Using Satellites to Monitor Soil Texture in Typical Black Soil Areas and Assess Its Impact on Crop Growth

Abstract

Soil texture is an important physical property of soil. Understanding the spatial distribution of cultivated soil texture in black soil areas is crucial for precise agricultural management and cultivated land protection in these zones. This study utilizes the random forest algorithm, Landsat-8 satellite remote sensing data, and climate- and terrain-related environmental covariates to map the spatial distribution of soil texture and analyze its impact on crop growth. The results show that (1) the order of prediction accuracy differs for different soil texture types; April is determined to have the highest prediction accuracy for silt and sand, while May exhibits the greatest accuracy for clay. (2) Adding environmental variables can significantly improve the accuracy of soil texture predictions; the root mean square error (RMSE) has decreased to varying degrees (silt: 0.84; clay: 0.04; sand: 0.85). (3) Soybean growth has the strongest response to soil texture; clay grain is the key factor affecting crop growth in drought scenarios, and sand grain is the dominant factor influencing flooding. This study improves the accuracy of the remote sensing mapping of soil texture through the combination of remote sensing images and environmental variables and quantitatively evaluates the impact of soil texture on crop growth.

1. Introduction

Black soil is generally in a high base state, with a relatively high cation exchange capacity and base saturation [1]. Additionally, the black soil region’s soil is highly fertile, making it ideal for agricultural activities. Since its reclamation in the 1950s, the black soil zone in northeastern China has emerged as a critical contributor to both domestic and global grain production. Its annual grain yield represents approximately 20% of China’s total output [2]. Affected by human activities and the East Asian monsoon climate, the vegetation covering the area has been destroyed on a large scale, making black soil more susceptible to erosion. This seriously threatens agricultural production and food security in black soil areas [3]. Consequently, tracking the physical and chemical characteristics of farmland soil in black soil regions is crucial for ensuring national food security. Soil texture, a key factor influencing soil’s physical characteristics, plays a critical role in regulating numerous physical, chemical, and biological processes within the soil environment. It significantly impacts properties such as thermal capacity, hydraulic conductivity, moisture retention, and the transport of solutes. Furthermore, soil texture is integral to various applications, including climate modeling, ecological studies, hydrological simulations, precision agriculture, and the management of soil contamination. The spatial variability of soil texture, both horizontally and vertically, is often pronounced across different regions and landscapes. In recent years, there has been a growing need for precise soil texture data to address pressing global and national challenges, including climate change, soil erosion, water resource management, environmental pollution, agricultural productivity, and the sustainability of ecosystems [4].

The physical characteristics of soil, particularly its texture, play a significant role in determining pore structure and porosity, which in turn influence the soil’s capacity to retain water and facilitate air circulation. Moreover, soil texture is closely associated with its ability to store nutrients and moisture, making it a critical factor in agricultural practices such as irrigation management, nutrient availability, and root development. In recent years, the expansion of agricultural land has led to a decline in natural vegetation in black soil regions. This has resulted in an increase in sandy and loamy soils, as well as areas rich in organic matter, while the extent of clay-rich soils has reportedly diminished [5]. Consequently, analyzing the spatial variability of soil texture in cultivated black soil regions is essential for implementing precision agriculture and ensuring the sustainable protection of farmland. Soil texture, along with other physical properties, interacts with fertility and environmental factors to influence crop growth and yield. The concentration of water-absorbing polymers (WAPs) in the soil is closely linked to its texture, which in turn affects key parameters such as available water content, the duration until wilting occurs, and the frequency of irrigation required. These factors collectively impact the overall growth and health of crops [6]. In agricultural practices, the appropriate selection of soil texture is critical for enhancing both the productivity and quality of farmland. Different crops exhibit varying levels of adaptability to soil textures; for instance, certain crops thrive in sandy soils, whereas others perform better in clay-rich soils. Consequently, understanding the interplay between soil texture and crop growth is essential for optimizing management strategies related to planting, fertilization, and irrigation.

Geostatistical analysis is a widely utilized technique for predicting and visualizing the spatial distribution of soil texture. Grounded in the principles of regionalized variable theory and variogram modeling, geostatistics examines natural phenomena characterized by spatial interdependence, combining both random variability and structured spatial patterns [7]. However, the accuracy of such predictions is often constrained by the density and variability of sampling points, necessitating a large number of closely spaced samples to generate reliable spatial distribution maps of soil texture. In recent years, remote sensing technology has gained popularity for estimating key soil properties, owing to its advantages of being non-invasive, cost-efficient, time-saving, and capable of providing direct measurements [8]. Multispectral and hyperspectral satellite images have been proven effective in predicting soil texture [9]. This lays the foundation for the development of the remote sensing prediction of soil texture. Typical remote-sensing approaches analyze the changes in the soil spectral reflectance of remote-sensing images and ground-based data at sampling points, establish a soil texture prediction model, and realize the spatial estimation and visualization of its distribution.

Currently, numerous research methods and prediction models have been established to investigate soil texture. Bahrami et al. [10] developed a partial least squares regression (PLSR) approach for the prediction of soil texture. Yudina et al. [11] performed soil texture classification based on the particle size distribution (PSD) using the laser diffraction method. Scholars have also employed digital images captured from smartphones and applications to predict soil texture [12]. Chagas et al. [13] employed the RF and multiple linear regression models to predict the spatial distribution of soil texture in semi-arid areas, revealing the RF model to exhibit a higher accuracy. The RF model is reported to have a higher inversion accuracy compared to the regression tree model [14].

Individual image features may reflect abnormal conditions in some areas, resulting in the abnormal reflectivity of ground objects and reducing the stability of the soil texture inversion model and accuracy [15]. Consequently, current studies on soil texture prediction using remote sensing typically utilize spectral indices derived from remote sensing imagery as input variables. These studies aim to predict soil texture and investigate how these indices influence the accuracy of the predictions [16]. However, different remote sensing images obtained in the same study area can produce differences in soil texture inversion accuracy. This can be attributed to the influence of factors such as precipitation, straw coverage, surface morphology, etc. on the image. Therefore, incorporating environmental covariates can better supplement soil information, thereby improving the accuracy of soil texture prediction.

Most of the studies related to the effect of soil texture on crop growth are small-scale or laboratory research experiments, with limited guidance for large-scale agricultural cultivation. The effect of different soil texture indices (silt, clay, and sand) on the growth and development of peanuts was investigated using the carton planting method [17]. Wang et al. [18] tested the effect of these three soil textures on cotton growth and development under combined drainage and irrigation and conventional drip irrigation via test pit trials.

To tackle the aforementioned challenges, this research focused on Youyi Farm, located in a representative black soil region, as the study area. Leveraging the random forest algorithm, reflectance data from Landsat-8 imagery across various months were used as input variables to identify the most suitable month for predicting soil texture. Subsequently, additional environmental covariates related to climate and terrain were incorporated to assess changes in prediction accuracy and evaluate the significance of input variables in the soil texture prediction process. This facilitated the spatial estimation and mapping of soil texture. Furthermore, the predicted soil texture results in the study area were correlated with the normalized difference vegetation index (NDVI) during the growing season to examine the relationship between soil texture and crop growth. The study aimed to investigate the spatial distribution of soil texture and its effects on the growth of various crops.

2. Materials and Methods

This study first predicted soil texture using April–June remote sensing imagery, selecting the optimal images based on prediction accuracy. The best-performing imagery was integrated with environmental covariates to develop an enhanced prediction model for mapping soil texture spatial distribution. Finally, correlations between the soil texture map and multi-year crop growth patterns were analyzed to assess soil texture’s impact on crop growth. The technical roadmap of this study is shown in Figure 1.

2.1. Study Area

The study area is located in Youyi Farm in the Sanjiang Plain in northeast China (Figure 2). As the first state-owned farm in China, its agricultural infrastructure is robust and comprehensively promotes the standardization of agricultural planting and breeding, the mechanization of operations at an appropriate scale, the intensification of production, the socialization of services, and the marketing of brands. Moreover, the farm contributes significantly to fostering the sustainable and high-quality advancement of modern agricultural practices. It is regarded as one of the nation’s key commercial grain production centers, particularly in the realm of modern agriculture. Youyi Farm (131°27′–132°15′ E, 46°28′–46°59′ N) is located in Shuangyashan City, in the hinterland of the Heilongjiang, Songhua River, and Wusuli River plains. The terrain in the interior is relatively flat, the southwest zone is hilly, and the northeast is low-lying, sloping from the southwest to the northeast. Some areas of Youyi Farm suffer from severe water and soil erosion, excessive fertilization usage, and a vicious cycle of water and fertilizer loss. These problems have resulted in the serious degradation of soil and farmland quality. The study area encompasses 1800 km² of cultivated land, characterized by fertile soil and a strong physical structure. Predominant soil types include meadow, black, and swamp soils. The region experiences a mid-temperate continental monsoon climate, influenced by the southeast monsoon during summer and polar cold air masses in winter. The average annual temperature is 2.5 °C, with an average annual precipitation of 500 mm. These climatic and soil conditions provide an ideal environment for cultivating a variety of crops. The agricultural season spans from April to October, with primary crops being rice, soybeans, corn, and wheat.

2.2. Data Acquisition and Processing

Using ArcGIS software (version 10.6), we first standardized the coordinate system of all datasets (remote sensing imagery, environmental covariate rasters, and sampling point data) to “WGS 1984 UTM Zone 52N”. All data were then resampled to a uniform 30 m spatial resolution. Subsequently, the “Extract Multi Values to Points” tool was applied to extract pixel values from both remote sensing images and environmental covariates to corresponding sampling points, which served as model inputs.

2.2.1. Sample Point Data Acquisition

This research was gathered during field surveys conducted between 22 March and 5 April 2021. During this period, Youyi Farm was in the bare soil phase, with the ground fully exposed and free from crop residues or snow cover. GPS positioning was conducted within the study area. The sample points were chosen from cultivated soil with an open view, relatively flat terrain, and uniform geological conditions. Additionally, information such as terrain characteristics and different soil types were considered. In total, 188 field-measured samples were obtained, and their coordinates, terrain, and other relevant information were recorded. The distribution of soil samples is illustrated in Figure 2. After collection, the samples were placed in cloth soil bags and transported to the laboratory for air drying to eliminate stones, weed roots, and other impurities. The soil was weighed indoors, air-dried, and sieved through a 2 mm mesh. The particle size distribution (mass fraction) of the samples was measured using a Malvern MS-A 2000 laser particle size analyzer (Malvern Instruments Co., Ltd., Worcestershire, UK), which has a residual error of less than 1%.

Figure 3 presents the descriptive statistics related to soil texture. The proportions of silt, clay, and sand were 18.31–81.35%, 2.03–19.89%, and 1.02–79.66%, with average proportions of 66.10%, 13.75%, and 20.15%, respectively (Figure 3). The coefficient of variation (CV) analysis revealed that the sand fraction exhibited the highest variability (CV = 81.89%), whereas the silt fraction showed the lowest (CV = 20.83%). The clay fraction had a CV of 22.47%. Based on the CV values, variability can be categorized into three levels: low (CV < 15%), moderate (15% < CV < 35%), and high (CV > 35%). Regions with high soil variability (elevated CV values) may offer ideal conditions for calibrating predictive models [19]. As illustrated in Figure 3, the statistical analysis indicates that the mean and median values for silt particles (mean = 66.10%, median = 71.05%), clay particles (mean = 13.75%, median = 14.55%), and sand particles (mean = 20.15%, median = 15.33%) show minimal variation. This suggests that each soil texture fraction (STF) follows a distribution that is close to normal [20].

2.2.2. Image Acquisition and Treatment

In recent years, a growing number of Earth observation platforms have emerged, offering cloud-based processing and online access to analytics-ready data. Notably, Google Earth Engine (GEE; https://earthengine.google.com/; accessed on 10 Marth 2024) enables users to retrieve data from satellite imagery and other Earth observation databases, while also providing robust computational resources to analyze these datasets [21].

The Landsat-8 satellite, launched and managed by the National Aeronautics and Space Administration (NASA) in 2013, provides imagery for the study area starting from 2014. The images have a spatial resolution of 30 m and a revisit cycle of 16 days. We obtained Landsat-8 imagery using GEE from April to June 2014–2022 (bare soil period). Following cloud and shadow masking, the images were synthesized into median composite images according to the month. Three multi-year composite images for the months of April to June were generated using data from 2014 to 2022. Seven spectral bands—blue, green, red, near-infrared, and two short-wave infrared bands—were utilized from the imagery to predict soil texture.

Sentinel-2 is a high-resolution multispectral imaging satellite equipped with 13 spectral bands. It offers ground resolutions of 10 m, 20 m, and 60 m, along with a revisit cycle of 5 days. We obtained Sentinel-2 imagery from GEE from 2019 to 2021 (31 August 2019; 23 July 2020; and 17 August 2021). The Sentinel-2 imagery was used to extract NDVI.

2.2.3. Environmental Covariate Acquisition and Treatment

To accurately predict soil texture in the study area, supplementary data sources beyond remote sensing imagery were also incorporated, we also used four environmental covariates related to climate and terrain [22,23,24], with a spatial resolution of 30 m (

Table 1. Environmental covariates and remote sensing data.

Class	Attribute	Description	Unit
Climate	PRE	Precipitation	mm
	AT	Average Temperature	°
Terrain	DEM	Digital Elevation Model	m
	SL	Slope	Degree
Remote Sensing	Band1	Coastal/Aerosol	Reflectance factor
	Band2	Blue	Reflectance factor
	Band3	Green	Reflectance factor
	Band4	Red	Reflectance factor
	Band5	Near Infrared	Reflectance factor
	Band6	Short Wave Infrared−1	Reflectance factor
	Band7	Short Wave Infrared−2	Reflectance factor

).

• Multiyear average precipitation (PRE): Precipitation plays a direct role in determining soil moisture levels, which subsequently impacts the transformation, movement, leaching, and accumulation of minerals, organic matter, and their byproducts within the soil [25]. This research utilizes the “total_precipitation” variable from the “ERA5 Monthly Aggregates” dataset available in Google Earth Engine (GEE) to compute the monthly average precipitation for the study area from 1979 to 2020, which serves as the multiyear average precipitation index.
• Multiyear average temperature (AT): Air temperature influences soil temperature and a wide range of physical, chemical, and biological processes. Moreover, variations in temperature lead to changes in soil properties and behavior [26]. In this research, the “mean_2m_air_temperature” variable from the “ERA5 Monthly Aggregates” dataset available in Google Earth Engine (GEE) is employed to compute the monthly average temperature for the study area from 1979 to 2020, which serves as the multiyear average temperature index.
• Elevation (DEM): Elevation indirectly influences soil properties by driving the redistribution of matter and energy in mountainous regions. As temperature, precipitation, and humidity vary with altitude, distinct climate and vegetation zones emerge, leading to pronounced vertical stratification in soil composition and physicochemical characteristics [27]. In this study, the “Elevation” dataset from the “NASADEM: NASA NASADEM Digital Elevation 30 m” collection, accessible through Google Earth Engine (GEE), is employed as the elevation input for the study area.
• Slope (SL): In steep terrains, limited infiltration and intense erosion hinder the accumulation of soil organic matter. Conversely, flat areas with slow water flow, reduced hydraulic erosion, and minimal loss of topsoil and nutrients provide more favorable conditions for organic matter accumulation [28]. The research employs the “Elevation” dataset from the “NASADEM: NASA” collection available in Google Earth Engine (GEE) to derive the “Slope” parameter, which serves as the slope input for the study area.

In addition to the inherent suitability of environmental covariates for soil texture prediction, they can also provide complementary information to remote sensing data, thereby improving prediction accuracy. While remote sensing data primarily capture instantaneous surface information, climate and topographic variables offer long-term, stable environmental context. This helps compensate for the limitations of remote sensing in characterizing deep soil properties or historical pedogenic processes. Specifically: temperature and precipitation reflect the climatic drivers of soil formation. Elevation and slope characterize spatial heterogeneity in soil properties.

2.3. Prediction Model

The random forest (RF) algorithm is known for its high prediction accuracy, robustness to outliers and missing values, and ability to handle high-dimensional data with multicollinearity. It requires minimal manual intervention and is resistant to overfitting. Previous research has demonstrated the effectiveness of the RF model in spatially predicting soil texture, with studies showing superior accuracy compared to other models, such as regression trees [14]. The RF modeling process is carried out using the random forest package in R Studio software (version 4.2.1). During execution, the package ranks the importance of independent variables influencing soil texture. The larger the value, the greater the impact of the independent variable on soil texture and the stronger the correlation. This study selects the number of optimal regression trees (ntree) and number of split nodes (mtry) in the RF model by observing the out-of-bag error, thus establishing the optimal RF prediction model. Through multiple training iterations, the optimal parameters were determined as follows: “ntree” was set to 1200, “mtry” to one-third of the number of input variables, and “nodesize” to 8. This configuration resulted in a relatively stable model performance. The seven-band information of monthly multi-year composite images of Landsat-8 in different months and four environmental covariates are used as input quantities for the prediction modeling of soil texture, respectively. Ten-fold cross-validation is employed to assess model performance. The 188 samples are randomly split into ten groups, with nine groups serving as training data and one group as validation data. A fixed seed value of 1000 ensures consistent division of training and test datasets for texture mapping, facilitating easier comparison across different models. Model stability is evaluated using the coefficient of determination (R²), while accuracy is measured by the root mean square error (RMSE) [29].

$$R^2=1-{\displaystyle \frac{{\sum }_{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{y}}_i-{\widehat{y}}_i\right)}^2}{{\sum }_{i=1}^n{\left(y_i-\overline{y}\right)}^2}}$$

(1)

$$RMSE=\sqrt{{\displaystyle \frac{1}{\mathrm{n}}}{\displaystyle \sum _{i=1}^n{\left(y_i-{\widehat{y}}_i\right)}^2}}$$

(2)

where n represents the number of samples; y_i denotes the observed value at the i-th sampling point; and ŷ_i refers to the predicted value at the i-th sampling point. The R² value typically ranges between 0 and 1, with values closer to 1 indicating a stronger alignment between predicted and actual measured values.

2.4. Extraction of Crop Growth Status

Precipitation data from April to October, which significantly influences crop growth and development, is extracted from the monthly average precipitation records of the study area over the past decade (2013–2023). These records are sourced from NCEI NOAA (https://www.ncei.noaa.gov/data/global-summary-of-the-day/archive/; accessed on 10 March 2024) and are used to assess the annual conditions of the study area, classifying them as flood, normal, or drought. Images during the growth period of different years are selected to extract NDVI as an indicator for the crop growth status under different years. After visual interpretation, we obtain the planting structure of the study area under different years. ArcGIS 10.6 is employed to perform mask extraction on the crop growth status to obtain the growth status of different crops in different years.

2.5. Correlation Analysis

Crop growth responses to soil texture may vary with time and the vegetation type [30]. Therefore, to clarify the corresponding relationship between crop growth and soil texture, we compare the correlation between different soil texture components and different crops in different years. Following the requirements of the Chinese Farmland Protection Department for the layout of sample points in the cultivated land quality monitoring area and the practical factors of Youyi Farm, a fishing net was established with dimensions 500 m × 500 m. After removing the null values, a total of 4715 sample points were obtained. The sample points were divided according to the crop zones planted in different years to analyze the relationship between NDVI and the prediction results of silt, clay, and sand in different years. We considered correlations with p < 0.01 as significant.

$${\mathrm{r}}_{xy}={\displaystyle \frac{{\displaystyle {\sum }_{i=1}^n\left(x_i-\overline{x}\right)\left(y_i-\overline{y}\right)}}{\sqrt{{\displaystyle {\sum }_{i=1}^n{\left(x_i-\overline{x}\right)}^2{\displaystyle {\sum }_{i=1}^n{\left(y_i-\overline{y}\right)}^2}}}}}$$

(3)

where r is Pearson’s correlation coefficient, with a range of [−1, 1]. A negative value signifies an inverse relationship, while a positive value indicates a direct relationship, and 0 indicates no correlation. The closer r is to 0, the weaker the correlation; the closer it is to −1 or 1, the stronger the correlation. $X_i$ and $Y_i$ are the observed values of two variables respectively. $\overline{X}$ and $\overline{Y}$ are their averages. The numerator represents the covariance, which evaluates whether the two variables exhibit similar trends in variation. The denominator is the product of the standard deviations of the two variables, serving to normalize the result.

2.6. Analysis of the Influence of Soil Moisture Content on the Prediction Accuracy of Soil Texture

To analyze the influence exerted by soil moisture content on soil texture prediction, the land surface water index (LSWI) is utilized to represent soil moisture content. The LSWI is calculated as follows:

$$LSWI={\displaystyle \frac{{\rho }_{\mathrm{nir}}-{\rho }_{\mathrm{swir}1}}{{\rho }_{\mathrm{nir}}+{\rho }_{\mathrm{swir}1}}}$$

(4)

where ${\rho }_{\mathrm{nir}}$ is the surface reflectance of the Landsat-8 near-infrared band (B8), and ${\rho }_{\mathrm{swir}1}$ is the surface reflectance of the shortwave infrared band 1 (B11).

The Land Surface Water Index (LSWI) incorporates reflectance data from shortwave infrared bands, which are highly responsive to both soil and vegetation moisture. This index is widely used to track changes in soil moisture levels, with higher LSWI values indicating increased moisture content [31]. The root mean square error (RMSE) of the Land Surface Water Index (LSWI) and soil texture prediction process can highlight the impact of soil moisture content on the accuracy of soil texture predictions.

3. Results

3.1. Analysis of the Soil Reflection Spectral Characteristics

Soil spectral reflectance is strongly correlated with the particle size distribution of topsoil, a key component of soil texture. Figure 4 depicts the reflection spectrum characteristics of soil particle composition across the same period. The Landsat-8 spectral curve shows similar trends across different particle sizes, but notable variations in spectral reflectance are evident. Reflectance increases with wavelength from the visible light to the shortwave infrared 1 bands (1565–1655 nm), whereas it decreases from the shortwave infrared 1 to the shortwave infrared 2 bands (2100–2280 nm). Additionally, spectral reflectance declines as the mass fraction of silt and clay particles rises, while it increases with a higher mass fraction of sand particles. This is due to the mixing effect of different-sized particles [32]. In addition, the proportion of minerals also has an impact on the spectral reflectance. Soil reflectivity is reduced by the layered silicates contained in clay particles, while the large amounts of quartz contained in sand particles enhance it [33].

3.2. Optimal Temporal Window for Soil Texture Prediction

Table 2. Comparison of soil texture prediction accuracy in different periods.

Month	Silt		Clay		Sand
	R2	RMSE (%)	R2	RMSE (%)	R2	RMSE (%)
April	0.73	2.55	0.60	0.74	0.73	3.09
May	0.62	3.01	0.65	0.70	0.67	3.45
June	0.51	3.43	0.54	0.79	0.50	4.23

reports the changes in the soil texture prediction accuracy of Landsat-8 multi-year monthly synthetic images in a single month. The image in April exhibits the highest silt and sand prediction accuracy (RMSE of 2.55% for silt, RMSE of 3.09% for sand), while the image in June has the lowest (RMSE of 3.43% for silt, RMSE of 4.23% for sand). For the clay predictions, the image in May is the most accurate (RMSE of 0.7%), while the image in June is the least (RMSE of 0.79%).

3.3. Impact of Environmental Covariates on Soil Texture Prediction Accuracy

After evaluating the prediction accuracy of bare soil period imagery, climate, topography, and hydrology covariates were integrated with the remote sensing images showing the highest accuracy in assessing soil texture prediction performance (

Table 3. Comparison of soil texture prediction accuracy in different periods after adding environmental covariates.

Combination	Silt		Clay		Sand
	R2	RMSE(%)	R2	RMSE(%)	R2	RMSE(%)
Remote Sensing	0.73	2.55	0.65	0.70	0.73	3.09
Climate	0.87	1.78	0.68	0.67	0.85	2.28
Terrain	0.83	2.04	0.69	0.66	0.82	2.52
RS + Climate + Terrain	0.88	1.71	0.68	0.66	0.86	2.24

). The inclusion of environmental covariates enhances the accuracy of soil texture predictions. The image in April has the highest prediction accuracy for silt and sand particles after adding climate and terrain data (RMSE of 1.71% for silt, RMSE of 2.24% for sand), while the image in May has the highest prediction accuracy for clay particles after adding terrain data (RMSE of 0.66%). Figure 5 shows the performance of modeling approaches.

3.4. Spatial Distribution Map of Soil Texture

The April imagery, integrated with climate and terrain data, was employed to predict the spatial distribution of silt and sand particles. In contrast, the May imagery, combined solely with terrain data, was used to predict the spatial distribution of clay particles. As the selected sampling points were all cultivated soil samples, the selected images were cropped within the boundary range of the cultivated land in Youyi Farm. Figure 6 illustrates the spatial distribution of predicted soil texture across the study area. The silt, clay, and sand contents range from 29.8 to 77.7%, 4.4 to 16.0%, and 7.3 to 65.8%, respectively. The findings reveal that silt and clay particles are predominantly concentrated in the northeast, north, and south regions of Youyi Farm, while their presence is relatively lower in the central and southwest areas. Sand particles exhibit the opposite trend, particularly in the central part of Youyi Farm, where the sand content is generally higher.

3.5. Analysis of the Impact of Soil Texture on Crop Growth

The cumulative precipitation from April to October during crop growth and development was extracted based on the monthly precipitation in Shuangyashan City from 2013 to 2023 (Figure 7). The highest level of precipitation was observed in 2019 (846 mm) and the lowest in 2021 (397 mm). The three-year precipitation dataset from 2019 to 2021 indicates that 2019 was a flood year, 2020 was a normal year, and 2021 was a drought year.

The NDVI was extracted using the remote sensing images captured during the growing seasons (July or August) of different years. The NDVI values were then clipped based on the crop classification for the years from 2019 to 2021, revealing distinct crop growth stages across different years (Figure 8). Crop growth remains relatively stable during flood years but declines significantly in drought years compared to normal years.

Correlation analysis of soil texture components and crop growth was performed for different years and different crops (

Table 4. Correlation between soil texture and crop growth in different crop zones in different years.

Year	Crop	Silt	Clay	Sand
2019	soya	−0.607	−0.505	0.579
	rice	0.361	−0.304	−0.354
	corn	−0.455	−0.434	0.454
2020	soya	−0.647	−0.525	0.613
	rice	0.337	0.47	0.324
	corn	−0.548	−0.498	0.54
2021	soya	−0.394	−0.45	0.32
	rice	0.354	0.453	0.348
	corn	0.632	0.628	−0.632

). The findings indicate that soil texture shows strong consistency with the observed data from nine crop growth conditions. However, its effect on crops differs based on the year and crop type. Soybean crop growth has the strongest correlation with soil texture, showing a significant correlation in both normal and flood years. However, this correlation is reduced greatly in drought years. The correlation between sand grains, silt grains, and crop growth is not significant. Moreover, the correlation between rice crop growth and soil texture is relatively low, while in drought years, the correlation is more significant. Corn crop growth is significantly correlated with different soil textures in different years.

4. Discussion

4.1. Effect of Imaging Time on the Soil Texture Mapping Accuracy

The soil texture remote sensing inversion results reveal R² values of 0.73–0.88, 0.63–0.69, and 0.73–0.86 for silt, clay, and sand particles, respectively. This suggests that remotely sensed imagery available for soil texture inversion varies with the surface morphology and month. We determined the LSWI of three images from April to June [34]. The inversion process of silt, clay, and sand are affected by soil moisture content to varying degrees (Figure 9). In Figure 8a, the April and May images with higher soil texture prediction accuracy have lower water content. This indicates that the Landsat-8 images with higher soil texture prediction accuracy have lower soil moisture content. Moreover, the RMSE generally increases with the LSWI (Figure 9b). Soil moisture significantly influences spectral characteristics by altering reflectance, absorption features, temperature effects, and surface properties, further impacting the prediction accuracy of soil texture. Water exhibits stronger absorption in the shortwave infrared range, leading to a decrease in reflectance. Near 1450 nm and 1900 nm, water has distinct absorption bands, and an increase in soil moisture enhances these absorption features. Higher moisture content increases the soil’s heat capacity, resulting in lower daytime temperatures and higher nighttime temperatures, which affects the radiative properties in the thermal infrared band. Moisture also changes the surface roughness of the soil, influencing the scattering and reflection of light. Therefore, soil moisture affects spectral reflectance to some extent, thereby influencing the prediction accuracy of soil texture.

In summary, during the optimal time window for soil texture prediction, soil moisture levels are relatively low. As soil moisture increases, the accuracy of remote sensing predictions for soil texture gradually declines. Therefore, soil moisture content is a critical factor influencing soil texture prediction. Soil texture is strongly correlated with soil reflectance in the visible near-infrared to shortwave infrared range (400–2500 nm). Within this spectral range, soil moisture significantly impacts the accuracy of soil texture retrieval through remote sensing. Swain et al. [35] utilized Sentinel-2 imagery from various collection dates to spatially predict soil texture. The study revealed that soil moisture content significantly influenced the accuracy of soil texture inversion. Similarly, subsequent research using Landsat-8 remote sensing images to estimate soil organic matter identified straw coverage and soil moisture content as key factors contributing to variations in prediction accuracy across different periods [36]. Furthermore, studies have shown that the variability of input data plays a critical role in determining the final accuracy of soil texture predictions [37]. The relatively higher prediction accuracy of sand can be attributed to the greater variability of sand content in the soil samples compared to that of silt and clay.

4.2. Role of Environmental Covariates in Soil Texture Prediction

This research incorporated four categories of environmental covariate data (multiyear average precipitation, multiyear average temperature, elevation, and slope). Numerous studies have already demonstrated that these four environmental covariates can well reflect the environmental conditions for soil development. Elevated temperatures typically accelerate the breakdown of soil organic matter and chemical weathering, leading to lighter soil textures and reduced fertility [38]. Higher precipitation levels can alter the particle size distribution within soil texture, accelerate soil erosion, and ultimately influence the formation and properties of soil texture [39]. Variations in elevation can affect soil drainage and erosion intensity, thereby influencing the formation and spatial distribution of soil texture [18]. Higher slopes can exacerbate soil erosion and water loss, thereby influencing the formation and spatial distribution of soil texture [39]. Including environmental covariates such as temperature, precipitation, DEM, and slope can provide additional information and explanatory power for the soil texture prediction model, thereby improving its accuracy and reliability.

This study systematically analyzed the impact of environmental covariates on soil texture prediction models, revealing that incorporating environmental covariates can significantly enhance model accuracy. As demonstrated by the random forest variable importance analysis in Figure 10, different types of environmental covariates exhibit distinct influences on soil texture prediction. Specifically, environmental covariates play a particularly crucial role in predicting soil silt content, where the predictive importance of air temperature (AT) even surpasses all remote sensing indices. Similarly, for sand content prediction, both air temperature (AT) and precipitation (PRE) show significantly higher importance than most remote sensing features. Model validation results indicate that incorporating climate-related covariates improved the R² of silt and sand content prediction models by 0.14 and 0.12, respectively. Notably, while air temperature and precipitation show relatively lower importance in clay content prediction, they still contribute to improvement in prediction accuracy, suggesting that climate factors universally enhance predictions for all soil texture types. Although terrain-related covariates demonstrate less importance than climate factors in soil texture prediction, they complement the topographic information of the study area and further improve prediction accuracy. After incorporating terrain-related covariates, the prediction accuracy of silt, clay, and sand content all showed modest improvements. These findings provide important evidence for optimizing soil texture prediction models and highlight the necessity of including environmental covariates in multi-source data fusion modeling approaches.

4.3. Response of Soil Texture to Different Crop Growth Conditions

Table 4 reveals soybean crop growth and soil texture to exhibit the highest correlation among all soil texture and crop parameters, indicating that soybean responds most strongly to soil texture. Additionally, soybean growth shows a positive correlation with silt and clay particles while exhibiting a negative correlation with sand particles. This indicates that silt and clay particles promote the growth of soybean plants, while sand particles exert a negative impact on the growth of soybeans. In flood years, the promotion or inhibitory effect of soil texture on soybean crop growth remains unchanged, and the response of soybean growth to soil texture becomes weaker. In drought years, silt and clay particles promote the growth of soybean plants. Thus, sand particles have an inhibitory effect during these years, and the key soil-texture-related factors affecting the growth of soybean crops are clay particles. The response of corn crop growth to soil texture is similar to that of soybean, yet the degree of response is slightly lower. The correlation between rice crop growth and soil texture is low, and the response to soil texture is relatively unobvious. The key factor affecting rice in normal and drought years is clay. In flood years, the soil texture changes, and the key influencing factor is sand. Moreover, the response degree of rice to soil texture changes slightly with the year. This is because rice is submerged in water all year round, and, as a consequence, floods and droughts have less of an impact on the growth of rice crops. Therefore, compared with corn and soybean, rice yield is more stable [40]. Since the crop growth of soybeans responds most clearly to soil texture, Figure 11 presents a schematic diagram of the effect of soil texture on crop growth by taking soybeans as an example. The reason for the above phenomenon is as follows: in drought years, the water required for crop growth and development mainly comes from groundwater, and clay particles are more conducive to soil moisture retention, promoting crop growth. In contrast, during flood years, soil moisture primarily comes from rainfall. Clay particles hinder rainwater infiltration, leading to waterlogging, which is detrimental to crop growth. On the other hand, sandy soils allow for better drainage, reducing the impact of waterlogging on crop growth.

As shown in Table 4, there is a strong negative correlation between silt and sand. Thus, silt and sand often simultaneously affect crop growth. Silt grains and sand have a stronger impact on the growth of corn and soybeans, while clay grains mainly have a greater impact on the growth of rice. In flood years, the impact of sand on crop growth is enhanced, while in drought years, the impact of clay on the growth of all three crops increases. This is because during drought years, soil moisture decreases, and crop growth relies on shallow groundwater. Capillaries rise to provide water and transport nutrients [41] and clay has better water storage capacity. Therefore, in drought years, the role of clay in crop growth is improved to varying degrees.

4.4. Limitations and Future Research

Our results reveal that the uncertainty of soil texture prediction mapping using images from the crop growth period remains high. This study is based on a typical black soil area and is generally applicable to one-year ripening areas. Other ripening areas still require further research [26]. Soil texture is a complex variable, and the three distinct components jointly exert an impact on the process of crop growth and development. This study only examines the influence of individual soil texture components on crop growth vigor. In subsequent studies, the synchronous prediction method for different soil texture components will be explored and its impact on crop growth vigor will be analyzed. Existing research employs multi-task learning (MTL) combined with the SHAP model to simultaneously predict multiple soil characteristics, proving the feasibility of the simultaneous prediction of composite variables [42]. Sentinel-2 offers superior temporal and spatial resolution compared to the Landsat series satellites and has demonstrated significant potential for soil mapping in numerous laboratory simulations [43]. Moreover, the selection of environmental covariates is associated with several limitations. It is particularly important to select environmental variables that can effectively describe the production environment, farming intensity, and productivity in the study area [44]. In this study, due to limitations in the satellite data, remote sensing images of different months were selected for crop growth in different years. This may have an impact on the research results. Future research on soil texture mapping should prioritize multi-sensor fusion and address the influence of crop type, variety, and phenology on crop growth. Additionally, further studies are needed to explore the response of soil texture to drought and flood conditions, particularly the mechanisms by which different soil texture types affect crop growth and development across varying years. The conclusion of this study has not been applied in the field planting structure at present. Our subsequent research stage will clarify the impact of the proposed method on different groups, such as scholars and farmers.

To address these limitations, our future research will utilize higher spatial resolution Sentinel-2 satellite imagery combined with relevant environmental covariates, including climate, topography, and hydrological factors, as well as appropriate spectral indices such as SAVI, NDMI, and EVI as input variables for multi-task learning (MTL) prediction models to estimate soil texture. This integrated approach is expected to further improve the prediction accuracy of soil texture while simultaneously predicting silt, clay, and sand content. Through analytical methods, we will further investigate the effects of soil texture on crop growth to determine critical effect thresholds, thereby providing more precise guidance for smart agriculture and precision farming practices in the study area to ultimately enhance crop yields.

5. Conclusions

This research assessed the predictive performance of soil texture by integrating remote sensing imagery with environmental covariates. Subsequently, the response of soil texture to drought and flood conditions was analyzed. We observed the prediction time windows of silt, clay, and sand to be different. After adding the environmental variables, the prediction accuracy of soil texture significantly improved. Climate-related environmental covariates exhibited the most obvious effect on improving the prediction accuracy of soil texture. This effect was even more prominent when combined with terrain-related environmental covariates. Correlation analysis between soil texture and crop growth revealed that different crops respond differently to soil texture, and soil texture affects crop growth to varying degrees in different years. This research highlights that the accuracy of soil texture mapping depends on the selection of suitable remote sensing imagery and environmental variables. To mitigate the effects of drought and flood disasters on agricultural productivity, crop planting strategies should be optimized according to soil texture predictions and annual variations.