Research on dynamic prediction of vegetation coverage by precipitation-evapotranspiration in arid regions based on CNN-LSTM hybrid model Authors: Kai Peng, Yanfei Zhang, Tiejun Liu, Zijing Li, Wentao Liang, Hualin Liu, Yawen Bai Published: 2025 ABSTRACT AbstractAiming to address the challenges of phenological lag and the dominant role of water factors (precipitation and evapotranspiration) in dynamically predicting vegetation coverage in arid regions, this study proposes a hybrid CNN-LSTM-based prediction method. By employing the Standardized Precipitation Evapotranspiration Index (SPEI) to classify the study area into different arid region types, it provides diverse samples of vegetation coverage changes under various drought conditions for training the CNN-LSTM model. The model integrates precipitation-evapotranspiration data from multiple arid region types along with target prediction data. The CNN extracts spatial features of precipitation and evapotranspiration, while the LSTM captures long-term temporal dependencies, enabling dynamic vegetation coverage prediction. Furthermore, the study quantifies both the single-year and comprehensive dynamic change indices of vegetation coverage to characterize its temporal variations. Experimental results demonstrate that the model effectively identifies differences in rainfall characteristics across areas with different drought severity levels. In arid regions, the contribution rate of heavy rainfall to the total rainfall anomaly (25.15%) is substantially higher than that in semi-arid regions (1.55%), with statistical tests confirming a significant difference (P < 0.05). On the test set, the Pearson correlation coefficient between the model’s predictions and the measured data reaches 0.95, indicating strong agreement with reality. The dynamic change indices accurately capture phased variations in vegetation coverage from 2000 to 2022.Keywords: CNN-LSTM hybrid model, Standardized precipitation evapotranspiration index, Precipitation-evapotranspiration in arid regions, Vegetation coverage, Dynamic prediction, Dynamic change indexSubject terms: Environmental sciences, Environmental social sciences Introduction Arid regions have extremely fragile ecosystems, characterized by scarce precipitation and high evaporation rates. Vegetation, as a crucial component of ecosystems, serves not only as an important natural resource but also as a sensitive environmental indicator1. Fractional Vegetation Cover (FVC), which quantifies the proportion of a statistical area occupied by the vertical projection of vegetation, is a key parameter for surface cover and an essential indicator of ecological and environmental changes2. It describes the coverage of vegetation on the ground surface and serves as one of the key metrics for characterizing vegetation status and the regional ecological environment. Under the global warming trend, climatic factors in arid regions, such as rainfall patterns and evaporation rates, have exhibited significant spatiotemporal variability3. The spatiotemporal distribution of rainfall has shifted, with increased frequency and intensity of extreme rainfall events, while rising temperatures have led to higher evaporation rates, exacerbating water loss. These changes affect vegetation growth, distribution, and coverage, altering vegetation phenology and growth cycles and leading to the degradation or death of some vegetation. In arid regions, water resources are scarce, and precipitation is the primary water source, while evaporation is an important pathway for water loss. Precipitation and evaporation directly influence the water supply required for vegetation growth, while vegetation coverage affects surface evaporation and water retention. Studying the dynamic relationship between precipitation, evapotranspiration, and vegetation coverage is of great significance for understanding water balance in arid regions and facilitating rational water resource utilization, providing a scientific basis for water resource management and ecological restoration. Therefore, in-depth exploration of this dynamic relationship and achieving accurate dynamic predictions of vegetation coverage have become urgent priorities in ecological research. Previously, numerous scholars have conducted a series of studies on vegetation cover, achieving certain results in aspects such as the analysis of its influencing factors and its spatiotemporal characteristics. Hideki et al. design a new vegetation loss model that considers transmission frequency and seasonal characteristics to measure and analyze the seasonal changes in deciduous tree vegetation in Japan. When electromagnetic waves of different frequencies (such as microwaves and millimeter waves) penetrate or bypass vegetation, their signal intensity is attenuated due to the blocking, scattering, and absorption effects of vegetation. By transmitting signals of specific frequencies in different vegetation areas and under different vegetation density conditions, and then receiving and measuring the signal attenuation, vegetation cover can be estimated. In addition to vegetation, atmospheric water vapor, aerosols, and other objects on the ground may also cause scattering, absorption, or other interference to electromagnetic wave signals, affecting the accuracy of signal attenuation measurements and thereby reducing the estimation accuracy of vegetation cover4. Batnyambuu et al. study a vegetation cover estimation method combining ground vegetation data with remote sensing, utilizing the complementary integration of field measurements and remote sensing imagery to establish a quantitative relationship between vegetation cover and spectral characteristics, thereby enabling large-scale, high-precision vegetation cover estimation. Radiative transfer models (such as PROSAIL) require input parameters such as Leaf Area Index (LAI), Leaf Angle Distribution (LAD), and soil reflectance, which are difficult to obtain accurately in practical applications, often relying on empirical values or simplified assumptions, leading to discrepancies in model simulation results5. Mishra et al. study the correlation between rainfall data and vegetation loss using Geographically Weighted Regression (GWR). They use the normalized difference vegetation index (NDVI), enhanced vegetation index (EVI), and rainfall distribution from global precipitation measurement data to generate an influence map of cyclones on vegetation. However, the geographically weighted regression model assumes that spatial non-stationarity changes smoothly with spatial continuity. In reality, vegetation loss may be influenced by discrete geographic features (such as rivers, roads, and land use boundaries), resulting in abrupt spatial variations that the model struggles to capture6. Thomas et al. investigate a method for generating vegetation feature maps based on satellite hyperspectral images, leveraging the high spectral resolution of hyperspectral data to capture leaf-canopy-level spectral absorption and reflection characteristics. By applying physical models or data-driven algorithms, they infer vegetation physiological and biochemical parameters, structural traits, and coverage status, ultimately producing spatially continuous vegetation feature maps for coverage analysis. However, hyperspectral data primarily reflect instantaneous spectral properties, while seasonal vegetation changes exhibit phenological lag, making single-timepoint data inadequate for tracking dynamic processes. Multi-temporal data fusion can mitigate this issue, though spectral stability across time (e.g., sensor calibration drift) may introduce errors7. Given this context, the study focuses on arid regions, with the aim of comprehensively utilizing multi-source data and advanced analysis techniques to deeply explore the dynamic influence mechanisms of precipitation and evapotranspiration on vegetation coverage and to construct a more accurate and effective method for predicting vegetation coverage dynamics. The findings are expected to provide theoretical and decision-making support for the scientific management and sustainable development of arid region ecosystems, thereby contributing to their conservation and restoration. Dynamic prediction of vegetation coverage in arid regions Multi-type early warning level analysis based on standardized precipitation evapotranspiration index (SPEI) The Standardized Precipitation Evapotranspiration Index (SPEI) integrates precipitation and potential evapotranspiration data, providing a comprehensive assessment of regional water balance that is particularly suitable for analyzing drought characteristics under climate change conditions8. In this study, SPEI serves as the basis for drought classification prior to predicting dynamic vegetation cover changes in arid regions. Since vegetation cover status varies significantly across different drought severity zones, this classification step is essential for achieving accurate vegetation cover predictions across multiple drought zone types9. Calculation of potential evapotranspiration (PET) Potential evapotranspiration (PET) is calculated using the Hamon method (Eq. 1): In Eq. (1), represents the proportionality coefficient, a dimensionless constant used in the Hamon model that adjusts the calculation results to better match actual conditions10; denotes the day length; indicates the monthly average temperature; and represents the saturation vapor pressure variable, calculated based on monthly mean temperature to reflect temperature’s influence on vapor pressure and consequently on the evapotranspiration process. Constructing cumulative moisture deficit series and probability distribution for the study area A cumulative moisture deficit series for different time scales is constructed in the study area, with the calculation formula as follows: In Eq. (3), represents the precipitation amount; represents the number of calculations. A three-parameter log-logistic probability distribution function is introduced to calculate the probability distribution of the cumulative moisture deficit series . In this study, a time-series cross-validation strategy is adopted for SPEI drought classification to ensure threshold fairness. The specific steps are as follows: For any year T in the study period (for example, 2005), when calculating its SPEI value and classifying the drought level, only the historical data up to year T (i.e., 2000–2005) are used to construct the cumulative water deficit sequence and fit a three-parameter log-logistic probability distribution function. The drought classification threshold for that year is determined entirely based on its past climate background, thereby eliminating the use of future information. When preparing training and verification samples for the CNN-LSTM model, the SPEI drought grade label corresponding to each sample is independently calculated according to the aforementioned "rolling base period principle". This means that when the model learns the data from 2005, the drought label used is based on data from 2000–2005; when learning the data from 2015, the labels used are based on data from 2000–2015. Through this method, the drought classification at each time point strictly depends on its historical climate series, simulating the condition of "using only known information" in real prediction scenarios, thus completely avoiding potential data leakage caused by using a global threshold and ensuring the reliability and fairness of the model evaluation results. The log-logistic probability density function is expressed as: In Eq. (4), , and represent the scale parameter (affecting the extent of stretching), the shape parameter (which determines the probability density function’s characteristics including curve steepness and peak sharpness), and the location parameter (defining the probability distribution’s horizontal axis position), respectively. They are calculated as follows: In Eqs. (5)–(7), , and are intermediate variables obtained through statistical calculations of the cumulative moisture deficit series, used to determine the specific parameter values of the probability distribution function11. The three-parameter log-logistic probability distribution function is expressed as: The cumulative probability for a given time scale is calculated according to . The SPEI is calculated based on the cumulative probability of the cumulative moisture deficit series when , . In Eq. (9), , and are the constant coefficients used in the calculation process of SPEI, which are used to construct the calculation formula of SPEI and determine the specific functional relationship between SPEI and the intermediate variable . Similarly, , and are additional constant coefficients employed in the SPEI computation, working together with , and to determine the calculation logic of SPEI. When , . Based on the Standardized Precipitation Evapotranspiration Index (SPEI), the study area is classified into different drought levels (Table 1). Table 1Classification of drought and flood levels.IndexLevel ≥ 2Extremely humid[1.5, 2]Semi humid zone[1, 1.5]Mild moisture[− 1, 1]Normal[− 1.5, − 1]Mild drought[− 2, − 1.5]Semiarid ≤ − 2Drought Classification of drought and flood levels. Based on the Standardized Precipitation Evapotranspiration Index (SPEI) values, drought classification criteria are established for the study area, enabling the labeling and categorization of training samples for vegetation coverage prediction in arid regions. This process clarifies the drought level of each sample, providing clear guidance for the model to learn the dynamic changes in vegetation coverage under different drought conditions. Dynamic prediction model of vegetation coverage based on CNN-LSTM Feature extraction of environmental data such as rainfall and evapotranspiration based on CNN Convolutional Neural Networks (CNN) serve as a core algorithm of deep learning, effectively extracting spatial features from input environmental data such as rainfall and evapotranspiration12. The CNN architecture mainly consists of two parts: the feature extractor (containing input layers, convolutional layers, and pooling layers) and the classifier (with fully connected layers that nonlinearly combine extracted features to accomplish learning objectives). The analyzed precipitation and evapotranspiration data from arid regions (classified by Standardized Precipitation Evapotranspiration Index (SPEI) values in "Multi-type early warning level analysis based on standardized precipitation evapotranspiration index (SPEI)") section serve as training samples, providing reference information for the CNN-LSTM model. When extracting features from environmental data such as precipitation and evapotranspiration, both the SPEI and the precipitation and evapotranspiration data samples used for prediction are input into the CNN, uniformly set as sample . Feature Extractor Feature Extractor Convolution layer: in the convolution layer, the convolution kernels scan the feature maps of the previous layer in the form of horizontal sliding windows13. Through convolution operations, batch normalization, and nonlinear transformations, deep spatial features of are learned. During convolution calculations, weight sharing is used to limit the number of network parameters, making the connections between neurons in the convolution layers sparse, which facilitates model training and avoids getting stuck in local optima14. The spatial feature extraction process for precipitation and evapotranspiration data across varying drought intensity levels in the convolutional layer operates as follows: In Eq. (11), represents the feature map of for the -th convolution kernel in the -th convolutional layer; represents the output of the -th layer; is the weight of the -th convolution kernel in the -th layer; is the bias vector of the -th convolution kernel in the -th layer; and is the activation function, which performs nonlinear transformations on the results of the convolution calculations, enhancing the model’s expressive capability. Pooling Layer: situated between two convolutional layers, it employs max-pooling operations to reduce feature map dimensionality while maintaining translational invariance of the extracted precipitation-evapotranspiration spatial patterns (downsampling). Each precipitation-evapotranspiration feature map in the pooling layer connects to its corresponding convolutional layer feature map, with the output feature map computed as follows: In Eq. (12), and represent the weight and bias in the pooling output, respectively; denotes the max pooling operation (downsampling function). (2)Classifier Classifier The spatial features of environmental data (e.g., precipitation-evapotranspiration) extracted from the final convolutional layer are flattened into a one-dimensional vector, transforming complex spatial patterns into sequential numerical representations suitable for fully connected layer processing. In the fully connected layer, neurons establish direct connections with all neurons in both adjacent layers to derive global environmental data characteristics. The softmax function is used to achieve multi-class (MC) classification of precipitation-evapotranspiration data15, with the corresponding output as follows: In Eq. (13), and are the weight matrix and bias vector corresponding to the fully connected layer. Dynamic prediction of vegetation coverage based on LSTM Long Short-Term Memory (LSTM) networks, as an improved type of recurrent neural network, can effectively capture long-term dependencies in time series16. In the dynamic prediction of vegetation cover, the gating structure and memory units can process vegetation-related data, thereby achieving long-term preservation and utilization of historical information on vegetation cover. LSTM adds three control units for information transmission to the RNN structure, namely the forget gate, input gate, and output gate, as shown in Fig. 1. This effectively solves the gradient disappearance or explosion defects that occur in RNNs in long sequences17. The forget gate selectively retains relevant historical information while suppressing noise propagation; the input and output gates respectively regulate data incorporation and processed information flow to subsequent time steps18. The specific calculation formulas are as follows: LSTM structure for dynamic prediction of vegetation coverage. In Eqs. (14)–(19), represents the output value of the forget gate, determining which information in the previous cell state is retained (a value close to 1 indicates retention) and which information is forgotten (a value close to 0 indicates forgetting). For example, when processing precipitation-evapotranspiration spatial feature sequences, the forget gate may assign near zero weights to historical information irrelevant to current predictions, effectively excluding such data from current cell state computations. represents the output value of the input gate, controlling how much information from the current input spatial feature is written into the cell state . For instance, when new precipitation-evapotranspiration spatial feature data becomes critical for vegetation cover prediction, the input gate permits its incorporation into the cell state with elevated weighting. denotes the input node’s output computed via the hyperbolic tangent function (tanh) with a value range of [-1, 1]. This output derives from a nonlinear transformation of both the current input and the previous hidden state , generating an intermediate representation of current input information. This intermediate value subsequently combines with the input gate’s regulatory effect for integration into the cell state. represents the output value of the output gate, obtained from the sigmoid function, determining how much information from the cell state will be output to the hidden layer state . For example, during vegetation cover prediction, when critical precipitation-evapotranspiration spatial features stored in the cell state require temporal filtering, the output gate attenuates their influence through weight reduction19. and represent the weight matrix and bias vector of the forget gate. The matrices , , , , and represent the weights of the respective gates multiplied by inputs and ; , and are the bias vectors corresponding to the respective gates multiplied by inputs and ; indicates the variation of the sigmoid function . The dynamic prediction result of vegetation coverage based on the spatial features of precipitation-evapotranspiration data is given by: In Eq. (20), and represent the prediction weight and bias vector for vegetation coverage , respectively. Table 2 provides details on the classification of vegetation coverage prediction results combined with the spatial features of precipitation-evapotranspiration. Table 2Classification of vegetation coverage levels.Vegetation coverageLevel < 0.08Low degree[0.08,0.32)Medium low degree[0.32,0.4)Moderate[0.4,0.64)Medium height > 0.64Height Classification of vegetation coverage levels. During the training process, a loss function is defined to measure the difference between the predicted vegetation coverage value and the actual vegetation coverage value , with the mean squared error commonly used as the loss function: In Eq. (21), represents the number of training samples. The parameters in the model, including the weights and biases of each layer of the CNN and LSTM, are updated through the backpropagation algorithm, continuously reducing the value of the loss function to make the model’s predictions more accurate20. The CNN-LSTM hybrid model adopts a series hybrid structure. First, the spatial features of the input data are extracted by the CNN module, and then the extracted feature sequence is input into the LSTM module to capture its time-series dependencies. Finally, the predicted value of vegetation coverage is output through the fully connected layer. As a spatial feature extractor, the CNN module is specifically configured as follows: Input layer: receiving pre-treated multi-dimensional environmental data such as precipitation and evapotranspiration. Input layer: receiving pre-treated multi-dimensional environmental data such as precipitation and evapotranspiration. Convolution layer: there are two layers. The first layer uses 64 convolution kernels with the size of 3 × 3, and the second layer uses 128 convolution kernels with the size of 3 × 3. The ReLU activation function is used to introduce nonlinearity after convolution of each layer. Convolution layer: there are two layers. The first layer uses 64 convolution kernels with the size of 3 × 3, and the second layer uses 128 convolution kernels with the size of 3 × 3. The ReLU activation function is used to introduce nonlinearity after convolution of each layer. Pooling layer: each convolution layer is followed by a max pooling layer with the size of 2 × 2, to reduce the dimension of the feature map, enhance the translation invariance of features and control overfitting. Pooling layer: each convolution layer is followed by a max pooling layer with the size of 2 × 2, to reduce the dimension of the feature map, enhance the translation invariance of features and control overfitting. Dropout layer: it is added after the second pooling layer, and the discard rate is set to 0.3 to further improve the generalization ability of the model. Dropout layer: it is added after the second pooling layer, and the discard rate is set to 0.3 to further improve the generalization ability of the model. Flattening layer: flattening the pooled multi-dimensional feature map into a one-dimensional vector as the input of the LSTM module. Flattening layer: flattening the pooled multi-dimensional feature map into a one-dimensional vector as the input of the LSTM module. As a time series relation modeler, the LSTM module is specifically configured as follows: LSTM layer: it contains an LSTM layer, and the number of hidden units is set to 100. The number of units aims to balance the model complexity and the ability to capture long-term dependencies. LSTM layer: it contains an LSTM layer, and the number of hidden units is set to 100. The number of units aims to balance the model complexity and the ability to capture long-term dependencies. Dropout layer: it is added after the LSTM layer, and the discard rate is also set to 0.3. Dropout layer: it is added after the LSTM layer, and the discard rate is also set to 0.3. The output of the last LSTM hidden state is transmitted to a fully connected layer, which contains one neuron, and the predicted value of vegetation coverage is directly output using a linear activation function. The key hyperparameters in the process of model training are as follows: the optimizer uses the Adam optimizer, with its initial learning rate set to 0.001, and the batch size set to 128. The model is trained for 200 epochs, and the early stopping method is used to prevent overfitting, meaning that when the validation set loss does not decrease for 10 consecutive epochs, the training is terminated early. Dynamic quantification of vegetation coverage To further quantify the dynamic changes in vegetation coverage, a single vegetation coverage dynamic change index and a comprehensive vegetation coverage dynamic change index are introduced. Single Vegetation Coverage Dynamic Change Index Single Vegetation Coverage Dynamic Change Index The interannual variation in vegetation cover is quantified by the ratio of the difference between two consecutive years’ vegetation cover values to the previous year’s baseline value, with the calculation formula expressed as follows: In Eq. (22), and represent the vegetation coverage for the time period and the vegetation coverage for the time period , respectively. The dynamic change index of single vegetation coverage reflects the sensitivity and stability of the annual change of vegetation coverage. A continuous positive index indicates that vegetation is in a stage of recovery or growth; a sustained negative index indicates degradation or increased stress; and its fluctuation reflects the severity of ecosystem response to external environmental disturbances (such as drought and grazing). For example, a rare rainstorm or a severe spring drought may leave obvious peaks or valleys on the index. (2)Comprehensive vegetation coverage dynamic change index Comprehensive vegetation coverage dynamic change index This index considers the changes in vegetation coverage over a certain time period (e.g., years) to comprehensively measure the long-term dynamic change trend, calculated using the difference between the vegetation coverage at the end of this time period and the vegetation coverage at the starting year, with the formula: In Eq. (23), represents the vegetation coverage in the year. The comprehensive dynamic change index of vegetation coverage reveals the macro-evolution trend of the vegetation ecosystem on a long time scale (such as 5 years). The sign of the index directly indicates whether the overall trend of vegetation coverage is improving or deteriorating during the study period, and its numerical value quantifies the intensity of change. It is very important for evaluating the effect of long-term ecological restoration projects (such as returning grazing to grassland) or for judging the overall response direction of the regional ecological environment under the background of climate change. By combining short-term (single index) and long-term (comprehensive index) perspectives, the index quantification system can provide ecologists and managers with a comprehensive tool to evaluate vegetation dynamics, helping identify key turning points of ecological changes and reveal the driving mechanisms behind them. Experimental analysis Experimental setup Data sources and data sets The Gobi region constitutes a vital component of Inner Mongolia’s ecosystem, spanning most of Alxa League—including portions of Ejina Banner, Alxa Left Banner, and Alxa Right Banner—with its northern boundary extending to the China-Mongolia border. To the east, it neighbors desert areas of Bayannur City, while its southern periphery adjoins Gansu Province’s Hexi Corridor. Characterized by an arid climate with scarce precipitation and intense evaporation, this region presents severe hydrological constraints on vegetation growth, making it an ideal study area for examining precipitation-evapotranspiration impacts on vegetation cover. Accurate prediction of vegetation cover dynamics here holds significant implications for developing targeted ecological conservation strategies, rational livestock management planning, and climate change adaptation. The study area encompasses the Inner Mongolian Gobi region (Fig. 2), incorporating Ejina Banner, Wulate Rear Banner, Wulate Middle Banner, and adjacent zones. This hyper-arid climate zone supports primarily drought-resistant shrubs and herbaceous plants within a fragile ecosystem highly sensitive to precipitation-evapotranspiration variations. Gobi area in Inner Mongolia. The meteorological data source includes annual precipitation data from the China Meteorological Science Data Sharing Service Network (http://cdc.cma.gov.cn/). Annual evaporation (i.e., potential evapotranspiration) data are obtained from the Resource and Environment Data Cloud Platform of the Chinese Academy of Sciences (http://www.resdc.cn/). Daily data from 118 meteorological stations in the Inner Mongolia Autonomous Region and its surrounding areas from 2000 to 2022 are adopted. Strict quality control is carried out on the original data, including eliminating abnormal values and filling short-term missing data at individual stations using spatial interpolation (such as the inverse distance weighting method). Finally, the daily data are aggregated into monthly values, and annual values are calculated for subsequent analysis. Vegetation coverage data are obtained from the Resource and Environment Data Cloud Platform (http://www.resdc.cn/) of the Chinese Academy of Sciences, and represent a product derived from inversion based on remote sensing images such as MODIS. These data are used as the target variable (true value) for model prediction. Land use data also come from the Resource and Environment Data Cloud Platform of the Chinese Academy of Sciences, and are used to help understand the underlying surface conditions of the study area. Rainfall evaporation data is a comprehensive concept, which mainly refers to two key water balance factors: precipitation and potential evapotranspiration. These factors serve as the fundamental variables for calculating the Standardized Precipitation Evapotranspiration Index (SPEI) and are directly input into the forecasting model. After rigorous preprocessing and spatialization, the precipitation and evapotranspiration data, as the core driving factors, are transformed into accurate predictions of vegetation coverage through the CNN-LSTM hybrid model. To train and evaluate the CNN-LSTM hybrid model, all the data from 2000 to 2022 are randomly divided into three parts according to the ratio of approximately 7:1:2: training set, validation set, and test set. The training set is used for directly training the model and learning parameters; the validation set is used to adjust the model hyperparameters and monitor the training process to prevent overfitting; after the model training and tuning, the test set is used to evaluate the generalization performance of the model finally and independently. Role and processing of precipitation and evapotranspiration data In this study, rainfall evaporation data serve as the core input features driving the CNN-LSTM forecasting model. Since the CNN requires regular grid‑like spatial data as input, whereas the original meteorological data are collected from discrete stations, the data are spatialized. The term "rainfall evaporation data" here primarily refers to two key water‑balance factors: precipitation and potential evapotranspiration. To convert point‑based observations from 118 meteorological stations into spatial grid data suitable for the CNN, the Kriging interpolation method is adopted. This method fully accounts for spatial autocorrelation and geographical structure in meteorological variables, providing unbiased and optimal spatial estimates. In practice, precipitation and potential evapotranspiration data for each time step are interpolated separately to generate continuous spatial grid maps covering the entire study area. The interpolation is performed in an ArcGIS/Python environment, and the resolution adopted matches that of the vegetation cover remote sensing products used, ensuring that all data layers are perfectly aligned and consistent in space. After interpolation, the precipitation and potential evapotranspiration at each time step form spatial distribution maps. These two raster layers are treated as independent channels, which together constitute a multi‑channel spatial data cube that serves as input to the CNN at the corresponding time step. For each geographical grid unit (pixel), time series of precipitation and potential evapotranspiration from 2000 to 2022 are extracted, and a sliding‑window sample format acceptable to the model is constructed. Role in the model: (1) In the CNN: The multi-channel spatial data cube constructed previously is fed into the CNN. The convolutional kernels of the CNN automatically learn and extract the spatial distribution patterns, spatial synergy characteristics, and local spatial structures of precipitation and potential evapotranspiration by sliding over the grid data. (2) In the LSTM: The spatial feature sequence for each time step, extracted by the CNN, is then input into the LSTM to capture the time-lag effects and long-term dependencies of water conditions on vegetation. In summary, the site data are successfully converted into grid data suitable for the CNN through Kriging spatial interpolation, which establishes the prerequisite for the CNN to extract spatial features. After rigorous preprocessing and spatialization, the precipitation and evapotranspiration data, serving as the core driving factors, are transformed into accurate predictions of vegetation coverage via the CNN-LSTM hybrid model. Experimental data analysis Analysis of arid region classification based on standardized precipitation evapotranspiration index (SPEI) In this experiment, the method presented in this paper utilizes the Standardized Precipitation Evapotranspiration Index (SPEI) to classify the study area into a distribution map of different arid regions, as shown in Fig. 3. Spatial distribution of arid regions classified by the standardized precipitation evapotranspiration index (SPEI). As shown in Fig. 3, different drought-prone areas are delineated using the Standardized Precipitation Evapotranspiration Index (SPEI). The SPEI incorporates both precipitation and potential evapotranspiration, thereby reflecting the moisture conditions of the study area. Based on an analysis of this index, the study area is categorized into different climate zones, which aids in understanding the spatial climatic differences in Inner Mongolia. From a spatial distribution perspective, arid regions dominate most of the western part, semi-arid regions are distributed across the central and parts of the eastern areas, while semi-humid areas are concentrated along the eastern fringe. This zoning clearly illustrates the spatial variation in moisture conditions across Inner Mongolia. The results, derived from the SPEI—which comprehensively accounts for both precipitation and potential evapotranspiration—accurately reflect the region’s actual water balance. Analysis of precipitation and evapotranspiration feature extraction results The method proposed in this paper employs CNNs to extract precipitation and evapotranspiration data from arid and semi-arid regions as training samples and to capture their spatial characteristics. For rainfall, the precipitation anomaly percentage is used to characterize its properties, as it effectively reflects both the degree of deviation of rainfall in the study area from the long-term average and whether rainfall in a specific period is excessive or insufficient relative to that average. In this study, the "long-term average" used for calculating this percentage is based on the mean values from the study period of 2000 to 2022. The classification of different rainfall grades follows the standards established by the China Meteorological Administration: light rain (0.1–9.9 mm/day), moderate rain (10.0–24.9 mm/day), and heavy rain (25.0–49.9 mm/day). The denominator for all calculations of percentage deviation in Table 3 is the long-term average of the corresponding variable during the reference period from 2000 to 2022. Specifically: Table 3Rainfall anomaly percentage.PartitionSemiarid /%Drought /%Total rainfall2.654.55Small rainfall0.950.35Moderate rainfall5.25 − 1.05Heavy rainfall1.5525.15Total days2.05 − 0.35Light rain day1.75 − 0.75Moderate rain day4.35 − 0.35Heavy rain day3.8519.35Total strength1.355.55Small intensity − 0.451.75Medial0.85 − 1.15High strength − 1.8514.75 Rainfall anomaly percentage. In Eqs. (24, 25), represents the percentage of "total rainfall/total days/total intensity". Here, denotes the total value for the current period, and denotes the long-term average of the total value over the base period. Meanwhile, represents the percentage of "rainfall/days/intensity" for each grade; denotes the grade value for the current period, and denotes the long-term average value of that grade over the reference period. The results are presented in Table 3. According to the data in Table 3, the total rainfall anomaly percentage in semi-arid regions is 2.65%, with light, moderate, and heavy rainfall contributing 0.95%, 5.25%, and 1.55%, respectively. Calculating the contribution rate of each rainfall grade to the total rainfall anomaly (i.e., the rainfall anomaly value of a specific grade divided by the total rainfall anomaly value, multiplied by 100%) reveals that the contribution of moderate rainfall is prominent, while the total rainfall days also exhibit an anomaly. The total intensity anomaly percentage is 1.35%, with moderate-intensity rainfall being the main positive contributor. In contrast, the total rainfall anomaly percentage in arid regions is 4.55%, within which the anomaly percentage for heavy rainfall is 25.15%, and its contribution rate to the total rainfall anomaly is absolutely dominant, whereas the average rainfall anomaly is -1.05%. The total rainfall days anomaly is -0.35%, indicating a decrease in light and moderate rain days and an increase in heavy rain days by 19.35%. The total intensity anomaly percentage is 5.55%, with heavy rainfall contributing significantly. A statistically significant difference exists between the two groups (p < 0.05). Overall analysis indicates that the distribution of rainfall characteristics in semi-arid regions is relatively balanced across grades, whereas the rainfall increment in arid regions is significantly concentrated in high-intensity heavy rain events, reflecting a more extreme pattern. Overall analysis indicates that rainfall characteristics in semi-arid regions are relatively balanced, whereas rainfall in arid regions is concentrated in heavy rainfall events characterized by extreme intensity. The spatial distribution characteristics of potential evapotranspiration in the Gobi border region of Inner Mongolia, as extracted using the CNN, are presented in Fig. 4. Spatial distribution characteristics of potential evapotranspiration. Analysis of Fig. 4 reveals that potential evapotranspiration exhibits distinct spatial differentiation, with annual potential evapotranspiration values varying significantly across regions. The southwestern region contains a high-value zone where annual potential evapotranspiration exceeds 1,400 mm, which may be influenced by arid conditions, strong solar radiation, and high wind speeds that promote intense evapotranspiration. In contrast, the northeastern region features low-value zones with annual potential evapotranspiration below 700 mm, likely due to relatively high precipitation, elevated humidity, and lower temperatures, which collectively constrain potential evapotranspiration. From northeast to southwest, the potential evapotranspiration generally increases, reflecting the combined influence of spatial variations in regional climate conditions on potential evapotranspiration. In summary, the method proposed in this study employs a CNN to extract rainfall characteristics—including rainfall amount, rainfall days, and rainfall intensity—thereby providing a comprehensive and detailed characterization of rainfall patterns across different climate zones and offering substantial information for a deeper understanding of regional rainfall regimes. The method also effectively extracts the spatial distribution characteristics of potential evapotranspiration in the Gobi Desert border region of Inner Mongolia, accurately capturing the significant regional variations in annual potential evapotranspiration and clearly revealing its increasing trend from the northeast to the southwest. Analysis of dynamic prediction performance for vegetation coverage After confirming the method’s capability to extract precipitation-evapotranspiration spatial characteristics in arid regions, the measured vegetation coverage data of the study area from 2000 to 2022, shown in Fig. 5, are used as test samples to further evaluate the method’s ability to predict vegetation coverage using rainfall-evaporation spatial characteristics. The proposed method utilizes a CNN to extract precipitation-evapotranspiration spatial features, which are subsequently fed into an LSTM to predict dynamic vegetation coverage changes in the study area; the prediction results are presented in Fig. 6. Measured vegetation coverage data for model testing (2000–2022). Prediction results of vegetation coverage in the study area. Figure 5 presents the measured vegetation coverage data used for model verification, spanning from 2000 to 2022 and covering the Gobi area of Inner Mongolia (identical to Fig. 2). The data originate from a resources and environment cloud platform provided by the Chinese Academy of Sciences, representing a vegetation coverage product derived from remote sensing image inversion; these data serve as the basis for verifying the accuracy of the prediction results of the proposed method. Analysis of Fig. 6 indicates that between 2000 and 2022, the average vegetation coverage in the study area fluctuates. Although variations occur during this period, an overall upward trend is observed, reflecting an improvement in vegetation coverage over the study duration. The specific vegetation coverage prediction results are as follows: Initial fluctuation (2000–2008): the average vegetation coverage in 2000 is at a relatively low level, followed by fluctuating increases and decreases in subsequent years, indicating that vegetation coverage during this stage is influenced by multiple factors and has not yet formed a stable trend. Growth phase (2008–2010): the average vegetation coverage shows a significant increase, likely resulting from relatively favorable climatic conditions (particularly precipitation and evaporation patterns) during this period or the successful implementation of ecological conservation measures, both of which contributes to enhanced vegetation growth. Second fluctuation (2010–2016): during this period, vegetation coverage fluctuates multiple times, reaching a relative low point around 2014 before rapidly rebounding. This suggests that vegetation coverage is subject to combined influences of natural factors (such as extreme weather events) and human factors (such as land use changes), resulting in limited stability. Later high-level fluctuations (2016–2022): the average vegetation coverage remains at a relatively high level with minor fluctuations, indicating that although the regional vegetation ecosystem generally improves during the later period, it still experiences certain external disturbances that prevent it from maintaining a consistently high coverage state. The red squares in the figure correspond to changes in vegetation coverage grades and average vegetation coverage. In years with lower average vegetation coverage, such as 2000–2006, vegetation coverage grades are predominantly at medium-to-low levels. In contrast, during years with higher average vegetation coverage, such as 2010–2012 and 2016–2022, vegetation coverage grades are mostly at medium-to-high and high levels. This indicates that changes in the quantity of vegetation coverage in the study area are associated with improvements or declines in its quality grades, collectively reflecting the evolution of the regional vegetation ecological condition. The methods employed in this study further quantify the vegetation coverage data by randomly selecting the single vegetation coverage dynamic change index for the years 2017–2021 and the comprehensive vegetation coverage dynamic change index for the years 2010–2021. The quantitative results are presented in Tables 4 and 5. The comprehensive vegetation coverage dynamic change index considers changes in vegetation coverage over five-year intervals to comprehensively assess long-term dynamic trends. Table 4Single-year vegetation coverage dynamic change index.YearDynamic change index of single vegetation coverageCorrelation of prediction results20170.0470.9520180.0220.9520190.0220.952020 − 0.0640.952021 − 0.0230.95 Single-year vegetation coverage dynamic change index. Table 5Comprehensive vegetation coverage dynamic change index.Start yearEnd yearDynamic change index of comprehensive vegetation coverageCorrelation of prediction results20102014 − 0.0680.95201120150.1190.95201220160.1190.95201320170.0980.95201420180.1460.952015201900.9520162020 − 0.0430.952017202100.95 Comprehensive vegetation coverage dynamic change index. Analysis of the data in Table 4 shows that the dynamic change index of single-year vegetation coverage exhibits fluctuating characteristics between 2017 and 2021. In 2017, the index is 0.047, indicating a certain degree of increase in vegetation coverage compared to the previous year. In 2018 and 2019, the index values are relatively small, at 0.022 and 0.022 respectively, indicating only slight increases in vegetation coverage during these two years and relatively stable changes. In 2020, the index is − 0.064, and in 2021, it is − 0.023, indicating decreases in vegetation coverage in these two years, with a relatively larger decrease in 2020. Furthermore, the Pearson correlation coefficient between the predicted values of the single-year vegetation coverage dynamic change index and the actual dynamic changes in vegetation coverage reaches 0.95, indicating a high match between the predicted results and the actual situation. Analysis of the data in Table 5 reveals that, based on the comprehensive vegetation coverage dynamic change index calculated for different start and end years, distinct phases of significant change are evident. The index for 2010–2014 is − 0.068, indicating a declining trend in vegetation coverage during this period. In contrast, the indices for 2011–2015 and 2012–2016 are both 0.119, suggesting a noticeable increase in vegetation coverage over these intervals. The index for 2013–2017 is 0.098, and for 2014–2018, it is 0.146, indicating continued yet slower growth in vegetation coverage. The indices for 2015–2019 and 2017–2021 are 0, while the index for 2016–2020 is -0.043, reflecting relative stability in vegetation coverage during certain periods and a slight decline in others. Furthermore, the Pearson correlation coefficient for the predicted values of the comprehensive vegetation coverage dynamic change index also reaches 0.95, which further confirms the accuracy of the prediction results. The single-year and comprehensive vegetation coverage dynamic change indices are mutually complementary: the former captures short-term interannual fluctuations, while the latter reflects long-term overall trends. A combined analysis of these indices enables a more comprehensive understanding of vegetation coverage dynamics in the study area, providing critical data support and decision-making references for ecological environment assessment, vegetation ecosystem management, and the development of scientifically based ecological protection and restoration strategies. To validate the superiority of the CNN-LSTM hybrid model proposed in this paper, it is compared with several benchmark models on the same test set. The selected benchmark models include: a multiple linear regression model, representing classical statistical methods; a support vector machine regression model, representing traditional machine learning methods; a single CNN model, used to verify the necessity of introducing the LSTM module for capturing temporal dependencies; and a single LSTM model, used to verify the effectiveness of introducing the CNN module for extracting spatial features. All models use the same training and test sets. The root mean square error (RMSE) and Pearson correlation coefficient (R) are employed as performance evaluation metrics, and the comparison results are presented in Table 6. Table 6Prediction performance comparison of different models.ModelsRoot-mean-square errorPearson correlation coefficientsMultiple linear regression0.0890.82Support vector machine regression0.0750.87Single CNN model0.0710.88Single LSTM model0.0650.91CNN-LSTM model in this paper0.0420.95 Prediction performance comparison of different models. From the analysis of Table 6, it can be observed that the CNN-LSTM hybrid model proposed in this paper achieves the best performance in both evaluation metrics, with the lowest RMSE (0.042) and the highest correlation coefficient (0.95), demonstrating significant superiority over other benchmark models and fully validating the effectiveness and advancement of integrating CNN’s spatial feature extraction capability with LSTM’s time-series modeling ability. The performance of the single CNN model and the single LSTM model is better than that of traditional machine learning methods, indicating the potential of deep learning in such complex spatio-temporal prediction problems. The CNN-LSTM model outperforms each individual model: compared with the single CNN model, the RMSE of the hybrid model is reduced by approximately 40%, highlighting the importance of incorporating LSTM to capture temporal dynamics; compared with the single LSTM model, the RMSE of the hybrid model is reduced by approximately 35%, confirming the effectiveness of introducing CNN to extract spatial features. The comparative experimental results demonstrate that the hybrid model framework proposed in this paper can fully exploit the spatio-temporal information of environmental factors such as precipitation and evapotranspiration, thereby enabling more accurate dynamic prediction of vegetation coverage. Conclusion Arid regions constitute a substantial portion of Earth’s terrestrial surface, characterized by fragile ecosystems where vegetation coverage serves as a crucial indicator of ecological quality. Vegetation significantly influences regional climate patterns, soil conservation, and biodiversity maintenance while playing a pivotal role in water cycle processes. The accurate prediction of vegetation coverage dynamics is essential for comprehending ecosystem structures and functions in arid regions, as well as for developing scientifically sound ecological conservation and restoration strategies. This study addresses the inherent complexities in predicting vegetation coverage dynamics across arid regions by proposing an innovative CNN-LSTM hybrid modeling approach, with the following key research outcomes: A hybrid prediction framework integrating SPEI-based arid region classification with CNN-LSTM is developed, utilizing CNN to extract spatial patterns of precipitation and evapotranspiration while employing LSTM to capture temporal dependencies, thereby overcoming traditional methods’ limitations in modeling phenological lags and multi-factor interactions. The established dynamic change index system enables precise quantification of vegetation coverage variations, with both single-year and composite indices demonstrating a 0.95 correlation coefficient with observed vegetation dynamics, accurately reproducing the temporal evolution of vegetation coverage from 2000 to 2022. Although the CNN-LSTM hybrid model proposed in this study demonstrates strong performance in dynamically predicting vegetation coverage in the Gobi region of Inner Mongolia, several limitations remain that should be addressed in future work: The model construction and validation are based on the specific environmental conditions of the Gobi region in Inner Mongolia. Although the methodology is generalizable, its direct applicability and predictive accuracy in other types of arid regions (such as the Sahel region in Africa or desert areas in Central Asia) require verification with more extensive geographical datasets. The model construction and validation are based on the specific environmental conditions of the Gobi region in Inner Mongolia. Although the methodology is generalizable, its direct applicability and predictive accuracy in other types of arid regions (such as the Sahel region in Africa or desert areas in Central Asia) require verification with more extensive geographical datasets. The current model primarily considers two core climatic driving factors—precipitation and evapotranspiration. However, vegetation dynamics are also significantly influenced by soil texture, groundwater, human activities (such as grazing and afforestation), and more detailed topographic factors. Insufficient consideration of these elements may affect prediction accuracy in certain areas. The current model primarily considers two core climatic driving factors—precipitation and evapotranspiration. However, vegetation dynamics are also significantly influenced by soil texture, groundwater, human activities (such as grazing and afforestation), and more detailed topographic factors. Insufficient consideration of these elements may affect prediction accuracy in certain areas. The meteorological data and remote sensing products used in this study are subject to certain spatial resolution limitations. Higher-resolution data (such as satellite-based downscaling data and UAV remote sensing data) could help reveal vegetation-water relationships at a finer scale, thereby improving model performance in highly heterogeneous areas. The meteorological data and remote sensing products used in this study are subject to certain spatial resolution limitations. Higher-resolution data (such as satellite-based downscaling data and UAV remote sensing data) could help reveal vegetation-water relationships at a finer scale, thereby improving model performance in highly heterogeneous areas. Despite the aforementioned limitations, the methods and findings of this study retain important universal value and scientific significance: this study successfully validates the effectiveness of the CNN-LSTM hybrid model framework in addressing the "climate-vegetation" coupling prediction problem. This technical framework itself is transferable and can be adapted to other regions or similar ecological prediction tasks, such as predicting soil moisture and net primary productivity. The study reveals that in extremely arid regions, high-intensity, low-frequency rainfall events represent key hydrological processes driving vegetation changes. This finding underscores that, against the background of increasing extreme precipitation events due to future climate change, understanding the ecological effects of such pulse precipitation events is crucial for predicting the evolution of dryland ecosystems. As simple yet effective quantitative tools, the proposed single-year and comprehensive vegetation coverage dynamic change indices can be widely applied in long-term dynamic monitoring and evaluation of vegetation in other regions, assisting ecological managers in promptly identifying key change stages and long-term trends. Building upon the achievements and limitations of this study, future work will focus on the following aspects: Integrating multi-source data—such as soil properties, topographic indices, and even human activity indices—into the model to construct a more comprehensive prediction framework; Integrating multi-source data—such as soil properties, topographic indices, and even human activity indices—into the model to construct a more comprehensive prediction framework; Utilizing data with higher temporal and spatial resolution for modeling and exploring the model’s transferability across different geographical units; Utilizing data with higher temporal and spatial resolution for modeling and exploring the model’s transferability across different geographical units; Developing a near real-time forecasting system by integrating it with satellite data streams such as MODIS, thereby providing real-time decision support for dynamic management and climate adaptation of arid region ecosystems. Developing a near real-time forecasting system by integrating it with satellite data streams such as MODIS, thereby providing real-time decision support for dynamic management and climate adaptation of arid region ecosystems. Author contributions K.P.: Conceptualization, Methodology, Investigation, Formal analysis, Writing—Original Draft, Writing—Review & Editing, Funding acquisition, Project administration, Supervision. Y.Z.: Methodology, Investigation, Formal analysis, Writing—Original Draft, Writing—Review & Editing, Validation. T.L.: Investigation, Data curation, Formal analysis, Writing—Original Draft, Visualization. Z.L.: Resources, Data curation, Software, Validation. W.L.: Resources, Validation, Writing—Review & Editing. H.L.: Writing—Review & Editing, Formal analysis. Y.B.: Writing—Review & Editing, Supervision. Funding The study was supported by “China Institute of Water Resources and Hydropower Research Basic Research Business Fund Special Project (Grant No. MKSZ2024JK014), the National Natural Science Foundation of China (U24A20573)”. Data availability Data is provided within the manuscript. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement Data is provided within the manuscript.