Modeling with Hysteresis Better Captures Grassland Growth in Asian Drylands

Abstract

Climate warming hampers grassland growth, particularly in dryland regions. To preserve robust grassland growth and ensure the resilience of grassland in these arid areas, a comprehensive understanding of the interactions between vegetation and climate is imperative. However, existing studies often analyze climate–vegetation interactions using concurrent vegetation indices and meteorological data, neglecting time-lagged influences from various determinants. To address this void, we employed the random forest machine learning method to predict the grassland NDVI (Normalized Difference Vegetation Index) in Asian drylands (including five central Asia countries, the Republic of Mongolia, and Parts of China) from 2001 to 2020, incorporating time-lag influences. We evaluated the prediction model’s performance using three indexes, namely the coefficient of determination (R²), root-mean-square error (RMSE), and Mean Absolute Error (MAE). The results underscore the superiority of the model incorporating time-lag influences, demonstrating its enhanced capability to capture the grassland NDVI in Asian drylands (R² ≥ 0.915, RMSE ≤ 0.033, MAE ≤ 0.019). Conversely, the model without time-lag influences exhibited relatively poor performance, notably inferior to the time-lag-inclusive model. The latter result aligns closely with remote sensing observations and more accurately reproduces the spatial distributions of the grassland NDVI in Asian drylands. Over the study period, the grassland NDVI in Asian drylands exhibited a weak decreasing trend, primarily concentrated in the western region. Notably, key factors influencing the grassland NDVI included the average grassland NDVI in the previous month, total precipitation in the current month, and average soil moisture in the previous month. This study not only pioneers a novel approach to predicting grassland growth but also contributes valuable insights for formulating sustainable strategies to preserve the integrity of grassland ecosystems.

1. Introduction

As one of the most widely distributed vegetation types in terrestrial ecosystems [1], grasslands play a vital role in regulating the global carbon cycle and sustaining the structure and function of ecosystems [2,3]. Grassland ecosystems also serve as a key indicator of global climate change, with physiological processes such as photosynthesis and respiration linking natural moisture, atmosphere, and soil [4]. However, the growth of grasslands is strongly influenced by the external environment, especially in dryland areas [5,6]. In the context of climate warming, frequent extreme events profoundly impact grassland growth [7,8]. Recent extensive research on the impacts of climate change on grassland growth indicates that rising temperatures and increased drought could hinder growth or even lead to stagnation [9,10,11,12]. Therefore, obtaining reliable and objective information on grassland growth in a timely manner is crucial [13].

To safeguard grassland growth in dryland areas, it is imperative to accurately simulate the vegetation growth process and predict growth conditions. However, predicting vegetation growth using current tools is challenging [14]. Among the available tools, process-based ecosystem models play a vital and beneficial role in predicting vegetation growth [3]. Achieving accurate predictions of vegetation growth requires these models to reproduce multiple processes with greater precision and realism [15,16]. However, process-based ecosystem models face limitations in accurately replicating key ecosystem processes such as photosynthesis and respiration [16,17]. Recognizing these limitations, machine learning methods have emerged as promising alternatives, capable of comprehensively considering influential factors reflecting ecosystem processes. Machine learning methods effectively compensate for the shortcomings of process-based ecosystem models [18]. Techniques such as support vector machines, random forests, and gradient-boosted trees are employed to elucidate relationships between variables and predictor variables [11,18]. However, research on predicting vegetation growth in Asian drylands remains relatively limited [11,19]. Many scholars have investigated the prediction of vegetation NDVI growth using this method. For instance, Li et al. employed the extreme gradient boosting machine learning method to establish a relationship model between vegetation NDVI and meteorological factors in mainland China. They discovered that the machine learning model exhibits strong performance in replicating the spatial and seasonal variations in satellite-derived NDVI throughout China [11]. Peng et al., on the other hand, utilized a novel machine learning method to forecast the NDVI growth of vegetation during the growing season in China under extreme conditions. The findings indicate that the machine learning method demonstrates stability and high accuracy even amidst severe drought conditions [19]. These studies successfully replicated the spatial and temporal distributions of vegetation growth in China, employing machine learning to assess its accuracy in predicting growth under extreme drought conditions. Machine learning was an important tool for predicting vegetation growth, enhancing our overall predictive capabilities. It is conducive to the timely acquisition of reliable, objective, and timely information on vegetation growth.

Asian drylands boast abundant grassland resources, playing a crucial role in the sustainable development of the local livestock economy [20,21]. For centuries, the indigenous communities have relied on the plentiful resources of the grasslands for the growth of animal husbandry, establishing them as a pivotal industry essential for their survival [22]. Consequently, forecasting information regarding Asian dryland grasslands plays a crucial guiding role in shaping local animal husbandry practices. In-depth research has studied the response mechanism of the grassland NDVI and climate change. A discernible time-lag effect was observed between the grassland NDVI and various climate elements, indicating a delayed response [23,24,25,26]. In Asian drylands, the grassland NDVI in the current month showed a stronger response to solar radiation and the previous month’s soil moisture, without significant lag effects from temperature and total precipitation [25]. This suggests that changes in climatic factors could impact the grassland NDVI within a few months. However, there are few studies that consider the influence of time lag in predicting the grassland growth of Asian drylands, highlighting a notable gap in the current scientific literature. Employing the random forest machine learning method, we developed an NDVI prediction model for Asian dryland grasslands, incorporating both time-lag influences and their absence. Our main research objectives were (1) to evaluate the accuracy of the grassland NDVI predictions from 2001 to 2020, accounting for and excluding time-lag influences; (2) to investigate the spatiotemporal distributions of the NDVI in Asian grasslands from 2001 to 2020 using the two distinct models; and (3) to identify the primary climate factors influencing the NDVI in arid Asian regions based on the outcomes derived from the two models.

2. Materials and Methods

2.1. Study Area

The research area is the Asian drylands, ranging from 46.49° to 126.08°E and 31.54° to 55.43°N, geographically covering six countries: Kazakhstan, Kyrgyzstan, Tajikistan, Turkmenistan, Uzbekistan, and Mongolia, as well as certain regions of China, namely Inner Mongolia, Shaanxi, Ningxia, Gansu, Qinghai, and Xinjiang (refer to Figure 1, adapted from MODIS land cover product, specifically MCD12C1, accessible at https://lpdaac.usgs.gov/pro-ucts/mcd12c1v006/ (accessed on 18 May 2018)). Spanning over 10 million square kilometers, this region features a distinctive continental arid climate characterized by frequent droughts, low precipitation, and escalating temperatures [27,28]. The eastern Asian drylands experience fewer than 200 mm of annual precipitation, featuring a cold and dry winter and high temperatures in summer [29,30]. The western part (the five central Asia countries) holds an annual precipitation of 300–500 mm, and this region witnesses cold winters, hot summers, and substantial day-to-day temperature variations [29,31]. Under arid and semiarid climatic conditions [32], the main land use type is grassland, with sparse vegetation and extensive desert areas in the central region.

2.2. Data Resources

2.2.1. MODIS NDVI

The MODIS NDVI (Normalized Difference Vegetation Index) data are sourced from the National Aeronautics and Space Administration’s (NASA) Moderate Resolution Imaging Spectroradiometer (MODIS) Vegetation Index (VI) product (MOD13C2) (https://ladsweb.modaps.eosdis.nasa.gov) (accessed on 15 April 2022). This dataset spans the period from January 2001 to December 2020, featuring a spatial resolution of 0.05° × 0.05° and a temporal resolution of 1 month. This dataset underwent rigorous preprocessing, including geometric, atmospheric, and radiometric corrections, as well as validation procedures to reduce the impact of atmospheric factors such as cloud shadows and aerosols [33]. The monthly MOD13C2 data are widely recognized for their reliability and are extensively employed for monitoring vegetation conditions at both regional and global scales [34,35].

2.2.2. Climate Dataset

Climate variables include temperature, precipitation, scPDSI (self-calibrating Palmer Drought Severity Index), minimum air temperature, and maximum air temperature. These datasets were obtained from the monthly gridded Climatic Research Unit Time-Series version 4.05 (CRU TS4.05), which is provided by the Climatic Research Unit (CRU) at the University of East Anglia, UK (https://catalogue.ceda.ac.uk/uuid/c26a65020a5e4b80b20018f148556681) (accessed on 21 April 2022) [36]. It features a spatial resolution of 0.5° × 0.5° and a temporal resolution of 1 month, encompassing the period from January 2001 to December 2020. Currently, CRU is one of the most extensively utilized climate datasets. It integrates several well-known existing databases and uses angular distance weighted interpolation to generate global monthly data covering the land surface from 1901 to 2020.

Solar radiation data for model training and testing are sourced from the European Centre for Medium-Range Weather Forecasts (ECMWF) version 5 reanalysis (ERA5) dataset (https://cds.climate.copernicus.eu/cdsapp#!/home) (accessed on 27 April 2022). Representing the latest generation of ECWMF reanalysis data, ERA5 offers improved spatial and temporal resolution, as well as enhanced radiative transfer modeling compared with its predecessor, ERA interim reanalysis [11]. These data can be acquired from 1979 to the present, with a spatial resolution of 0.1° × 0.1° and a temporal resolution of 1 month. For this study, data from January 2001 to December 2020 are utilized.

Vapor Pressure Deficit (VPD) datasets are extracted from the Terra-Climate dataset provided by the University of Idaho (https://climate.northwestknowledge.net/TERRACLIMATE/index_directDownloads.php) (accessed on 5 March 2023). The Terra-Climate dataset provides vpd information at a spatial resolution of 1/24° × 1/24°, with a temporal resolution of 1 month. The data cover the time period from January 2001 to December 2020.

2.2.3. Soil Moisture

The daily soil moisture datasets are obtained from the soil moisture dataset within the European Space Agency (ESA) Climate Change Initiative Program (CCI) (https://www.esa-soilmoisture-cci.org) (accessed on 6 May 2022). It is featured with a spatial resolution of 0.25° × 0.25°, and the data span from January 2001 to December 2020. The soil moisture product was derived by synthesizing an active microwave soil moisture product produced by Bartalis et al. [37] and Wagner et al. [38] and a passive microwave soil moisture product developed by Vrije Universiteit Amsterdam in collaboration with NASA [39,40]. This composite product accurately represents soil moisture levels at a depth ranging from 0 to 10 cm. Due to limitations in satellite sensor coverage, there were significant gaps in the data, resulting in missing pixels. Thus, we removed the missing data points (7%) and computed monthly averages of soil moisture data, ensuring comprehensive coverage of the study area on a month-by-month basis.

2.2.4. Land Use Data

The land cover data were resourced from NASA’s MODIS land cover product MCD12C1 (https://lpdaac.usgs.gov/products/mcd12c1v006/) (accessed on 18 May 2018). This dataset features a spatial resolution of 0.05° × 0.05°. This particular research focused on identifying grassland image elements that remained consistent throughout the entire period from 2001 to 2020. This approach was employed to some extent to mitigate the land use impact.

2.3. Method

2.3.1. Trend Analyses

We used least squares linear regression to analyze the spatial trends of the grassland NDVI in the dryland areas of Asia from 2001 to 2020. F-test was used to analyze the significance of the trends [41].

$$\mathit{S}\mathit{l}\mathit{o}\mathit{p}\mathit{e}=\frac{\mathit{n}{\sum }_{\mathit{i}=\mathsf{1}}^{\mathit{n}}\left(\mathit{i}\times {\mathit{N}\mathit{D}\mathit{V}\mathit{I}}_{\mathit{i}}\right)-{\sum }_{\mathit{i}=\mathsf{1}}^{\mathit{n}}\mathit{i}\times {\sum }_{\mathit{i}=\mathsf{1}}^{\mathit{n}}{\mathit{N}\mathit{D}\mathit{V}\mathit{I}}_{\mathit{i}}}{\mathit{n}{{\sum }_{\mathit{i}=\mathsf{1}}^{\mathit{n}}{\mathit{i}}^{\mathsf{2}}-\left({\sum }_{\mathit{i}=\mathsf{1}}^{\mathit{n}}\mathit{i}\right)}^2}$$

(1)

where $\mathit{S}\mathit{l}\mathit{o}\mathit{p}\mathit{e}$ is the slope of the image element, ${\mathit{N}\mathit{D}\mathit{V}\mathit{I}}_{\mathit{i}}$ is the mean value of NDVI in the ith year, and n is the length of the study.

Although the observed NDVI indeed provides valuable data on current vegetation conditions over Asia’s drylands, there are several compelling reasons for focusing our trend analysis on predicted NDVI. Firstly, the predicted NDVI reveals possible future vegetation patterns under certain environments, allowing model accuracy assessment through comparison with actual observations. Secondly, the reliable NDVI prediction model helps identify links between vegetation growth and the environment, enhancing ecological understanding. Finally, the predicted NDVI supplements observed data, especially in data-scarce or observation-challenging areas, offering a fuller picture of vegetation trends.

2.3.2. Random Forest

The random forest algorithm, proposed and developed by Leo Breiman and Adele Cutler [42], stands out as a powerful tool for addressing both classification and regression problems. Its core concept diverges from the conventional approach of constructing a single, large decision tree using the entire training dataset. Instead, the random forest technique generates multiple decision trees using different subsets of training samples and attributes. These individual decision trees are then combined in a random fashion to create a more robust and effective model, enhancing the overall performance of the random forest model [43,44].

The MODIS NDVI product was used to represent vegetation growth, a choice commonly employed in previous research [4,5]. Previous studies identified a significant time-lag effect on the grassland NDVI in Asian drylands [25]. Because vegetation growth in a current month is influenced by the preceding month, the NDVI from the previous month was also incorporated into the model as an explanatory variable during training. This study utilized MODIS NDVI data from 2001 to 2020 in Asian drylands as the dependent variable, with various other factors serving as independent variables (See

Table 1. Model details: independent variable, dependent variable, and random forest parameter.

	Independent Variable	Dependent Variable	Temporal Resolution	Spatial Resolution	Time Span	Model Parameter
with time-lag	tmp_0, pre_0, rad_1, sm_1 tmx_0, tmn_0, vpd_0, NDVI_1	NDVI	One month	0.25°	January 2001–December 2020	70% training set, 30% validation set
without time-lag	tmp_0, pre_0, rad_0, sm_0 tmx_0, tmn_0, vpd_0, NDVI_1	NDVI	One month	0.25°	January 2001–December 2020	70% training set, 30% validation set

Notes: NDVI_1 represents NDVI in the previous month; tmp_0 represents temperature in the current month; pre_0 represents precipitation in the current month; sm_0 represents soil moisture in the current month; sm_1 represents soil moisture in the previous month; rad_0 represents solar radiation in the current month; rad_1 represents solar radiation in the previous month; vpd_0 represents vapor pressure deficit in the current month; tmn_0 represents minimum temperature in the current month; tmx_0 represents maximum temperature in the current month.

). Each image for each explanatory variable covered an area of 30,528 pixels, for a total of 240 images for a variable. The overall dataset for the random forest encompassed 7,326,720 samples, with 70% allocated for training and 30% for validation.

2.3.3. Statistical Analysis

Two models were developed using the random forest machine learning method to predict the grassland NDVI in Asian drylands. One model excluded time-lag influences, while the other incorporated these influences. The precision differences were then compared between the two models. R² was used to measure how effectively each model explained the variation in the data; RMSE served as a metric to measure the standard deviation between the true and predicted values; and MAE was used to measure the size of the gap between the true and predicted values [3,42].

$${\mathit{R}}^{\mathsf{2}}=\mathsf{1}-\frac{{\sum }_{\mathit{i}=\mathsf{1}}^{\mathit{n}}{({\mathit{y}}_{\mathit{i}}-{\widehat{\mathit{y}}}_{\mathit{i}})}^2}{{{\sum }_{\mathit{i}=\mathsf{1}}^{\mathit{n}}({\mathit{y}}_{\mathit{i}}-{\overline{\mathit{y}}}_{\mathit{i}})}^2}$$

(2)

$$\mathit{R}\mathit{M}\mathit{S}\mathit{E}=\sqrt{\frac{\mathsf{1}}{\mathit{n}}{\sum }_{\mathit{i}=\mathsf{1}}^{\mathit{n}}{({\mathit{y}}_{\mathit{i}}-{\widehat{\mathit{y}}}_{\mathit{i}})}^2}$$

(3)

$$\mathit{M}\mathit{A}\mathit{E}=\frac{\mathsf{1}}{\mathit{n}}{\sum }_{\mathit{i}=\mathsf{1}}^{\mathit{n}}|{\mathit{y}}_{\mathit{i}}-{\widehat{\mathit{y}}}_{\mathit{i}}|$$

(4)

3. Results

3.1. Evaluation of Model Performance in Predicting Grassland NDVI in Asian Drylands

The results reveal that the NDVI model incorporating time-lag effects outperforms the model without time-lag across all evaluation indices. The overall coefficient of determination for the model considering time-lag influences was 0.92, whereas the NDVI prediction model without time-lag influences achieved a coefficient of determination of 0.89 (Figure 2a,b). The R² values indicated that these models could explain 92% and 89% of the variations in the grassland NDVI, respectively. Further, the RMSE and MAE values of the model considering time-lag influences were 0.033 and 0.019, respectively, while values of the model without time-lag influences were 0.037 and 0.022, respectively (Figure 2a,b). The results from the three evaluation metrics collectively demonstrate that the model considering time-lag influences outperformed the model without time-lag influences in predicting the grassland NDVI in Asian drylands. As depicted in Figure 2, both models fitted well with observations, with the majority of NDVI values clustering between 0 and 0.25. This visual representation affirmed the reliability and accuracy of the developed models in capturing the dynamics of the grassland NDVI in the study region.

Figure 2. Density scatter plot showing the relationship between the observed MODIS and predicted NDVI values. Notes: (a) illustrates MODIS NDVI versus predicted NDVI with time-lag influences considered, while (b) depicts MODIS NDVI versus predicted NDVI without time-lag influences. The black line represents the 1:1 correspondence, and the red line indicates the regression line. According to the spatial distributions of R², the developed model for predicting grassland NDVI considering time-lag influences outperformed the model without time-lag influences (Figure 3). Regions exhibiting notably high R² values were mainly concentrated in Kazakhstan (Figure 3a,b). Regarding RMSE index, the grassland NDVI model considering time-lag influences demonstrated superior performance compared with the model without time-lag influences, particularly in the northeastern part of the Mongolian Plateau and Kazakhstan (Figure 3c,d). Evaluating the MAE values from the two models, the areas with relatively high MAE of the model considering time-lag influences were mainly distributed in the central part of the Mongolian Plateau and Kazakhstan (Figure 3e,f). The spatial distributions of these three indexes (R², RMSE, MAE) collectively indicate that both models were effective in predicting grassland NDVI in Asian drylands. However, the model considering time-lag influences exhibited superior performance, especially in the Kazakhstan region (Figure 3). These findings highlight the significance of considering time-lag influences to enhance the accuracy of predictions, particularly in specific geographical areas.

Figure 2. Density scatter plot showing the relationship between the observed MODIS and predicted NDVI values. Notes: (a) illustrates MODIS NDVI versus predicted NDVI with time-lag influences considered, while (b) depicts MODIS NDVI versus predicted NDVI without time-lag influences. The black line represents the 1:1 correspondence, and the red line indicates the regression line. According to the spatial distributions of R², the developed model for predicting grassland NDVI considering time-lag influences outperformed the model without time-lag influences (Figure 3). Regions exhibiting notably high R² values were mainly concentrated in Kazakhstan (Figure 3a,b). Regarding RMSE index, the grassland NDVI model considering time-lag influences demonstrated superior performance compared with the model without time-lag influences, particularly in the northeastern part of the Mongolian Plateau and Kazakhstan (Figure 3c,d). Evaluating the MAE values from the two models, the areas with relatively high MAE of the model considering time-lag influences were mainly distributed in the central part of the Mongolian Plateau and Kazakhstan (Figure 3e,f). The spatial distributions of these three indexes (R², RMSE, MAE) collectively indicate that both models were effective in predicting grassland NDVI in Asian drylands. However, the model considering time-lag influences exhibited superior performance, especially in the Kazakhstan region (Figure 3). These findings highlight the significance of considering time-lag influences to enhance the accuracy of predictions, particularly in specific geographical areas.

Figure 3. Spatial distributions of three indexes of grassland NDVI without time-lag influences and grassland NDVI with time-lag influences. Notes: (a,b) coefficient of determination (R²); (c,d) root-mean-square error (RMSE); (e,f) root-mean-square error (MAE).

Figure 3. Spatial distributions of three indexes of grassland NDVI without time-lag influences and grassland NDVI with time-lag influences. Notes: (a,b) coefficient of determination (R²); (c,d) root-mean-square error (RMSE); (e,f) root-mean-square error (MAE).

3.2. Comparing Spatiotemporal Variations between the Two Models

We reproduced the annual variation in the mean grassland NDVI in Asian drylands from 2001 to 2020 (Figure 4). The annual average grassland NDVI with time-lag influences was closer to the observed MODIS NDVI. The average NDVI for the grassland NDVI model, both with a time-lag influence (0.17) and without it (0.21), was higher than the annual mean NDVI derived from MODIS NDVI (0.14). Among the three NDVI values, all had consistent fluctuations in the grassland NDVI change (Figure 4). The time series of MODIS NDVI and NDVI from the two models showed a weak decreasing trend from 2001 to 2020.

We could observe that both models were able to reproduce the spatial distributions of the annual mean grassland NDVI in Asian drylands, and they were consistent with the spatial distributions of the observed MODIS NDVI (Figure 5a–c). The model considering time-lag influences was better than that without time-lag influences, as the former model was similar to the observed annual mean MODIS NDVI (Figure 5b). The annual mean NDVI decreased from the periphery to the center (both for the five central Asia countries and the Mongolia Plateau) (Figure 5a–c). Compared with the annual mean NDVI predicted by the model without time-lag influences, the annual mean NDVI predicted by the model considering time-lag influences was more accurate in the central and southern Kazakhstan and Inner Mongolia regions (Figure 5b,c).

3.3. Comparing NDVI Trend Spatial Distributions between the Two Models

We analyzed the spatial trends of the grassland NDVI in Asian drylands from 2001 to 2020 using least-squares linear regression. The spatial trends of the predicted grassland NDVI considering time-lag influences were closer to that of the MODIS NDVI (Figure 6a,b) compared with the model simulations without time-lag influences (Figure 6c). From the MODIS NDVI, 54% of the grassland regions had unchanged NDVI values in Asian drylands (marked in stable), mainly located in Kazakhstan, Inner Mongolia, and Qinghai of China. A fifth (22%) of the grassland areas featured with significantly decreased NDVI values were located in northern Kazakhstan, whereas most of the areas with significant increases were located in the northeastern Republic of Mongolia, Inner Mongolia, and Qinghai (8%) (Figure 6a). The distribution of the five spatial trend types in the model output, taking into account time-lag influences, was superior to that of the model output without considering time-lag influences (Figure 6). In terms of concrete values, the grassland NDVI remained stable in 54%, 48%, and 69% of the grassland areas for (i) MODIS NDVI, (ii) predicted NDVI with time-lag, and (iii) predicted NDVI without time-lag, respectively (Figure 6a). Conversely, the grassland NDVI significantly decreased in 22%, 22%, and 20% (Figure 6b) and significantly increased in 8%, 8%, and 3% (Figure 6c) of the grassland areas in Dryland Asia for (i) MODIS NDVI, (ii) predicted NDVI with time-lag, and (iii) predicted NDVI without time-lag, respectively (Figure 6b).

3.4. Importance of the Variables from Random Forest

Figure 7a,b illustrates the relative importance of the two models with and without considering the time-lag influences by model outputs. NDVI_1 was most important for the grassland NDVI in Asian drylands from 2001 to 2020, with importance values of 24.7 and 55.0 for the two models. The second most important factor was pre_0, with importance values of 10.3 and 17.2 for the two models. Except for NDVI_1 and pre_0, sm_1 and vpd_0 were important for the grassland NDVI in Asian drylands. For the model considering time-lag influences, NDVI_1 and moisture variables (pre_0, sm_1, and vpd_0) were the major determinants for the grassland NDVI in Asian drylands, and they accounted for about 86.3% of the overall importance. In comparison, for the model without considering time-lag influences, NDVI_1, pre_0, sm_0, and rad_0 were the major determinants, which accounted for about 87.9% of the overall importance. Hence, soil moisture from the previous month played a crucial role in influencing the current month’s grassland NDVI, as depicted in Figure 7a.

4. Discussion

Machine learning methods have been demonstrated in this study to possess remarkable capabilities in predicting the grassland NDVI, consistent with the findings of numerous previous studies [19,45,46]. Additionally, several studies have suggested that lag effects have a substantial impact on vegetation growth. Miao [25] investigated the response of the grassland NDVI to soil moisture and solar radiation with a one-month time lag in Asian drylands, while Chen [47] et al. identified a strong positive relationship between soil moisture and the NDVI, with the NDVI typically lagging behind soil moisture by one month. Two models were developed utilizing the random forest algorithm, with and without considering time-lag influences, for the prediction of the grassland NDVI in Asian drylands. We highlighted the strong performance of both models in capturing the spatial and temporal variations and trends of the grassland NDVI in Asian drylands, while the model considering time-lag influences outperformed the other one. Thus, the integration of time-lag influences and machine learning presents a novel perspective for predicting the grassland NDVI and significantly enhances prediction accuracy.

Whether considering time-lag influence or not, the prediction results for the entire Asian grassland exhibit consistency. Over the period from 2001 to 2020, the grassland NDVI in Asian drylands generally exhibited a weak decreasing trend. Grasslands in eastern Asia generally demonstrate a growth trend in the NDVI, while the NDVI in western Asian dryland grasslands shows a slightly declining trend, with the most pronounced decline observed in northern Kazakhstan. We postulate that the factors driving grassland decline can be attributed to two main factors: climate change and human activities. Research has demonstrated five Central Asia countries witnessing an increase in aridity, marked by the severest drought in the northwestern part of Kazakhstan and a gradual reduction in overall precipitation [4,32,48]. This shift in climate patterns was likely a primary driver behind the declining grassland NDVI in northern Kazakhstan. Meanwhile, the degradation of grassland in northern Kazakhstan could be attributed to human activities, including intensified overgrazing, economic challenges faced by countries after the Soviet Union’s dissolution, and suboptimal government policies for grassland preservation [49]. Concerning the regions experiencing grassland NDVI growth, we attribute this phenomenon primarily to climate factors, particularly in northeastern Mongolia, Inner Mongolia, and Qinghai. The results of numerous studies in this region also align with our findings, indicating a consistent increasing trend in the NDVI [50,51]. This was primarily due to precipitation on the Mongolian plateau originating from both the westerly belt and the East Asian monsoon [52]. The northern part of the Mongolian Plateau is influenced by the westerly belt, resulting in higher precipitation in the north and lower levels in the south. The eastern part is influenced by the East Asian monsoon, providing more abundant rainfall, and both rain and warmth coincided, leading to a gradual increase in precipitation from the west to the east [52]. Consequently, the varying trends in the grassland NDVI across different regions of Asian drylands can be attributed to their distinct natural conditions. Therefore, the prediction of grassland growth by machine learning is particularly significant for understanding vegetation physiology and managing human activities. Nonetheless, the prediction of grassland growth in Asian dryland areas was not only related to climatic influences but also to human activities, vegetation physiological processes, and natural hazards. These multifaceted elements might potentially affect the accuracy of grassland growth predictions. Therefore, future endeavors aimed at forecasting grassland growth must comprehensively address the confluence of climatic influences and natural hazards.

The grassland NDVI is typically influenced by various concurrent climatic factors, including temperature and precipitation [53,54]. However, our study reveals a distinctive pattern, as we observed that soil moisture in the preceding month exerts a robust influence on the grassland NDVI, surpassing the impact of temperature on the grassland NDVI in grasslands. Our study identified specific climate variables that played a particularly crucial role in determining the grassland NDVI in the study area, including precipitation in the current month, soil moisture in the previous month, and vapor pressure deficit in the current month. Moisture is often regarded as the primary factor controlling vegetation growth in arid and semiarid areas [55]. Previous studies have commonly used precipitation as a key factor of moisture availability [56,57], but they have often overlooked the underlying mechanisms governing moisture dynamics. In arid and semiarid areas, a significant portion of surface-layer soil moisture originates from precipitation [58], while changes in soil moisture in deeper layers were determined by fluctuations in groundwater depth [59]. When deeper soil layers are well-hydrated, hydraulic redistribution can enhance the efficiency of the root system to absorb and transport water [59], i.e., the conductive movement of water from the wetter part of the soil to the drier part of the soil via the plant’s root system [59]. This, in turn, benefits plants by optimizing resource utilization and extending the availability of soil moisture [60]. This mechanism likely explains why soil moisture from the previous month is critical for the grassland NDVI in the current month. Therefore, soil moisture serves as a more accurate indicator of water availability for grasslands in arid regions. In research pertaining to grassland drought, numerous scholars have acknowledged the significance of soil water. They have conducted analyses to examine its impact on vegetation growth, establishing thresholds based on varying soil moisture levels [61,62].

Understanding the interplay between the NDVI and soil properties, climate, and other environmental factors is imperative for predicting grassland growth and productivity. While spectral indexes, notably the Normalized Difference Vegetation Index (NDVI), serve as reliable indicators of vegetation growth and productivity, the augmentation of learning machine models with additional spectral indexes presents an opportunity for a more holistic comprehension of grassland ecosystems. For instance, the incorporation of the Normalized Difference Water Index (NDWI) could elucidate soil moisture content dynamics, whereas the inclusion of the Soil Adjusted Vegetation Index (SAVI) could mitigate soil background effects [59,60,61]. By incorporating these additional indexes, we can cultivate more realistic models that encompass a broader spectrum of environmental variables. This, in turn, could lead to more accurate predictions of grassland growth and productivity across diverse conditions, such as different temperature scenarios or grazing intensities. Furthermore, the integration of soil properties, climate data, and other environmental factors into these models unveils the potential to reveal previously unrecognized patterns and relationships. For example, factors such as soil nutrients, pH levels, or the presence of specific minerals may exert a substantial influence on grassland growth dynamics. By including these factors in the models, we can attain a deeper understanding of their impacts on vegetation dynamics [5,34].

5. Conclusions

To ensure the resilient growth of grasslands in Asian drylands, understanding the impact of time-lag influences on the grassland NDVI is crucial. Employing a random forest machine learning approach and incorporating seven meteorological elements with MODIS NDVI observations, we predicted the grassland NDVI in Asian drylands, considering or not considering time-lag influences. The grassland NDVI model considering time-lag influences demonstrated relatively higher accuracy, as indicated by R², RMSE, and MAE compared with the model without time-lag influences. The annual mean NDVI predicted by the model considering time-lag influences closely aligned with the MODIS NDVI. The annual mean NDVI in Asian drylands demonstrated a weak decreasing trend from 2001 to 2020. In terms of spatial variations, the grassland NDVI remained stable in the majority of areas, significantly decreasing in northern Kazakhstan and significantly increasing in northeastern Republic of Mongolia, Inner Mongolia, and Qinghai. Notably, more areas showed a decrease in the grassland NDVI than those exhibiting an increase. The main factors affecting the grassland NDVI were NDVI in the previous month, the current month’s total precipitation, and the previous month’s soil moisture. The precise prediction of the grassland NDVI serves as a valuable scientific reference for both grass growth and livestock development in Asian drylands.