate, indicating that it is feasible to be extended to the research of remote sensing mapping of aeolian landform types in other regions. Deep learning is currently the most advanced method of remote sensing classification and is widely used. Compared with traditional supervised classification and object-oriented classification, the accuracy is increased by 46% [28,39].

The construction of a deep learning model involves multiple hyperparameters (such as the division ratio of data used for training, verification, and testing of the model, the size of the labeled data set image, the size of the convolution kernel, etc.), and the settings of the model are different, and the performance of the built model is also different [40,41]. In the experiment, the hyperparameter setting of the deep learning model is mainly selected based on the researcher’s experience, which is highly subjective. Experts and scholars have conducted effective research on how to determine the hyperparameters of deep learning models, but their study objects are mostly on cultivated land and forests [42–44], and the studies on the types of sandy landforms are insufficient. Therefore, an in-depth research is urgently needed. Related research work to further improve the accuracy of drawing. The deep learning method in this article can be extended to the high-precision classification of other sandy landforms in arid areas.