Evaluating the Efficacy of Sentinel-2B and Landsat-8 for Estimating and Mapping Wheat Straw Cover in Rice–Wheat Fields

Abstract

Sustainable agriculture and soil conservation methods are integral to ensuring food safety and mitigating environmental impacts worldwide. However, crop residue/straw serves many vital functions from tillage to harvest, so that quantifying the appropriate amount of Crop Straw Cover (CSC) on the soil surface is crucial for monitoring tillage intensity and crop yield performance. Thus, a novel research study is conducted to develop an innovative approach for accurately estimating and mapping the Wheat Straw Cover (WSC) percentage through two different multispectral satellites (Sentinel-2B MSI and Landsat-8 OLI-TIRS), using remote sensing-based techniques in Changshu County, China. The field measurements were collected from 80 distinct sites and eight images were acquired through both satellites for the analysis process by applying Crop Residue Indices (CRIs). The results indicate that the coefficients of determination (R²) of the Normalized Difference Tillage Index (NDTI) computed by Sentinel-2 and Landsat-8 were 0.80 and 0.70, respectively, and the root-mean-square deviation (RMSD) values were in the range from 6.88 to 12.04% for CRIs for both satellite data. Additionally, the comparative analysis of the developed model revealed that NDTI was R² = 0.85 and R² = 0.77, followed by STI, R² = 0.82 and R² = 0.80 and NDRI, R² = 0.69 and R² = 0.56 for Sentinel-2B and Landsat-8 data, respectively. Hence, the correlation strength of NDTI, STI and NDRI with WSC percentages was markedly superior by using Sentinel-2B spectral data compared to Landsat-8 ones. Moreover, the NDTI of Sentinel-2B data was the most accurate in mapping the WSC percentage in four categories, with an overall accuracy of 86.53% (κ = 0.78), surpassing the other CRI indices. Therefore, these findings suggest that the multispectral imagery of Sentinel-2B bolstered with enhanced temporal and spatial data was superior for precisely estimating and mapping the WSC percentage compared to Landsat-8 data over a large-scale agricultural region.

1. Introduction

The soil plays a crucial role in supporting human life by providing food, fiber, timber, storing and distributing water, filtering pollutants and storing carbon [1]. The amount of Crop Straw Cover (CSC), Fractional Vegetation Cover (FVC) and Bare Soil (BS) in croplands is temporally variable throughout the year depending on agricultural practices from tillage to harvest. Straw serves multiple purposes, such as protecting against erosion, loosening soil and increasing soil nutrients [2,3,4]. In agricultural production, retaining straw could positively influence water infiltration, evaporation and soil temperature [5] and increase soil fertility, improving soils’ water-holding capacity and soil quality [6]. Many scientists found that the percentage of residues left on the soil surface can have a significant impact on reducing the level of soil erosion within an agricultural field, while 15% of corn straw cover can reduce the effect of erosion by 75%, compared to bare soil [5].

In addition, obtaining a precise determination of Crop Residue Cover (CRC) is an essential metric that holds a significant value [7] for inclusion in soil erosion models, i.e., in order to determine the “C” factor in the Universal Soil Loss Equation [8,9]. Crop Straw Cover (CSC) holds a significant importance as an input parameter within models utilized for predicting the effects of agricultural systems on SOC levels, the emissions of greenhouse gases and crop productivity [10,11]. While agroecosystem models have been designed to assess the carbon sequestration associated with different soil and crop management conditions, such as the implementation of tillage practices that retain crop straw to prevent soil erosion [12,13], these models also monitor the carbon sequestration resulting from changes in field management methods. Hence, it is essential to estimate CSC on agricultural land at a regional scale to properly manage straw and incorporate retaining percentages for higher crop production and sustainable agriculture.

At present, the calculation of CSC relies on conventional survey methods, including visual estimation [14], point intercept [15], meter stick [16], spiked wheel [17] and photographic techniques [18]. Another method is the line transect method [19], famous for estimating the percentage of straw cover under in situ field conditions. In this, the number will directly measure the percent cover for that field area. Shelton and Jasa [19] recommended the line transect method as one of the easiest, comparable and most accurate methods for determining straw cover percentage. Regrettably, conventional techniques utilized to measure CSC across extensive areas are unsuitable, due to their cost, unsuitability for large spatial coverage and time-consuming nature [20]. Alternatively, some scientists used RS methods, such as RS-based image-processing techniques, to acquire CSC on various agricultural resources at a regional scale. The methods utilized to measure CSC based on RS can be divided into three categories: empirical regression [21,22] based on crop residue indices (CRIs); spectral angle methods [2,23]; and spectral unmixing [24]. The most commonly employed technique among these approaches is empirical regression, wherein a linear or non-linear relationship is established between CRC and the CRIs, also referred to as the crop residue indices technique. Thus, the RS-based technique offers a consistent approach to model the spatial variation in residue cover across extensive areas [25], unlike ground-based methods, which depend on interpolation and sampling.

However, residues and the soil exhibit spectral similarities and typically differ primarily in amplitude within the relatively broad spectral bands of most multispectral sensors [26]. Crop residue may be darker/brighter than the soil, depending on the type, age and moisture content [27,28]. In contrast, straw and the soil are visible in the NIR spectral range (350–2500 nm) with equivalent spectral properties. However, crop straw is primarily observed in the vicinity of 2100 nm, where cellulose and lignin absorption occurs [29]. Several agricultural residue indexes have been developed with the aim of enhancing the accuracy of estimating CSC, such as the Cellulose Absorption Index (CAI) [30,31]. Moreover, the Shortwave Infrared Normalized Difference Residue Index (SINDRI), which relies on ASTER bands 6 and 7, has a strong relationship with CRC [26]. Several broadband indices have been devised for mapping and estimating CRC by utilizing multispectral satellite data. The widely recognized ASTER CRIs encompass the Lignin Cellulose Absorption Index (LCA) and the Shortwave Infrared (SWIR), derived from SWIR bands [32], and these indices excel in quantifying CRC, by relying on their precise assessment of cellulose and lignin absorption characteristics [33]. The above-mentioned indices are designed to enhance the spectral signal differences between crop residues and soils for laboratory and field studies using a ground-based spectroradiometer and a hyperspectral satellite [34,35]. The hyperspectral indices linearly correlate with CRC and remain consistent across different soil types. Unfortunately, some hyperspectral satellite sensors offer a limited spatial and temporal coverage.

The Normalized Difference Tillage Index (NDTI) [36,37] has been found to be the most effective RS-based tillage index for determining residue cover. This is due to its ability to utilize the contrast in reflectance between the two Shortwave Infrared (SWIR) bands centered around 1600 and 2300 nm. However, previous studies have shown that there is a strong correlation between the Simple Tillage Index (STI) and the NDTI [37]. The mini NDTI is used to estimate CSC based on a time series of Landsat-7 (ETM+) data (R² = 0.89) [38]. Furthermore, these indices leverage the relative difference in broadband reflectance between soils and crop residues, so that they are suitable for estimating the crop residue percentage cover within an agricultural environment utilizing multispectral data across different residues from crops, including barley, corn, soy, sunflower, wheat and rice [39,40]. Most of the above indices primarily rely on normalized difference ratios derived from two bands. The formulation of these indices, e.g., simple ratio and normalized difference ratio, along with the selection of bands, affect the estimation performance [41].

Several satellites equipped with broadband multispectral sensors (Sentinel, Landsat Gaofen, ASTER and Worldview-3) provide a global coverage of agricultural lands [27]. However, the Landsat-8 with 30 m spatial resolution, compatible with CRIs, depends on spectral broad contrasts and the temporal resolution is 16 days, compared to Sentinel-2 (5 days), which has a 10 m spatial resolution [42].

The temporal resolution of satellite images can strongly influence crop straw percentage estimation, due to straw moisture content, straw age, soil condition, soil brightness and the existence of vegetation [33,43]. Moreover, there is not sufficient research to assess the capability of Sentinel-2B data in estimating crop straw cover [10,44]. The Sentinel-2B satellite was launched in 2017 and is equipped with a Multispectral Instrument (MSI) consisting of 13 spectral bands [45]. Notably, with the exception of the three red-edge bands, the spectral regions covered by the MSI bands are closely aligned with those of the Landsat-8 multispectral satellite. Moreover, Sentinel-2B imagery possesses higher spatial and temporal resolutions, which is more appropriate for agricultural monitoring [42]. Additionally, rice (Oryza sativa) and wheat (Triticum aestivum L.) are recognized to be essential food crops worldwide, representing approximately 32% of the total rice crop area and 42% of the wheat cropping regions. In Jiangsu Province, China, the combined cultivation areas of these crops represent about one-fifth of the nation’s total crop area [46,47]. However, the annual crop residues/straw produced from these food crops is estimated as 1430 million tons globally, while China ranks first in the production of wheat and rice residues, with an estimated total yearly production of 600–800 million tons.

In this context, previous research studies have left vast gaps. There is a need to clearly describe which multispectral satellite data are more accurate to estimate WSC values on the soil surface under various residue management strategies within a given agricultural region using RS- and GIS-based multispectral indices. Thus, there is a strong need to compare the efficacy of multispectral satellites through crop residue RS indices to develop a model, in order to precisely estimate the crop straw cover percentage on a regional scale. The specific objective of this study is to compare Sentinel 2B and Landsat-8 spectral indices with CRIs for accurately estimating Wheat Straw Cover (WSC) percentage and mapping it by using RS and GIS techniques. The expected results of this research provide a reference for and comparative analysis of two different satellite imagery data for accurately estimating and mapping WSC over a regional scale for long-term agricultural sustainability.

2. Materials and Methods

2.1. Study Area

The research was conducted in various villages inside Changshu County, located at coordinates 31.656° N and 120.753° E (Figure 1), and belonging to the Suzhou district in the Southern part of Jiangsu Province, China, in the Tai Lake plain. This county encompasses a total area of 1264 km², of which 386 km² comprises water bodies. The remaining land space is mainly used for agricultural purposes, with a total arable ploughable land area of 42,265 ha. Within this area, 27,451 ha are dedicated to paddy fields, 5600 ha are allocated for irrigation and 9214 ha are dry land. The area is located in a transitional region characterized by subtropical monsoon, humid climate with recorded yearly rainfall ranging from 1100 to 1200 mm and a temperature of 16 °C [48]. Moreover, the predominant cropping systems within the study area are rice (9.355 tons ha⁻¹) and wheat (5.03 tons ha⁻¹).

2.2. Ground Position Data

A survey was conducted in the study area during May and June 2019 by using a GNSS receiver to sense and record each sampling point’s position. The field data were collected from 80 sampling points that were randomly selected from various villages within the study area. The spatial distribution of these sampling points is shown in Figure 1. However, before the field visit, a questionnaire was prepared to collect the qualitative and the necessary information data. The questionnaire included coordinates (longitude, latitude and altitude), preceding crop, tillage practices, straw management methods, retained straw amount, residue type and height. Moreover, the ground survey included a complete inventory for measuring wheat straw cover and other field information.

2.3. Measurement of Straw Cover Percentage

The Wheat Straw Cover (WSC) percentage at the field sites was calculated using the line transect method [17] between 24 May and 6 June 2019, whereas each sample plot covered an area of approximately 1500 m² and exhibited a relatively uniform homogeneity. It is an easy and reliable method to determine crop straw/residue cover. A 50 ft (15.24 m) line-point transect was used to measure each test plot (Figure 2). This transect was divided into 100 evenly spaced markers with a 0.5 ft distance. The number of markers that intersected the crop residue was recorded and the percentage of the cover was subsequently determined by counting the number of markers that intersected or directly overlapped a piece of residue under in situ conditions [38]. The GNSS receiver was used to sense and record the precise position of each farm site, in order to produce a simple map of each study parameter and analyze the temporal changes in straw cover percentage through remote sensing techniques.

2.4. Remotely Sensed Satellite Data

For the estimation and temporal variability of straw cover percentage in large areas, remote sensing is the most reliable method using different indices. In this study, it was possible to acquire Landsat-8 (OLI-TIRS) and Sentinel-2B (MSI) multispectral satellite images of the study area, which were downloaded from different platforms.

Table 1. Characteristics of Landsat-8 (OLI-TIRS) and Sentinel 2B (MSI) satellite optical sensors.

Sensors	LANDSAT-8 (OLI-TIRS) Operational Land Imager and Thermal Infrared Sensors	SENTINEL-2B (MSI) Multispectral Instrument
Launch	11 February 2013	7 March 2017
Geographic reference system	Worldwide Reference System-2 (WRS-2)	Universal Transverse Mercator 1984
Orbit period/orbit length	16 days/705 km	10 days/768 km
Orbit inclination	98.22°	98.62°
Swath width	185 km	290 km

briefly describes the satellite sensors available for this study, i.e., OLI-TIRS on Landsat-8 and MSI on Sentinel-2B.

OLI-TIRS scenes were obtained from Landsat-8 through the USGS EarthExplorer platform, accessible at https://earthexplorer.usgs.gov/, (accessed on 30 August 2023) under WRS-2, Path 119, Rows 38 with UTM zone 51N. The Sentinel-2B data were downloaded using the Copernicus Open Access Data Hub platform on the European Spatial Agency’s (ESA) official website at https://scihub.copernicus.eu/ (accessed 30 August 2023). Both satellites (Figure 3) offer valuable data for various applications, including agriculture and forestry, enabling the prediction of crop yields and more. However, the Landsat-8 satellite achieves a complete Earth surface coverage in 16 days, while the Sentinel-2B satellite accomplishes the same task in a shorter period, i.e., 10 days.

2.5. Acquisition of Satellite Images

In the current study, eight images from various time periods were selected (Figure 4) using both satellites. These images were acquired from satellite sensors from mid-April to June to estimate straw cover percentage (

Table 2. Summary of the satellite images used for the estimation and mapping of Wheat Straw Cover (WSC).

No.	Acquisition Dates	Satellite	Sensor	Day of Year (DOY)
1	15 April 2019	Landsat-8	OLI-TIRS	105
2	1 May 2019	Landsat-8	OLI-TIRS	121
3	4 May 2019	Sentinel-2B	MSI	124
4	14 May 2019	Sentinel-2B	MSI	134
5	24 May 2019	Sentinel-2B	MSI	144
6	2 June 2019	Landsat-8	OLI-TIRS	153
7	3 June 2019	Sentinel-2B	MSI	154
8	23 June 2019	Sentinel-2B	MSI	174

). However, we selected images after the crop harvest to estimate straw coverage percentage because, in the mid-growing season, vegetation covering the soil and crop residue complicates the differentiation between soil and crop straw. This is due to the spectral similarity between straw and soil, with the main distinguishing factor being the amplitude at a specific, limited wavelength [49].

It was observed that, immediately after the harvest, the crop residues often exhibit a higher level of brightness compared to the soil and then, as the process of decomposition takes place, the crop residues may exhibit varying degrees of brightness in relation to the soil, with some residues appearing brighter and others appearing darker [28].

In addition, the wavelength values of all the selected satellites (Landsat-8 and Sentinel-2B) are shown in

Table 3. Details of the spectral band wavelengths of Sentinel-2B (MSI) and Landsat-8 (OLI-TIRS) satellites (Lillesand et al. [50] and Roy et al. [51]). NIR represents the Near-Infrared Band, 1-SWIR/2-SWIR are the Short-Wave Near-Infrared bands and 1-TIRS/2-TIRS are the Thermal Infrared bands.

Sentinel-2B		Landsat-8
Band Resolution(m)	$\mathbf{S}\mathbf{p}\mathbf{e}\mathbf{c}\mathbf{t}\mathbf{r}\mathbf{a}\mathbf{l} \mathbf{R}\mathbf{e}\mathbf{g}\mathbf{i}\mathbf{o}\mathbf{n}{ }_{ \mathbf{W}\mathbf{a}\mathbf{v}\mathbf{e}\mathbf{l}\mathbf{e}\mathbf{n}\mathbf{g}\mathbf{t}\mathbf{h}\left(\mu \mathbf{m}\right)}$	Band Resolution (m)	$\mathbf{S}\mathbf{p}\mathbf{e}\mathbf{c}\mathbf{t}\mathbf{r}\mathbf{a}\mathbf{l} \mathbf{R}\mathbf{e}\mathbf{g}\mathbf{i}\mathbf{o}\mathbf{n}{ }_{ \mathbf{W}\mathbf{a}\mathbf{v}\mathbf{e}\mathbf{l}\mathbf{e}\mathbf{n}\mathbf{g}\mathbf{t}\mathbf{h}\left(\mu \mathbf{m}\right)}$
b1 60m	Coastal Aerosol 0.43–0.45µm	b1 30m	Coastal Aerosol 0.43–0.45µm
b2 10m	Blue 0.45–0.53µm	b2 30m	Blue 0.45–0.5µm
b3 10m	Green 0.54–0.57µm	b3 30m	Green 0.53–0.59µm
b4 10m	Red 0.65–0.68µm	b4 30m	Red 0.63–0.67µm
b5 20m	1-Vegetation Red Edge 0.69–0.71µm	b5 30m	NIR 0.85–0.87µm
b6 20m	2- Vegetation Red Edge 0.73–0.74µm	b6 30m	1-SWIR 1.56–1.65µm
b7 20m	3- Vegetation Red Edge 0.77–0.79µm	b7 30m	2-SWIR 2.10–2.28µm
b8 10m	NIR 0.78–0.90µm	b8 15m	Panchromatic 0.50–0.67µm
b8A 20m	Narrow NIR 0.85–0.87µm	b9 30m	Cirrus 1.36–1.38µm
b9 60m	Water Vapor 0.93–0.95µm	b10 100m	1-TIRS 10.60–11.19µm
b10 60m	SWIR/Cirrus 1.36–1.39µm	b11 100m	2-TIRS 11.50–12.51µm
b11 20m	1-SWIR 1.56–1.65µm
b12 20m	2-SWIR 2.10–2.28µm

.

2.6. Image Processing and Vegetation Indices Analysis

All selected satellite images underwent pre-processing steps, such as radiometric calibration and atmospheric correction, implemented by means of various geospatial software. The images were subjected to atmospheric correction and subsequently masked for clouds and their associated shadows by means of a multi-temporal algorithm. However, for Landsat-8 datasets, Digital Numbers (DN) were transformed to TOA (Top-of-Atmosphere) reflectance by means of the Impact Portable GIS toolbox from the European Commission [52]. In contrast, for Sentinel-2B datasets, we kept the original Level-1C TOA product according to ortho-rectified surface reflectance values. Moreover, for estimating WSC through CRIs, the spectral bands available for both sensors were considered. Therefore, the B4, B5, B6 and B7 bands of Landsat-8 images were selected and integrated with the B4, B8, B8a, B11 and B12 bands of Sentinel-2B for a more comprehensive analysis. All the images were finally resized to the exact spatial resolution of 20 m, because the selected Landsat-8 spectral bands possess a 30 m spatial resolution. Accordingly, the dataset was rescaled to a 20 m spatial resolution, by employing the nearest point interpolation method and, for enhanced precision, the resample function within the raster package by using the bilinear interpolation method. Later, all the images underwent procedures to align them to the designated study sites. All these operations were conducted by means of the ENVI software, ver. 5.3.

Afterwards, the processing used for the estimation of straw cover percentages through satellite images was performed in three phases: (i) data preparation and extraction of Crop Residue Indices’ (CRIs) values; (ii) the extraction of multi-indices’ values and results onto a spreadsheet for further analysis; and (iii) the production of a crop straw cover map by means of the supervised maximum likelihood classification method. Moreover, the entire analysis process workflow of Vegetation Indices (VIs) and CRIs was performed using the ArcGIS Pro software ver. 3.0, Model Builder tool (Figure 5), including four primary processes: (a) the import of whole satellite scenes with the selected bands; (b) the extraction of spectral band values for the Region of Interest (ROI); (c) the analysis of the crop residues’ RS-based indices; and (d) the export of indices’ results onto a spreadsheet. The model utilized in this study was developed using the Model Builder of the ArcGIS software [53]. Its primary purpose was to optimize the process of calculating the various indices’ values for each satellite image. Before the calculations, statistical analysis and value quantification were conducted, while during the development of this model, several inputs were designated as parameters to allow users to modify them based on their region and the specific information. This was achieved by means of the Extract by Mask tool to reduce disk space and increase the processing performance.

Numerous indices have been developed to estimate crop straw cover, but two categories of indices were considered in this paper to determine the relationship between spectral properties and the straw cover percentage. The first indices were primarily derived from Landsat-8 spectral bands, while other comparable indices were developed for Sentinel-2B.

However, distinguishing crop straw from soils in Landsat imagery data was difficult, due to the similarity of their bands’ responses in the visible (VIS) and NIR spectral wavelengths, particularly under moist conditions [5]. In this study, crop residue/straw and Vegetation Indices were employed to estimate the straw cover, as shown in

Table 4. Selected RS-based indices for both multispectral satellites for the estimation of the straw coverage percentage.

Index/Abbreviation	Landsat-8	Sentinel-2B	Reference
Simple Tillage Index (STI)	b6/b7	b11/b12	[37]
Normalized Difference Tillage Index (NDTI)	(b6 − b7)/(b6 + b7)	(b11 − b12)/(b11 + b12)	[37]
Normalized Difference Index 5 (NDI5)	(b5 − b6)/(b5 + b6)	(b8 − b11)/(b8 + b11)	[14]
Normalized Difference Index 7 (NDI7)	(b5 − b7)/(b5 + b7)	(b8 − b12)/(b8 + b12)	[14]
Shortwave Red Normalized Difference Index (SRNDI)	(b7 − b4)/(b7 + b4)	(b12 − b4)/(b12 + b4)	[54]
Normalized Difference Soil Adjusted Vegetation Index (NDSAVI)	(b6 − b4)/(b6 + b4)	(b11 − b4)/(b11 + b4)	[55]
Normalized Difference Residue Index (NDRI)	(b4 − b7)/(b4 + b7)	(b4 − b12)/(b4 + b12)	[40]
Shortwave Green Normalized Difference Index (SGNDI)	(b3 − b7)/(b3 + b7)	(b3 − b12)/(b3 + b12)	[34]

Note: the letter “b” denotes the band of each satellite.

.

2.7. Calibration and Validation Datasets

The field data that were measured were randomly subdivided into two datasets: a calibration dataset, encompassing 65% of the total data; and a validation dataset, constituting 35% of the overall data. However, the calibration data were used for constructing regression equation models, which were subsequently employed in a validation dataset to assess the constructed model and evaluate the agreement between the measured and predicted WSC values. The data points in the calibration set were excluded from the validation dataset, and each dataset’s statistics for the WSC estimates are shown in

Table 5. Statistics of the field data for the WSC percentage across the study area.

Dataset	No. of Samples	Maximum	Minimum	Mean	Standard Error (SE)
Calibration dataset (65%)	52	90	25	61	2.10
Validation dataset (35%)	28	85	20	63	3.07

.

Moreover, the model performance was assessed by calculating the root-mean-square deviation (RMSD) between the measured and observed WSC values. Usually, the evaluation of a regression model’s performance involves assessing differences between the R² and RMSD values. The RMSD is determined by:

$$RMSD=\sqrt{\frac{1}{n}{\left(x-y\right)}^2}$$

(1)

where n is the number of the recorded WSC sampling points, x is the sum of the measured WSC values and y is the sum of the predicted values of WSC.

2.8. Mapping of the Wheat Straw Cover Percentage

During this stage, it was possible to conduct a sequence of intricate operations involving segmentation and classification procedures applied to the unclassified satellite scenes obtained from the study region. Several unsupervised and supervised classification techniques were considered in previous studies for mapping crop residue percentages in different classes. All of these approaches must address the challenges associated with imagery and data quality [56]. Using object-oriented approaches that integrate multispectral data, like Sentinel-2B satellite data, with an enhanced spatial resolution of 10 m proves even more advantageous to mitigate the aforementioned issues [57]. In this study, it was possible to use Sentinel-2B images to perform image classification for the mapping of WSC. This process starts with selecting sampling points by using ground survey data. The sampling points represent the different classes of the wheat residue cover percentages identified during the field survey. The ArcGIS Pro software allows us to identify these points by means of the ‘Draw Polygon’ tool, where polygons are drawn over the corresponding areas of the Sentinel-2B image. Once all sampling points are identified, a spectral signature for each class is generated. A spectral signature comprises the reflectance characteristics of a particular class. This is achieved by means of the ‘Image Classification’ toolbar in the ArcGIS ver. 10.8 software. Then, the supervised maximum likelihood classification (SMLC) technique was implemented. It is a statistical method based on probability and allows us to calculate the mean vectors calculated from the signature file of the ground samples. It assigns each satellite image pixel to the different WSC percentage classes with the highest likelihood. This was performed in the ArcGIS Pro Package by means of the Classify tool within the Image Classification toolbar, by generating signature files regarding different WSC percentages based on spectral bands’ remote sensing based on CRIs’ results. The above approach computes the statistics for each class within each band following a normal distribution and, then, calculates the likelihood of a given pixel being associated with a certain class. In order to verify the accuracy of the classification, a set of test points (different from the sampling sites and collected during the field survey) was compared with the classification result by means of a confusion matrix. The ‘Error Matrix’ tool was used to determine the users’ accuracy of the classification. Finally, the classified image was converted into a map representing different classes of wheat straw cover percentage in the study area.

2.9. Statistical Data Analysis

Linear and non-linear regression methods were employed in the development of prediction models for WSC. The objective of linear regression is to analyze the distribution of a response variable through a linear relationship involving one or more predictor variables. In order to evaluate the performance of the developed models, it was possible to calculate the coefficient of determination (R²) and root-mean-square Deviation (RMSD) for predicting the crop straw coverage percentage. The processed satellite images and recorded field survey data were subjected to analysis through the SPSS ver. 23 software [58].

2.10. Software and Hardware

ESA SNAP 7.0 (Sentinel Application Platform), ENVI 5.3, ArcGIS Pro and ArcGIS 10.8 were mainly used for image processing and the interpretation of the data for results. At the same time, IBM SPSS v23 was the mathematical computation tool selected for statistical data analysis. The required software was installed on an Intel (R) Core (TM) i7-2310 CPU (64 bit) with 16 GB RAM with Windows 10 platform, which was tested for successful performance.

3. Results

3.1. Field Data Analysis

The study area is mainly cultivated with wheat and rice rotational fields. The WSC percentage was calculated using a line-point transect method from 80 different sampling locations. The highest recorded WSC was 90%, and the lowest was higher or equal to 25%. However, for analyzing the field data, the sampling points were split into the calibration dataset, which was employed to construct regression models. Subsequently, the derived equation was applied to predict the residue cover percentage in the study area by using the validation dataset.

3.2. Performance of the RS-Based CRIs Models for WSC Estimation

The best performance of the model relationship of CRIs is shown in Figure 6, while the regression of the other indices is depicted in Figure 7. It was possible to acquire eight cloud-free satellite images, including three Landsat-8 and five Sentinel-2B images of the study area during 20 days, before and after field measurements. Two images were selected regarding suitable analysis for estimating the WSC from each satellite, i.e., 2 June 2019 and 3 June 2019 for Landsat-8 and Sentinel-2B, respectively. At the same time, the highest and lowest R² values were observed in NDTI and NDSAVI, respectively, using both satellite datasets.

Moreover, the coefficient of determination (R²) values of NDTI using Sentinel-2B and Landsat-8 were found to be 0.80 and 0.70, respectively, while NDSAVI values were 0.43 and 0.36, respectively.

However, the higher to lower values of R² concerning WSC were ranked as NDTI, STI, NDRI, NDI7, SRNDI, NDI5 and NDSAVI (

Table 6. Performance of the Crop Residue Indices (CRIs) in estimating the WSC percentage from the calibration dataset (n = 52 of the measured WSC sampling points).

Sr. No	Residue Indices	Satellite	Coefficient of Determination (R2)	Regression Equation	RMSD (%)
1	NDTI	Sentinel-2B	0.80 **	y= 189.07x0.4984	6.88
		Landsat-8	0.70 **	y = 48.195ln(x) + 152.51	8.56
2	STI	Sentinel-2B	0.78 **	y = 123.22x − 92.482	7.00
		Landsat-8	0.65 **	y = 106.98x − 85.417	8.76
3	NDRI	Sentinel-2B	0.64 **	y = 146.14x + 93.66	9.07
		Landsat-8	0.57 **	y = 162.18x + 88.683	9.86
4	NDI7	Sentinel-2B	0.60 **	y = 118.29x + 58.119	9.54
		Landsat-8	0.52 **	y = 111.79x + 45.773	10.42
5	SRNDI	Sentinel-2B	0.58 **	y = −233.55x + 106.72	10.02
		Landsat-8	0.60 **	y = −167.59x + 87.813	9.40
6	NDI5	Sentinel-2B	0.48 **	y = 90.056x + 65.13	10.84
		Landsat-8	0.40 **	y = 155.75x + 62.207	11.63
7	NDSAVI	Sentinel-2B	0.43 **	y = −139.93x + 104.04	11.32
		Landsat-8	0.36 **	y = −130.88x + 102.04	12.04

** Significant (p < 0.01 probability) and best-performing variables for each category are highlighted in bold.

), respectively. The results also indicate that the coefficient of determination (R²) value of SRNDI was marginally higher (0.60) by using Landsat-8 compared to Sentinel-2B (0.58). Moreover, the crop residue indices were fitted into various simple linear and non-linear regression models to accurately measure the WSC (Table 6). Thereby, the NDTIs for Sentinel-2B and Landsat-8 were fit into power and logarithm equations, respectively. All the other indices (STI, NDRI, NDI7, NDI5, SRNDI and NDSAVI) best fit a linear regression model. The RMSD values of the calibration models were found to be in the range from 6.88 to 12.04%, regarding different CRIs by using both satellite datasets.

3.3. Correlation between the Measured and Predicted WSC for Both Satellites

In order to assess the overall accuracy of each regression model used for the WSC estimation, the calibration dataset (n = 52) was compared to the validation dataset (n = 28) to find the correlation between the within-field measured and predicted WSC, as shown in Figure 8.

The performance of the relationship model between the measured and predicted WSC percentages showed that the highest and lowest coefficients of determination were calculated for NDTI and NDSAVI, respectively, for both data platforms. The positive variation obtained for NDTI was R² = 0.85 and R² = 0.77; followed by STI, R² = 0.82 and R² = 0.80; and NDRI, R² = 0.69 and R² = 0.56, compared to NDSAVI, R² = 0.53 and R² = 0.26 for Sentinel-2B and Landsat-8, respectively. However, the RMSD values of the calibration models ranged from 6.49 to 20.91%, regarding different CRIs (

Table 7. Performance of the developed crop residue indices models, associated with the measured and predicted wheat straw cover percentage for the validation dataset (n = 28 of the measured WSC sampling points) for both satellites.

Sr. No	Residue Index	Satellite	Coefficient of Determination (R2)	Regression Equation	RMSD (%)
1	NDTI	Sentinel-2B	0.85 **	y = 0.7651x + 13.979	6.49
		Landsat-8	0.77 **	y = 0.6091x + 25.042	8.22
2	STI	Sentinel-2B	0.82 **	y = 0.7419x + 15.704	6.80
		Landsat-8	0.80 **	y = 0.6405x + 23.947	7.69
3	NDRI	Sentinel-2B	0.69 **	y = 0.6206x + 21.707	9.27
		Landsat-8	0.56 **	y = 0.5369x + 24.717	11.51
4	NDI7	Sentinel-2B	0.63 **	y = 0.5403x + 28.777	9.86
		Landsat-8	0.49 **	y = 0.2706x + 29.265	20.91
5	SRNDI	Sentinel-2B	0.48 **	y = 0.4821x + 30.752	11.72
		Landsat-8	0.58 **	y = 0.516x + 29.561	10.52
6	NDI5	Sentinel-2B	0.56 **	y = 0.4681x + 32.058	10.84
		Landsat-8	0.46 **	y = 0.3576x + 40.138	11.97
7	NDSAVI	Sentinel-2B	0.53 **	y = 0.4612x + 34.426	11.05
		Landsat-8	0.26 **	y = 0.2246x + 50.057	13.83

** Significant (p < 0.01 probability) and best-performing variables for each category are highlighted in bold.

).

3.4. Mapping and Feature Classes of the WSC Percentages

The classification results (Figure 9) reveal that the Wheat Straw Cover (WSC) percentage was found to be in the range from 38 to 86%, while the Land Use Land Cover (LULC) indicated that the total paddy fields were 54.96%, while urban settlements, other agricultural fields and water bodies occupied areas of 25.14, 9.73 and 10.19%, respectively, in the study area. The results of estimating the WSC percentages demonstrated that 7.07, 16.29, 18.24 and 13.35% of the cropping area were classified as 0–40%, 41–60%, 61–80% and 81–100%, respectively. The maximum wheat straw cover percentage was determined in the category of 61–80% (18.24%) across the total wheat cropping region. Hence, these results are aligned with the field measurements of this work.

4. Discussion

Accurate Crop Straw Cover (CSC) information can contribute to determine numerous benefits, including regional-scale assessments and the effect of different amounts of straw used through a tillage method on cropping systems. Accurate measurements and robust methodologies used to estimate the Wheat Straw Cover percentage at large scales are still evolving by applying other multispectral satellites than Landsat. Presently, many multispectral satellites for extremely strong or scientifically valid applications are available for agricultural applications at a regional scale. This study aimed to address these gaps through the Crop Residue Indices (CRIs) and Sentinel-2B multispectral products that are either openly available with high temporal and spectral resolution satellites and are supportive of the RS scientists, in order to reduce the amount of data and processing time required for analysis. During this investigation, it was possible to measure the Wheat Straw Cover (WSC), by means of a line-point transect method in 80 rice—wheat rotational fields after crop harvest. The findings reveal that the straw cover ranged from 90 to 25%. This variation in coverage can be attributed to an accumulation of abundant Wheat Straw Cover at the sampling locations, due to the density of straw left on the ground after harvest. The results of this study are in agreement with the findings reported by Memon et al. [47] and Pforte et al. [59].

However, the relationship between CRIs and wheat straw cover was calculated by using remotely sensed indices (CRIs) for the calibration dataset (n = 52) and based on simple linear and non-linear regression models with coefficients of determination (R²). In addition, a positive relationship was observed between WSC and CRIs, such as NDTI, STI, NDRI, NDI7 and SRNDI, while an inverse relationship was observed between NDI5 and NDSAVI. In order to properly observe the relationships between WSC and CRIs, the R² regression values were adjusted. Moreover, the comparative analysis of model performance indicated that the performance of NDTI, STI and NDRI (Figure 8) by using Sentinel-2B was much better than that using Landsat-8 data as well as other CRIs (Figure 10) for predicting WSC over a homogeneous, agricultural regional area. A similar finding was obtained by Sharma et al. [60], who concluded that the linear correlation between the measured straw cover percentage and the predicted one through NDTI and STI was expressed by R² values of 0.89 and 0.78, respectively, from a validation dataset (n = 30) in Nebraska. Another research conducted by Ding et al. [42] indicated that the performance of CRIs for estimating the CRC from Sentinel-2 images was improved, and they found that NDRI had a positive correlation with CRC (R² = 0.77), followed by NDI7 and NDRI estimated from band-4 and band-12, by achieving a R² = 0.61 and RMSE = 6.663%. Moreover, Hively et al. [61] demonstrated a considerable concern related to residue indices, as evidenced by a roadside survey. Their study emphasized the vital role of vegetation in undermining the accuracy of straw estimation predictions. The findings indicated that the average NDTI surpassed the typical values associated with residue measurement in bare fields. This observation implies a significant susceptibility to the influence of green vegetation within the Worldview satellite imagery dataset.

Nevertheless, Thoma et al. [62] revealed that the Crop Residue Index Multiband (CRIM) exhibited a superior performance when applied to bean and corn residues, because it has different colors and texture characteristics along with distinct water absorption properties. Consequently, the dynamic range of the reflectance spectra for dry crop residues (absorption index of cellulose and lignin) on diverse soils was more remarkable for the absorption feature near 2.1 µm than for the 2.3 µm feature [7]. The structural components, i.e., cellulose, hemicellulose and lignin, within wheat straw, exert a predominant influence over the spectral characteristics of each sample at wavelengths exceeding 1.32 µm [29] and decrease near 1.7, 2.1 and 2.3 µm features [63]. Thus, the crop residue indices NDTI and STI were measured in this study using band-11 (1.57–1.65 µm) and band-12 (2.11–2.29 µm) of Sentinel-2B satellite and the data showed to be positivity correlated to the wheat straw cover estimation (Table 6). These results are also aligned with those of Jin et al. [54], who revealed that STI and NDTI were more appropriate for assessing the crop straw cover percentage with multi-temporal and spectral satellite data.

Additionally, the NDRI (R² = 0.69 and RMSD = 9.72%) for Sentinel-2B revealed a positive relationship for predicting the WSC in the study area, because the NDRI using bands 4 and 12 of Sentinel-2B decreased the effects of green vegetation on residue detection. These results agree with those of Gelder et al. [40], who concluded that the NDRI was chosen to resist vegetation impacts and performed the best with a R² = 0.81 and RMSE = 0.15, compared to other Vegetation Indices for Landsat imagery data. This trend is mostly associated with using Landsat-8, band-7, or Sentinel-2B, band-12. In fact, band-12 is likely to be a cellulose and lignin absorption band evident in crop residues/straw and plants but absent in soils. However, the results of this work for SRNDI show the highest R² = 0.58 in Landsat-8 compared to Sentinel-2B (R² = 0.48). The finding is strongly supported by the study of Jin et al. [54]. They developed a new index, i.e., the Shortwave Red Normalized Difference Index (SRNDI). This index was derived from Landsat band-7 and band-4 and evaluates the effectiveness of SRNDI with the existing CAI (Cellulose Absorption Index) and LCA (Lignin–Cellulose Absorption) indices in Jilin Province, China. Moreover, a study by Daughtry et al. [64] indicated that CAI outperforms other Vegetation Indices (VIs) for accurately predicting crop residues based on cellulose and lignin absorption features near 2100 nm. They further observed that SRNDI is highly affected by the maize residue cover, with a maximum R² of 0.71 (RMSE = 14%), compared to NDSAVI. In this way, Sentinel-2B data with RS-based crop residue indices (CRIs), such as NDTI, STI and NDRI, are more effective than Landsat-8 data for evaluating WSC.

Furthermore, in order to develop the distribution mapping of the WSC percentages and the extraction of a wheat cultivation area in the study area, the tested model’s linear equation of NDTI was used to estimate the WSC for the sampling points. Then, supervised maximum likelihood classification techniques were implemented to Sentinel-2B images acquired on 3 June 2019, because this classification method was as accurate as in the literature [65,66], even without the on-ground field samples. However, in this investigation aimed at mapping WSC (Figure 9), the straw cover percentages were divided into four distinct categories: (i) 0–40%; (ii) 41–60%; (iii) 61–80%; and (iv) 81–100%, although the water body, urban settlements and other agricultural fields were mentioned under different classes. Consequently, the overall precision of the estimated WSC percentage and other land cover classification was 86.53% and the kappa coefficient (κ) was 0.78. Therefore, the supervised classification algorithm demonstrated a satisfactory performance for accurately mapping crop straw cover in a large field. These results strongly agree with those of Memon et al. [47], who observed that the accuracy for mapping crop straw cover was 84.61% with κ = 0.76 by using Landsat-8 imagery in Changshu County. A similar research was conducted by Cai et al. [10], who performed the mapping of crop residue cover through Sentinel-2 images, and their findings indicated that only 0.74% of the farmlands had a CRC of 30–60% and 99.24% of farmlands had a CRC higher than 60% in the study area. They also concluded that NDTI was the best variable in estimating CRC in a large area and mapping it into different percentage classes. Additionally, Sharma et al. [60] and Tao et al. [67] preliminarily examined the possibility of mapping Crop Residue Cover (CRC) across various crops, and they discovered that the integration of multispectral satellite imagery with tillage indices proved to be a viable method for estimating and mapping CRC at the regional scale. Researchers additionally noted that the precise mapping of CRC into various percentage ranges offers many benefits, including the ability to conduct extensive assessments of diverse tillage methods, impact on soil quality, storage of soil organic carbon (SOC), analysis of crop phenology, policy formulation and crop simulation models for agricultural sustainability.

Although other studies by Thoma et al. [62] and Hively et al. [61] demonstrated that the utilization of Landsat satellite data for estimating crop residue coverage has an accuracy ranging from 61 to 69%, this accuracy was determined to be superior for estimates of straw coverage obtained by the Tillage Transect Survey (TTS). Yet, the human observers involved in the TTS frequently encounter challenges in distinguishing minor differences in the residue cover near the 30% threshold, which are employed for differentiating between conventional and conservation tillage practices [61,68]. Therefore, Sentinel-2B multispectral imagery can provide a uniform approach for estimating crop straw cover and develop a detailed WSC mapping with a higher accuracy at the regional scale.

5. Conclusions

In this study, it was possible to compare two satellites (Landsat-8 and Sentinel-2B) to estimate the Wheat Straw Cover (WSC) percentage. However, the used approach combined RS-based crop residue (CRIs) and Vegetation Indices (VIs) to develop a model for estimating the WSC in the study area. The results demonstrate that NDTI, STI and NDRI by using Sentinel-2B data were more strongly correlated with the WSC percentage than using Landsat-8 spectral data. The NDTI of Sentinel-2B performed the best in mapping the WSC percentage in distinct classes, with an overall accuracy of 86.53% (κ = 0.78), compared to other CRIs. NDTI had a positive regression relationship for Sentinel-2B and Landsat-8, expressed by an R² of 0.80 and 0.70, respectively, as well as RMSD values of 6.88 and 8.56%, respectively. NDTI, STI and NDRI also showed the highest coefficient of determination values of 0.85, 0.82 and 0.69, respectively, between the linear relationship for the measured and predicted wheat straw cover percentages. Thus, it is highly recommended that multispectral Sentinel-2B imagery, which has a better spatial and temporal resolution compared to those of Landsat-8, is used to estimate the WSC percentage over a large scale, using remote-sensing and GIS techniques. Additionally, this study provides a foundation for scientists to conduct future research on the effects of distinct crops, climate conditions, vegetation health and human activities by using remote sensing and machine learning techniques, which can validate the used technique and the obtained results to achieve comprehensive outcomes. These data can further assist scientists to monitor and understand how agricultural systems in various regions address the challenges of global climate change and food security in the context of agricultural sustainability.