2.1. Overview of Study Area
This study focuses on the Mu Us Sandy Land, located in the arid and semi-arid region of northern China. The Mu Us Sandy Land is the transition area between the Mu Us desert and its surrounding geomorphic types, generally situated in the south of the Hetao Plain and surrounded by the Yellow River on three sides of the northwest and east (as shown in
Figure 1c). The hydrothermal conditions of the Mu Us Sandy Land are better than other sandy lands in China. The eastern part of the Mu Us Sandy Land is characterized by typical grassland vegetation, while the western part is mainly desert steppe, featuring dry-denuded bedrock hills, tablelands, and a small amount of aeolian dunes. In contrast, the western part of Otuoke Banner, located near the Yellow River, is exposed bedrock. The core desert area of the Mu Us Sandy Land is characterized by the presence of mobile and coppice dunes, which mainly consist of barchan dunes and dune chains. The area also has various types of wind erosion pits and parabolic dunes. The beaches are distributed among sand belts composed of dunes, and saline-alkali lands are common on the beaches [
36]. The vegetation communities in the moderately and extremely severe desertification areas were mainly composed of herbaceous pioneer vegetation and large shrubs, while the dominant species in the area were
Artemisia ordosica Krasch. and
Corethrodendron fruticosum var.
mongolicum (Turcz.)
Turcz. ex Kitag. The building species in the mild desertification area were
Caragana sinica (Buc’hoz)
Rehder and
Salix gordejevii Y. L. Chang et Skv. while the coppice dune types were
Artemisia ordosica Krasch.,
Caragana sinica (Buc’hoz)
Rehder, and
Salix gordejevii Y. L. Chang et Skv. The main types in the Gobi desertification areas are composed of annual and perennial herbs [
43,
44].
The annual total precipitation in the Mu Us Sandy Land region decreases from southeast to west, ranging from 440 mm to 250 mm [
6]. As shown in
Figure 1b, the peak monthly precipitation occurs in August. The annual average precipitation is about 340 mm, which is higher than that of other sandy lands. From 1995 to 2019, the precipitation in 1999–2000 was significantly lower than the average precipitation of the past 60 years. The precipitation in 2005 was also very low, but in 2004 it was slightly above the long-term average. The annual average precipitation in 2010 and 2011 was basically the same as that of many years, while the annual average precipitation in 2014 and 2015 was slightly below the long-term average for two consecutive years. The Mu Us Sandy Land region has an annual average temperature ranging from 6 °C to 8.5 °C, and the temperature decreases from south to north. The region also belongs to the arid to semi-arid climate region, with a maximum wind speed of up to 28 m/s [
44].
2.2. NPV and PV Data Sources and Calculation Methods
This study utilized MODIS NDVI 8-day products and SWIR76 data to calculate the monthly average NPV, PV, and bare soil products from 2000 to 2015, based on linear spectral unmixing. This approach was developed by Guerschman et al. [
5] using hyperspectral data collected from over 1000 field sites. The SWIR76 index is defined as the ratio of the SWIR7 band (2130 nm) to the SWIR6 band (1640 nm) in MODIS data.
It is important to note that the spatial resolution of MODIS data, which is greater than 250 m, may result in inaccurate analysis of NPV and PV in the Mu Us desert area, where shrub desertification is predominant. Consequently, a key focus of this study is to develop NPV and PV estimation models using Landsat series imagery through machine learning techniques.
The construction process of the machine learning model in this study is divided into three parts: data collection for the model output, external validation of the model, and model selection. First, data collection for the model output used UAV imagery collected in the field, covering a total of 457 points from 2017 to 2023. The UAV imagery processing included radiometric correction and mosaicking, object-based multiscale segmentation, resulting in NPV, PV, and bare soil cover data for 457 points, each with a spatial coverage exceeding 30 × 30 m. Second, the input for model construction consisted of Landsat spectral data corresponding to the UAV sampling points. With both the output and input data combined to form the dataset, 70% of the data were used to build the model, and 30% were used for the test set. To minimize the impact of data distribution bias on accuracy, ten-fold cross-validation was employed to average the final accuracy. Notably, the data used for model construction were also split into a training set and a validation set at a 70% to 30% ratio. Finally, the model selection was based on the final accuracy obtained from the ten-fold cross-validation of the test set.
The formulas mentioned in the above methods are explained as follows. In Formulas (1) and (2),
is the predicted value,
is the true value,
is the average of the predicted values, and
n is the number of samples.
The ten-fold cross-validation method is commonly used to test the accuracy of algorithms. It involves dividing the dataset into 10 parts, then iteratively using 9 parts as training data and 1 part as test data. Each iteration yields an accuracy (or error) rate, and the average of the ten results provides an estimate of the algorithm’s precision.
The development of a machine learning model first requires the identification of its inputs and outputs. The model inputs will include indices such as NDVI (Normalized Difference Vegetation Index) [
18], Enhanced Vegetation Index (EVI) [
19], Normalized Difference Senescence Vegetation Index (NDSVI) [
20], Landsat 8-OLI Dead Fuel Index (OLI-DFI) [
21], Modified Soil-Adjusted Vegetation Index (MSAVI) [
22], Ci-green, and Normalized Difference Index (NDI) [
23].
In this study, the mesh search method is used in the parameter adjustment process of the machine learning model. The parameter settings for the model tested in this study are detailed as follows:
RF Parameters: To optimize for minimal mean squared error (MSE) and fast convergence, the parameters for the RF regression models, specifically for NPV and PV, were set with a minimum of five leaf nodes and requiring over 50 decision trees. Additional parameters, such as the depth of individual trees and the number of features considered by each tree, were fine-tuned based on MSE values during the algorithmic process.
BP Neural Network Parameters: The BP neural network comprises three layers. The transfer function from the input layer to the first hidden layer is the logsig function, with the number of neurons experimentally determined. From the first hidden layer to the second hidden layer, the transfer function is the softmax function, ensuring the sum of the three neurons in the output layer equals one. The transition from the second hidden layer to the output layer utilizes the purelin function. The output layer consists of three nodes representing the fractions of NPV, PV cover, and bare soil within a pixel. The error loss function employed is the mean squared error function, and the gradient descent is executed using the Levenberg–Marquardt algorithm. To ensure model convergence, the learning rate and maximum number of fail attempts were set to 0.01 and 10, respectively.
FCNN Parameters: Apart from the input and output layers, this deep learning network includes five hidden layers, each with 128 neurons and the rectified linear unit (ReLU) function as the transfer function between layers. The overall performance of each model is illustrated in
Appendix A Figure A1.
The preprocessing of the remote sensing images included radiometric correction, geometric correction, color balancing, and mosaicking. Atmospheric correction was performed using the FLAASH module in ENVI. Histogram matching was conducted for other images and test period images, followed by mosaicking of all employed images. This study downloaded all low-cloud-cover Landsat 5 TM images taken from 17 September to 17 October for the years 2000 to 2012. Considering the dates and cloud cover, the missing images were replaced with those from the same period in the previous or following year. For the data from 2013 to 2019, Landsat 8-OLI images from 17 September to 17 October were used. Similarly, missing images were substituted with those from the same period in the previous or following year, considering both dates and cloud cover. The frequency of image acquisition for each period and the corresponding sensors are documented in
Appendix A Table A1,
Table A2,
Table A3 and
Table A4.
The Landsat Operational Land Imager (OLI) and Thematic Mapper (TM) sensors have been found to show differences in reflection values for the same ground objects. For instance, in urban mapping, the reflectance difference of Landsat 8-OLI images for different ground objects is more noticeable than that of Landsat 5 TM [
45]. However, some studies have shown that there is a high correlation between vegetation indices and reflectance of the same ground object using different sensors in large and uniform areas [
46]. To account for these differences, a regression relationship between the Landsat 8-OLI and Landsat 5 TM bands was established before building the NPV and PV coverage detection models for Landsat 5 TM images. The reflection values of Landsat 8-OLI images corresponding to sample points were then back-calculated to reflect the values of Landsat 5 TM images through this regression relationship. Finally, a new machine learning regression fitting model was built using the NPV coverage and PV coverage data from UAV sampling points.
2.4. Temporal and Spatial Regression Analysis
In this study, the PV and NPV from September 2001 to September 2019 were analyzed using pixel-by-pixel regression with MODIS NPV and PV products having a resolution of 500 m. The annual total precipitation and annual mean temperature of the same resolution were used for the analysis. The time delay effect of the current year refers to the total precipitation and average temperature from January to September of the image year, while the time delay effect of the previous year refers to the total precipitation and average temperature from 10–21 months before the image acquisition month. Similarly, the time delay effects for the first two years (22–33 months), the first three years (34–45 months), and the first four years (46–57 months) were also analyzed.
Regression analysis was conducted at a regional scale to explain the response of NPV and PV coverage, calculated from Landsat series images, to precipitation and multi-scale SPEI. The monthly total rainfall data for the first 21 months of the NPV and PV coverage data of the current period were considered. For example, when the data for the period were from Landsat image with the strip number 127032 in 2005 (obtained on 7 October), the corresponding rainfall data were from January 2004 to September 2005. Similarly, the NPV and PV data were analyzed against the cumulative drought and moisture conditions on four-time scales of SPEI-1, SPEI-3, SPEI-6, and SPEI-12 using SPEI data. In this study, these datasets were used to analyze the response of spatial patterns of vegetation to precipitation in the geographical ecological region.
Based on the explanation of the above methods and data, the technical flowchart of this research is shown in
Figure 2.