Retrieval of Remotely Sensed Sediment Grain Size Evolution Characteristics along the Southwest Coast of Laizhou Bay Based on Support Vector Machine Learning

Abstract

Grain size is the basic property of intertidal zone sediment. Grain size acts as an indicator of sedimentary processes and geomorphological evolution under human and nature interactions. The remote sensing technique provides an alternative for sediment grain-size parameter monitoring with the advantages of wide coverage and real-time surveying. This paper attempted to map the distributions of three sediment grain size contents and the mean grain size with multitemporal Landsat images along the southwestern coast of Laizhou Bay, China, from 1989 to 2015. Considering the low correlations between the measured reflectance and grain-size parameters, we used a support vector machine (SVM) to develop a nonlinear calibration model by taking several band indices as input variables. Then, the performance of the back propagation neural network (BPNN) was determined and discussed with that of the SVM. The SVM performed better than the BPNN in calibrating the four grain-size parameters based on a comparison of R² and the root-mean-square error (RMSE). Moreover, an atmospheric correction algorithm originally proposed for case II water enabled the TM\ETM+ images to be precisely atmospherically corrected in this study. The SVM-mapped spatial-temporal grain-size variation showed a coarsening trend, which agreed with that obtained during in situ measurements in a former study. The changes in Yellow River discharge and precipitation associated with the coarsening trend were further analyzed. The yielded results showed that the coarsening trend and reduction in tidal flat area might be aggravated with overutilization. More reasonable planning would be necessary in this case.

1. Introduction

The intertidal zone is an important component of silt coastal landforms. Siltation occurs in the intertidal zone under the interaction of rivers and tides [1], and this process closely relates to hydrodynamics and sedimentology [2]. Changes in sediment grain-size parameters act as indicators of the change rate of sediment pollutants being discharged to the marine environment [3]. Similarly, these changes closely relate to nutrient cycling by affecting primary the production, survival, and growth of benthic animals, flatfish, and waders [4]. There is an urgent need for synoptic knowledge of the sediment grain-size parameter distribution in the intertidal zone [5].

Sediment grain-size parameters vary spatially and temporally in the intertidal zone [6]. In situ measurement is a general method for sediment grain-size parameter surveys. However, sampling sites following a well-defined strategy are inherently discrete and isolated. Whether spatially or temporally, it is difficult to gain an accurate and continuous change trend. In addition, the field survey is time-consuming and laborious. It is essential to develop a remotely sensed approach to periodically provide a quantitative sketch of sediment grain-size parameter distributions.

Intertidal sediment types have been classified by multispectral satellite remote sensing [7], multispectral airborne remote sensing [8], and hyperspectral airborne remote sensing [9,10]. For example, sediment types were classified into different types by supervised classification methods [11]. It was acknowledged that the validity of the training datasets dominated the supervised classification methods. The spectral features of these particular training datasets should typically work independently. Thus, user interaction is unavoidable. Unsupervised classifications with little user interaction were also adopted in producing different clusters by setting a different number of clusters and allowable dispersion around cluster centers [12,13]. The related studies focused on sediment types classification rather than grain-size parameter quantification.

Commercial satellite images with high spatial resolution have been effective for grain-size parameter retrieval, such as IKONOS [14] and WorldView-2 [15]. However, the related commercial satellite data are not free for most scholars. Sensors at low altitudes, such as aircraft and unmanned aerial vehicles (UAVs), have been validated in quantifying sediment/soil grain size by using their hyperspectral reflectance in the VNIR (visible/near-infrared) band or SWIR (shortwave infrared) band [16,17,18,19] or texture properties [20,21]. However, the narrow spatial coverage and inaccessibility of hyperspectral images decreased their utility in surveying the spatial-temporal variation in sediment grain-size parameter distributions. An alternative satellite remote sensing approach based on multispectral images rather than hyperspectral images is needed. Multispectral images were accessible and could provide periodic sketch data. However, the spectral resolution of multispectral images was inherently lower than that of hyperspectral images, resulting in low correlations between band reflectance and grain-size parameters [17]. Thus, it is difficult to develop a linear calibration model with a certain band of multispectral images. Advanced algorithms based on different band combinations were adopted. For example, multispectral satellite images have been adopted to map the sediment distribution in the intertidal areas of Wash, East England (UK) [22]. Multiple regression (MR) and spectral mixture modeling were used to quantify the amounts of each sediment type within a pixel. The yielded result was acceptable when compared with the measured sample sites. Similarly, MR calibration algorithms developed with the reflectance of TM2 and TM5 succeeded in mapping the mud content of the sediment in the Westerschelde (southwest Netherlands) with a maximum R² of 0.72 [23]. Regression-based algorithms were developed for monitoring the sediment grain-size of intertidal flats, using information from both space-borne microwave (SAR) and optical/shortwave infrared remote sensing [24]. Moreover, principal component analysis (PCA) and wavelet neural networks (WNNs) performed well in sediment grain-size parameter retrieval along the muddy coast of the central Jiangsu Province with HJ-1 satellite images [25]. These studies demonstrated the utility of multispectral satellite images in mapping grain-size parameters.

This paper attempted to map the spatial-temporal variation in sediment grain-size parameter distributions based on Landsat images on the southwestern coast of Laizhou Bay, China, from 1989 to 2015. The main goals are described as follows: to analyze the correlations between sediment grain-size parameters and in situ reflectance measurements; to develop a nonlinear calibration model with Landsat band reflectance by using a support vector machine (SVM), which was subsequently compared with a back propagation neural network (BPNN); to retrieve sediment grain-size parameter distributions based on the validated SVM model; to analyze the spatial-temporal changes in sediment grain-size parameter distributions associated with the changes in Yellow River discharge and precipitation over 26 years. A flow diagram of the methodological steps is sketched in Figure 1.

2. Materials and Methods

2.1. Study Area

The study area, the southwestern coast of Laizhou Bay, is jointly dominated by coastal currents, Yellow River discharge, and precipitation. This area lies in the northeastern part of the Yellow River Delta, a continental weak alluvial tidal delta. Laizhou Bay sea has significant continental climate characteristics, which are greatly affected by cold waves in winter and hot waves in summer [26]. Due to climate change, the fresh water supply in the southern coast of Laizhou Bay fluctuates, and the decrease in rainfall leads to the decrease in runoff into the sea and aggravates the intensity of seawater invasion, and then leads to continuous degradation of the whole coastal wetland ecosystem. The landforms along Laizhou Bay can be divided into three categories: terrestrial landforms, buried landforms, and coastal landforms. Among the landforms, the accumulation plain is the most widely distributed area. Buried landforms are mainly buried ancient channels, which are the main channels of seawater invasion inland. The coastal landform mainly includes bedrock coast, sandy coast, and silty silt coast [27]. The intertidal zone is roughly continuous along Laizhou Bay and unvegetated. The estuary, intertidal zones, and landforms in the Yellow River Delta have frequently changed in recent decades. With the inclusion of the Yellow River Delta Efficient Ecological Economic Zone in China’s National Strategic Planning Area, urban planning, industrial structure and layout, environmental protection, and resource utilization urgently need to understand future intertidal zone sedimentary processes and predict coastal landform erosion and siltation trends.

The Yellow River Delta coast has a typical weak tide. The coastal tide is mainly dominated by the M2 tidal wave. There is no tidal point in the nearshore seawater north of pile No. 5. The average tidal range along the deltaic coast is approximately 1 m (0.6–1.3 m). Most of the shores have irregular semidiurnal tides. Only the shores near Shenxiangou have irregular full-day tides. The estuary during the flooding period is dominated by the reciprocating flow parallel to the river channel. The coasts on both sides of the Yellow River mouth are reciprocating flows of roughly parallel shorelines [28]. Depending on the observations at the northern end of the Yellow River Harbor Terminal (water depth 7 m) in 1987, the wave frequency with a wave height greater than 1.5 m was 11.8%. The normal wave was SSE-S [29]. The strong wave was NE-NNE. The maximum wave was NE with a height of 5.3 m and a cycle of 8 s. The sediment, which is transported by the Yellow River from the Loess Plateau, is mainly composed of sand, silt, and clay. A grain size of less than 0.063 mm accounts for 94.2% of the total.

2.2. Field Measurements and Laboratory Analysis

We conducted 30 field surveys to measure the surface sediment reflectance from 2 March 2013 to 19 July 2014. Sediment samples were synchronously collected from the muddy intertidal zones on the southwestern coast of Laizhou Bay. In situ reflectance measurements and sediment sample collections were conducted during ebb tide or low tide off the southwestern coast of Laizhou Bay with the surface of the intertidal zone exposed. In this paper, a total of 238 sets of intertidal zone surface sediment samples and their reflectances were obtained (Figure 2). We divided the study area into two parts, the west coast and the south coast, to delineate the spatial variation in grain-size parameters. The Xiaoqing River, which is recognized as the boundary of the abandoned southern Yellow River subdelta (ASYS) and the plain of south Laizhou Bay (PSLB), was set as the boundary of the two parts (Figure 2).

2.2.1. Sediment Grain-Size Measurement

Sediment samples distributed in the soil layers at 5–10 cm depth were gathered. Each sample was mixed with five adjacent samples and placed in a numbered polymer bag. We used a Mastersize 2000 laser analyzer to measure grain size distribution, including clay, silt, sand, mean grain size, sorting coefficient, kurtosis, and skewness [30] (Folk and Ward 1957). The principle is described as follows: The intensity of scattered light was inversely proportional to the diameter of the particle under laser beam irradiation. The intensity of scattered light was logarithmically attenuated as the particle size increased. By accepting and measuring the energy distribution of the scattered light, the particle could be obtained. The measuring range of the particle size analyzer was from 0.02 to 2000 μm. In the measurement interval, a total of 100 particle size percentage data points could be generated. The percentage of each particle size was given. The particle size distribution curve was also provided, as were the cumulative curves. Compared with other particle size analyzers, the Mastersize 2000 laser particle size analyzer had a wide range of measurements, high precision, lower sample volume, faster data acquisition, higher measurement accuracy for small particles, and an overall error of less than 1%.

We added 10 mL of 10% H₂O₂ to the sediment sample (0.2~0.5 g) and heated it on a hot plate to fully react to effectively remove the organic matter. Next, we added 10 mL of 10% HCl and boiled it to react and fully eliminate carbonate. The beaker was filled with distilled water and permitted to stand overnight with the supernatant removed. Ten milliliters of a sodium hexametaphosphate dispersant with a concentration of 0.05 mol/L was added and shaken for 10 min in an ultrasonic cleaner. The dispersant was later measured by using a Mastersize 2000 particle size analyzer. In this experiment, the same treatment method was applied to all samples; that is, each sample was subjected to 5 repeated measurements. Several indices, including the mean grain size (Mz), sorting coefficient (σ₁), kurtosis (KG), and skewness (SK_I), were obtained with Equations (1)–(4), respectively.

Mz = (ϕ₁₆ + ϕ₅₀ + ϕ₈₄)/3

(1)

σ₁ = (ϕ₈₄ − ϕ₁₆)/4 + (ϕ₉₅ − ϕ₅)/6.6

(2)

SK₁ = (1/2)∗(( ϕ₈₄ + ϕ₁₆ − 2ϕ₅₀)/(ϕ₈₄ − ϕ₁₆) + (ϕ₉₅ + ϕ₅ − 2ϕ₅₀)/(ϕ₉₅ s− ϕ₅))

(3)

KG = (ϕ₉₅ − ϕ₅)/(2.44∗(ϕ₇₅ − ϕ₂₅))

(4)

where ϕ_i is a log transformation of grain diameter (D) with base two. The contents of clay, silt, and sand and the mean grain size were subsequently obtained, as shown in Figure 3. From the south coast of Laizhou Bay to the west coast, the clay content of sediment samples varied from 0 to 11.91% and showed an increasing trend (Figure 3a). The slit content of the sediment samples increased from 7.92% to 94.21% and showed a similar trend (Figure 3b). However, the sand content of sediment samples ranged from 90.95% to 1.14% and showed an obvious decreasing trend (Figure 3c). Similar to the clay content and the silt content, the mean grain size, which ranged from 1.69 φ to 6.05 φ, showed an increasing tendency (Figure 3d).

2.2.2. Measurement and Preprocessing of the Sediment Reflectance Spectrum

The in situ reflectance of sediment samples was measured by an ASD Field-Spec FR2500 Spectra-radiometer (Analytical Spectral Devices Inc., Boulder, CO, USA) with a spectral range of 350–2500 nm. A spectrum reflectance panel with 99% (Lab-sphere, Inc., North Sutton, NH, USA) served as the reference standard to adjust and optimize the spectrometer for incoming irradiation. Once the reflectance of one sediment sample was obtained, the spectrometer was optimized. The spectral resolution was 3 nm at 350–1050 nm and 10 nm at 1000–2500 nm. Each sediment sample was measured 10 times with a probe view angle of 25° and a probe distance of 15 cm on sunny and breeze days. Then, we used Savitzky–Golay smoothing to reduce the signal noise to obtain the averaged reflectance curve. The gained reflectance was subsequently matched to Landsat bands by using the spectral response function of the Landsat TM/ETM sensor [31].

2.2.3. Landsat Images and Preprocessing

In view of jet lag, eight cloud-free Landsat images with low water levels were selected (USGS, http://glovis.usgs.gov/, accessed from 13 February 1989 to 5 February 2015.) (

Table 1. Specification of the Landsat imagery used in this study.

Date	Sensor	Time (UTC)	Track/Path	Water Level (m NAP)	Tidal Stage
13 February 1989	TM	02:14	121/34	−1.60	incoming
26 March 1995	TM	01:51	121/34	−1.35	outgoing
10 December 1999	ETM+	02:34	121/34	−1.77	incoming
12 October 2001	ETM+	02:29	121/34	−1.80	outgoing
11 November 2006	ETM+	02:31	121/34	−1.55	incoming
3 November 2009	ETM+	02:32	121/34	−1.85	incoming
23 August 2012	ETM+	02:36	121/34	−1.93	incoming
5 February 2015	ETM+	02:40	121/34	−2.05	incoming

Tidal information accessed from 13 February 1989 to 5 February 2015 is from China Oceanic Information Network (https://ocean.cnss.com.cn/, accessed on 1 January 2022).

). Digital numbers in visible bands TM1/ETM+1, TM2/ETM+2, TM3/ETM+3, NIR band TM4/ETM+4, and shortwave infrared (SWIR) bands TM5/ETM+5 and TM7/ETM+7 were transformed to apparent reflectances in view of the gains and bias coefficients of scenes, solar azimuth angle, solar spectral irradiance, and Sun–Earth distance [32]. Then, the images were georeferenced to the World Geodetic System 1984 (WGS84) by using the nearest-neighbor resampling method with the selected ground control points. A pixel size of 30 m was retained in the rectified images.

Considering complex aerosol types at land–sea junctions, we adopted an improved atmospheric correction algorithm [33] rather than FLAASH (fast line-of-sight atmospheric analysis of spectral hypercubes). MODIS Terra data were adopted here to provide essential radiometric calibration information in eliminating air molecules and aerosol scattering from the apparent reflectance of Landsat images. Although the algorithm was originally developed for atmospheric correction of case II waters [33], it precisely described the information of the specific aerosol types in tidal flats due to the sensitivity and designed spectral bands of MODIS Terra data. Thus, it was more suitable for atmospheric correction in tidal flats than FLAASH, which had four restrictive aerosol models (rural, maritime, urban, and tropospheric). The flow is shown as follows.

The scattering radiance (L_r) due to air molecules (Rayleigh scattering) was calculated using the following methodology [34].

L_r(λ) = ρ_r(λ) F₀(λ) cos(θ₀)/π

(5)

where ρ_r(λ), F₀(λ), and θ₀ denote the Rayleigh scattering reflectance, extraterrestrial solar irradiance, and solar zenith angle, respectively. F₀(λ) could be obtained from the website (https://oceancolor.gsfc.nasa.gov/, accessed on 1 January 2022). θ₀ was presented in the header file of Landsat images. Then, MODIS Terra data were adopted to gain ρ_r(λ) of the Landsat images. The Rayleigh scattering reflectance ρ_r(λ)∝τ_r(λ)∝λ⁻⁴ [35]. The conversion factor β, from the MODIS/Landsat band response-averaged Rayleigh optical thickness (τ_r(λ)) to that of the center wavelength, was calculated as Equation (6):

β = τ_r(λ)/<τ_r(λ)>^MODIS =ρ_r(λ)/<ρ_r(λ)>^MODIS

(6)

where

τ_r(λ) = 0.008569λ⁻⁴ (1+0.0113λ⁻²+0.00013λ⁻⁴)

(7)

$$<{\tau }_r(\mathsf{\lambda }){>}^{\mathrm{MODIS}}=\int {\tau }_r\left(\mathsf{\lambda }\right)S\left(\mathsf{\lambda }\right)F_0\left(\mathsf{\lambda }\right)d\mathsf{\lambda }/\int S\left(\mathsf{\lambda }\right)F_0\left(\mathsf{\lambda }\right)$$

(8)

Then, β for MODIS bands were obtained. <ρ_r(λ)>^MODIS was accessible in the MODIS Rayleigh lookup tables (LUTs) (http://seadas.gsfc.nasa.gov/, accessed on 1 January 2022). ρ_r(λ) could be calculated by <ρ_r(λ)>^MODIS and β. L_r at the bands of Landsat images was derived by means of band response-averaged calculation [36] with Equation (5).

MODIS images acquired approximately synchronously with Landsat images were processed using the atmospheric correction algorithm [37] with the aerosol type provided in the aerosol LUTs. Aerosol scattering radiance (L_a(λ₅)) could be obtained after subtracting L_r(λ₅) from L_t(λ₅) under the assumption of black water at TM/ETM⁺ band 5 (λ₅). Then, ρ_a(λ₅) was calculated by Equation (5). The relationship between the single scattering reflectance (ρ_as(λ)) and ρ_a(λ), which were accessible in the aerosol LUTs, was applied to convert ρ_a(λ5) to ρ_as(λ5). Once ρ_as(λ5) was obtained, ρ_as(λ_i) (i = 1, 2, 3, 4) was interpolated with the spectral relationship in aerosol LUTs. Similarly, ρ_as(λ_i) was inversely converted to ρ_a(λ_i) using the relationship between (ρ_as(λ)) and ρ_a(λ). Then, L_a(λ_i) was derived from Equation (5). Surface reflectance (R_rs) could be derived from Equation (9):

R_rs(λ) = L_s(λ)/t_s(λ) ρ_r(λ) F₀(λ) cos(θ₀)

(9)

where surface radiance (L_s(λ)) and diffuse transmittance t_s(λ) are given in Equations (10) and (11), respectively.

L_s(λ) = L_t(λ₅) − L_r(λ₅) − L_a(λ₅)

(10)

t_s(λ) = exp (−(τ_r(λ))/2 + τ_oz(λ))/cos(θ₀))

(11)

where τ_oz(λ) is a multiplication between the ozone absorption coefficient at a specific wavelength (k_oz(λ)) [38] and the ozone content (U_oz). U_oz is accessed from 13 February 1989 to 5 February 2015 (http://oceandata.sci.gsfc.nasa.gov/Ancillary/Meterological/ (2012), accessed on 1 January 2022). The algorithm performed well by the comparison between mean spectra of 238 sites after atmospheric correction and that of in situ measurement (Figure 4).

2.3. Sediment Grain-Size Retrieval

2.3.1. Spectral Indices

The reflectance of the Landsat image bands (b1, b2, b3, b4, b5, and b7) was related to sediment grain-size parameters. The reflectance of the Landsat image bands (b1, b3, b4, and b5) was significantly correlated with sediment grain-size parameters (r > 0.141, p < 0.05) (Figure 5a). However, the correlation coefficients were not high enough to develop stable regression models, even if the highest correlation coefficient between b4 (center wavelength at 0.83 microns) and the sand content was adopted (Figure 5b).

Spectral indices, such as SI, DI, PI, RI, and NDI, were subsequently calculated based on the Landsat image band reflectance. The indices were also correlated with sediment grain-size parameters (

Table 2. Correlations between the band indices and the grain-size parameters.

Spectral Indices	Expressions	Correlation Coefficients(r)
		Mean Grain-Size		Sand		Silt		Clay
		rmax	rmin	rmax	rmin	rmax	rmin	rmax	rmin
R	original reflectance	0.55 *	0.02	0.58 *	0.07	0.58 *	0.07	0.52 *	0.07
SI	Ri + Rj	0.48 *	0.08	0.49 *	0.09	0.48 *	0.09	0.43 *	0.04
DI	Ri − Rj	0.65 *	0	0.71 *	0.04	0.70 *	0.04	0.64 *	0.01
PI	Ri × Rj	0.43 *	0.08	0.44 *	0.08	0.43 *	0.08	0.40 *	0.04
RI	Ri/Rj	0.52 *	0.01	0.59 *	0	0.59 *	0.01	0.50 *	0.01
NDI	(Ri − Rj)/(Ri + Rj)	0.14	0	0.15 *	0	0.16 *	0	0.10	0

* denotes a significant level (p < 0.05).

). By contrast, DI performed best in relating to the four grain-size parameters, with r_max values of 0.65, 0.71, 0.70, and 0.64. However, the correlation coefficients were still not high enough to promote robust linear regression models with corresponding R² values of 0.42, 0.50, 0.49, and 0.41. Thus, we used the SVM [39] to calibrate the grain-size parameters. Seventy-five percent of all samples were randomly selected as the calibration dataset, and the remaining 25% were used as the validation dataset.

2.3.2. Estimation Models

SVM was conducted by using R, SI, DI, PI, and RI as input variables. The basic idea of SVM is described as follows: to realize some nonlinear transformation by defining an appropriate kernel function, to map the input vector to a high-dimensional space, to find the optimal classification hyperplane in high-dimensional space, and to achieve the algorithm by the inner product in high-dimensional space [40]. Once an appropriate kernel function was selected, we obtained the classification function of high-dimensional space [41].

We used the scaleForSVM normalized preprocessing function to normalize the training dataset and validation dataset. All datasets were transformed into the range between ‘0’ and ‘1’. Because the radial basis kernel function (RBF) could solve the nonlinear problem of the sample into a higher dimensional space, it could solve the nonlinear problem that the linear kernel function could not solve [42]. Thus, RBF was adopted here. The core of the SVM model was to determine the best c and g combination so that the resulting root-mean-square error was reduced. The simplest solution to this issue was to obtain c and g discrete values in a certain range. We used the grid cross-validation method to locate the best combination of c and g here (Figure 6). The SVM optimal calibration and validation models for clay, silt, sand, and mean grain size are presented in Figure 7.

3. Results

SVM was subsequently used in mapping grain-size parameters with Landsat TM/ETM+ images from 1989 to 2015. The spatial and temporal distributions of the clay, silt, sand, and mean grain size contents are shown in Figure 8, Figure 9, Figure 10 and Figure 11, respectively.

3.1. Spatial Variation in Grain-Size Parameters

In 1989, the clay content increased perpendicular to the coastline on the west coast. In contrast, the clay content showed a declining trend on the south coast. The clay content was roughly the same coastwise in 1995, 1999, 2001, and 2006. In 2009, the clay content remained roughly the same coastwise on the west coast. However, the clay content showed a declining trend perpendicular to the coastline on the south coast. The spatial variation in clay content in 2012 was similar to that in 2009, as was that in 2015. The smaller difference was that the clay content in 2015 was slightly lower than that in 2012. In 1989, the silt content showed a similar trend as the clay content on the west coast and the south coast. The silt content remained roughly the same coastwise from 1995 to 2012. In 2015, the silt content was uniform coastwise, but there was a significant decrease from that in 2012. In 1989, the sand content also showed a similar trend as the clay content on the west coast and the south coast. From 1995 to 2009, the sand content remained roughly uniform coastwise. However, these values were higher in 1995 and 2009 than in 1999, 2001, and 2006. The spatial variation in sand content increased in 2012 and especially in 2015. The spatial variation in the mean grain size in 1989, 1995, 1999, 2001, and 2012 was higher than those in 2006, 2009, and 2015. Moreover, there was an obvious increase in 2015 from 2012.

3.2. Temporal Variation in Grain Size

Mean values of the retrieved grain-size parameters were calculated and applied to describe the temporal variation in grain size during 1989 and 2015 (

Table 3. Statistics of the SVM mapped grain-size parameters from 1989 to 2015.

	Clay (%)			Silt (%)			Sand (%)			Mean Grain-Size (φ)			Discharge (109 m3)	Precipitation (mm)
	Max	Min	Mean	Max	Min	Mean	Max	Min	Mean	Max	Min	Mean
1989	9.65	1.35	5.32	93.02	9.25	59.62	93.51	3.19	48.78	6.79	3.25	4.05	568.80	369
1995	5.47	4.40	5.03	76.52	30.82	54.18	72.18	20.40	45.65	5.40	2.48	4.11	567.90	728
1999	5.47	4.86	5.01	85.15	30.09	59.03	70.80	12.35	38.99	5.86	2.30	4.10	61.69	373
2001	5.62	3.85	5.02	78.49	19.98	55.46	76.82	15.23	41.45	5.54	2.48	4.14	40.89	414
2006	5.96	3.18	5.03	84.54	18.88	54.13	80.25	9.48	43.23	5.51	1.40	4.01	186.70	452
2009	6.25	2.62	4.89	83.76	21.75	51.12	75.72	7.89	45.57	5.91	2.54	4.27	132.90	653
2012	6.57	1.61	4.89	75.76	9.51	52.03	92.93	19.64	44.27	5.65	2.37	4.11	282.50	534
2015	5.07	2.89	4.64	75.65	5.60	37.42	94.19	16.86	58.46	5.47	1.94	4.10	133.60	595

). The clay content decreased slightly from 5.32% in 1989 to 4.64% in 2015 and showed a decreasing trend with an average annual reduction rate of 0.005. Similarly, the silt content decreased from 59.62% in 1989 to 37.42% in 2015 and showed a clear decreasing trend with an average annual reduction rate of 0.018. The sand content increased from 48.78% in 1989 to 58.46% in 2015 and showed an increasing trend with an average annual growth rate of 0.007. Therefore, the mean grain size increased from 4.05 in 1989 to 4.10 in 2015 and showed an increasing trend with an average annual growth rate of 0.005. In general, a coarsening trend occurred in the southwestern part of Laizhou Bay over 26 years. A previous study indicated that a similar transition from viscous sand to nonviscous sand occurred here, with a decrease in the clay content and viscosity/mud ratio from 2007 to 2013 [43].

The hydrodynamics of tides on the southwestern coast of Laizhou Bay have been stable since the Yellow River estuary changed north of Qingshuigou in 1996 [44]. Although the decreasing trend of the Yellow River annual discharge accessed form 1989 to 2015 (http://www.yrcc.gov.cn/, accessed on 1 January 2022) was opposite to that of mean grain size, the negative correlation between them was not significant, indicating that grain size coarsening had no correlation with the current Yellow River discharge (Table 3, Figure 12). Similarly, although the increasing trend of precipitation was similar to that of mean grain size, the positive correlation between precipitation and mean grain size was not significant, indicating that grain size coarsening had no correlation with the current precipitation, either (Figure 12).

4. Discussion

To minimize the influence of atmospheric path radiation, we adopted an improved atmospheric correction algorithm [33] rather than FLAASH in the case of the existence of complex aerosol types at land–sea junctions. MODIS Terra data described the specific aerosol types in tidal flats due to the sensitivity and designed spectral bands. The adopted atmospheric correction algorithm for TM/ETM+ images performed well with the aid of MODIS Terra data in this study (Figure 4).

Considering the low correlations between the measured reflectance and grain-size parameters, another advanced calibration algorithm, the BPNN, was tested here for comparison with SVM. The BPNN can learn and store the mapping relationship of a large number of input–output modes [45]. The BPNN could eliminate the mathematical formula for describing and revealing the required mapping relationship beforehand. The learning method of BPNN adopted the steepest descent method to continuously adjust the weight and threshold of the network through reverse transmission so that the square sum of errors in the network was minimized [46]. Additionally, we adopted R, SI, DI, PI, and RI as input variables. The nodal number of hidden layer neurons was set between 2 and 20 based on Zhang’s method [47]. The output layer transfer function of the single hidden layer model was placed at the Purelin function. The hidden layer transfer function was set as a tansig tangent s-type function [48]. The training functions were trained functions based on the gradient descent algorithm [49]. The maximum number of iterations, learning rate, and training accuracy were set to 1000, 0.05, and 1 × 10⁻⁵, respectively. The remaining parameters were set as default values. The optimal BPNN calibration models for grain-size parameter retrieval were obtained (Figure 13,

Table 4. Optimal node number combinations and calibration accuracies of BPNN model.

Grain-Size Parameters	Hidden Layer Number	Node Number		R2	RMSE
		1st	2nd
clay content	2	6	11	0.67	1.95
silt content	2	16	16	0.78	101.27
sand content	2	11	16	0.79	120.53
mean grain-size	2	17	18	0.77	0.13

). The SVM performed better than the BPNN based on a comparison of R² and RMSE. In addition, there were some inherent issues in the BPNN, such as how to determine local minimum points, network structure problems, overlearning, and underlearning [50,51]. The SVM transformed the problem to be solved as a problem of quadratic programming optimization. In theory, the global optimal solution was guaranteed, and the local solution and overlearning in the neural network training process were solved [52,53]. Thus, the SVM was finally adopted here in the grain-size parameter retrieval with multitemporal TM/ETM+ images over 26 years. Compared with multispectral satellite images, airborne images, such as simulated Daedalus 1268 Airborne Thematic Mapper (ATM) data, Compact Airborne Spectrographic Imager (CASI) data, and high-spatial-resolution satellite data (IKONOS), have been used in mapping grain-size parameters. Recombining the subpixel end member abundance through multivariate regression analysis significantly improved the image calibration for both sediment clay and sand content (R² > 0.8) in the ATM images [8]. Two robust adjustment techniques (MVE and MCD multivariate M-estimators of location and scale) provided acceptable algorithms for grain-size mapping with CASI-2 data in Santander Bay, Spain [9]. Furthermore, calibration algorithms provided an estimated R² of 0.91 for mud flat facies and 1 for sand flat facies with IKONOS data for the Hwangdo tidal flat, Cheonsu Bay, Korea. The calibration algorithms provided an estimated R² of 0.93 for sandy loam, 0.94 for silty loam, and 0.67 for clay loam [54]. The calibration accuracies in this paper were slightly below those relying on airborne images in the aforementioned literature. However, the estimated accuracies were acceptable, with R² values of 0.66 for clay, 0.81 for silt, 0.83 for sand, and 0.80 for mean grain size. In addition, the accessibility and cost-effectiveness of multispectral satellite images enabled the grain-size parameters to be periodically mapped with wide coverage. Notably, the effect of water on the reflectance was not considered in this paper. The influence was difficult to differentiate and eliminate from the reflectance. The reflectance of dried soil samples in the laboratory might develop an actual correlation between reflectance and grain-size parameters. However, the laboratory-acquired correlation was difficult to use to map grain-size parameters in the field. From another point of view, the water content was influenced by grain-size parameters. To some extent, there was an inverse correlation between grain size and water holding capacity [55,56]. The smaller the grain size of the soil is, the greater its water holding capacity. The effect of water content on the reflectance was inherent; however, it was affected by grain-size parameters, especially under moisture equilibrium conditions for the same sediment. A recent study had illustrated that the moisture of the same sand (a median diameter of about 250 μm) in a 1 m resolution laser image could differ from almost 0% (dry) to over 20% (fully saturated), with a strong effect on surface reflectance of a near-infrared band (1550 nm) [57]. However, the effect might be reduced here as the measured mean grain-size of water-holding sediment was less than 6.05 φ (from 2 to 66.26 μm) and the large pixel size (30 m) of the satellite imagery used here. We preferred to address this issue in further studies. Additionally, given the increasing availability of high-temporal- (almost daily) and spatial-resolution (3 m) and high-radiometric-resolution (12-bit) images from CubeSats such as PlanetScope imagery [58,59,60], the application of this imagery in mapping grain size based on spectral and textural features would be an alternative for further investigation.

The yielded results of the study showed that the area of tidal flats has been greatly reduced under the interaction of humans and nature in recent decades (Figure 8, Figure 9, Figure 10 and Figure 11). In addition, the remotely sensed coarsening trend of grain size indicated that erosion had overtaken siltation. The coarsening trend might be aggravated by the overutilization. More reasonable planning for the utilization of tidal flats is needed.

5. Conclusions

An experiment aimed at quantitatively mapping the sediment grain-size parameter distribution in the intertidal zone of the southwestern coast of Laizhou Bay is presented in this paper. A total of 238 sets of intertidal zone surface sediment samples and their reflectance spectra were measured during nearly 30 field surveys from 2 March 2013 to 19 July 2014. Then, in situ reflectance measurements were matched with the Landsat band reflectances by using the corresponding spectral response function. The generated Landsat band reflectances (b1, b3, b4, and b5) were significantly correlated with the clay, silt, sand, and mean grain size contents. However, the correlations between the original band reflectances and the sediment grain-size parameters were not high enough to develop a stable regression model, as were the correlations between several band indices (Table 2) and the sediment grain-size parameters. Then, two nonlinear fitting methods (SVM and BPNN) were compared in the sediment grain-size parameter calibration due to their efficient nonlinear fitting capability when the linear fitting did not work. However, there were still several issues to be solved in the BPNN, such as poor generalization ability and overfitting. Thanks to the SVM, the training error was minimized, and the generalization ability was maximized by setting an RBF. Moreover, in contrast with R² and RMSE, the SVM performed better than the BPNN did.

The spatial-temporal distributions of the clay, silt, and sand contents and mean grain size over 26 years from 1989 to 2015 were finally obtained with multitemporal Landsat TM/ETM+ images based on the validated SVM model. The yielded results showed that there was an obvious decreasing trend in the contents of clay and silt on the southwestern coast of Laizhou Bay. However, there was an increasing trend in sand content and mean grain size here. The coarsening trend agrees with the conclusion obtained through in situ measurements in a former study. Although the reduction in Yellow River discharge and the increase in precipitation were not related to the coarsening of grain size significantly, the coarsening trend and the reduction in tidal flat area might be aggravated by the overutilization. More reasonable planning is required.