Mapping desert shrub aboveground biomass in the Junggar Basin, Xinjiang, China using quantile regression forest (QRF)

Research area

This investigation focused on the western sector of the Junggar Basin in Xinjiang, China (Fig. 1A), specifically the ancient lake basin sediment area formed after the contraction of Manas Lake (Fig. 1B). This area hosts the largest Haloxylon ammodendron forest in Xinjiang (Fig. 1C). This region receives less than 80 mm of annual precipitation, with an annual evaporation rate exceeding 3,600 mm. The average temperature in July exceeds 27 °C, with absolute maximum temperatures surpassing 44 °C, while the average temperature in January falls below −19 °C, with absolute minimum temperatures dropping below −43 °C. The average elevation is around 300 m  (Yin, 1993). The terrain is relatively flat with a few small sand dunes less than 10 m high. The area’s vegetation is mainly composed of pure Haloxylon ammodendron forests (Fig. 1D). Over the past half-century, human activities and climate change have caused significant damage to the “desert forest”, which has yet to fully recover. The study area is characterized by high salinity soils, sparse Haloxylon ammodendron stands with an average coverage of 11.58%, and very few seedlings, indicating a declining population. In such an arid environment, the overall low coverage of Haloxylon vegetation and the high proportion of dead plants within the community (Liu et al., 2010; Si et al., 2011) present conservation challenges. To protect this Haloxylon ammodendron forest, a national public welfare forest reserve was established in 2006. However, there is a lack of corresponding biomass survey data in the national forest resource inventories of China. Studying the biomass of the Haloxylon ammodendron forest in this area is crucial for its conservation and for understanding the trends in natural environmental changes as well as the impact of human activities in this region.

Research data

Methods

This research used data from UAV-LiDAR field surveys, as well as Sentinel-1 and Sentinel-2 imagery, which included SAR backscatter coefficients, texture characteristics, spectral bands, and vegetation index attributes. Employing the quantile regression forest (QRF) prediction method, the Haloxylon ammodendron AGB in the western Junggar Basin was estimated to assess the model’s reliability and the level of uncertainty in the results. The workflow for the study is depicted in Fig. 2 and is expanded upon in the following sections.

Resampling strategy

This study used hexagonal grids as the units for AGB statistics and modeling. Hexagons are the shapes closest to circles that can tessellate a plane regularly. Compared to squares, they possess additional symmetry, which can reduce sampling bias caused by edge effects of grid shapes (Birch, Oom & Beecham, 2007; Birch, Vuichard & Werkman, 2000). Hexagonal grids have also been used in AGB survey research (Bruening et al., 2023).

The determination of grid cell size was influenced by several factors, the most significant being a noticeable misalignment in the registration of data between Sentinel-1 and Sentinel-2 images (Ye et al., 2021); Secondly, there are spatial coordinate registration errors between the UAV ground sampling area and images from Sentinel-1 and Sentinel-2. Expanding the spatial coverage of sample units can mitigate the effects of common registration errors on model precision (Hogland & Affleck, 2019). This refinement in granularity can also minimize variations arising from spatial mismatches between the sampling data and Sentinel data at the sub-pixel level (Shafeian, Fassnacht & Latifi, 2021). As the sampling grid radius expands, larger plots help maintain a heightened level of spatial correspondence between the ground reference datasets and the the remote sensing datasets, potentially improving the AGB model’s performance (Frazer et al., 2011). In this investigation, hexagonal grid cell sizes were tested from 100 square meters to 800 square meters. AGB estimates obtained through random forest prediction were compared to reference AGB using R² analysis. Figure 3 presents violin plots that depict the model performance metrics obtained during the iterative evaluation of the RF model for AGB estimation. The findings reveal that R² values vary from the highest resolution of 300 square meters to 800 square meters, with the highest mean R² at 600 square meters. However, the R² values then slowly stabilizes and remains roughly constant. Considering that finer spatial granularity is deemed beneficial, a grid cell size of 600 square meters was used for modeling.

Resample result

Using ArcPy, approximately 120,000 hexagonal grid cells were continuous generated within the boundary of the UAV sampling area, each with an area of 600 square meters. The height, density, and coverage data of Haloxylon ammodendron in each unit was calculated, as well as the biomass of Haloxylon ammodendron in each unit according to the allometric growth equation (Eq. (1)). Next, from the “Original Dataset”, the median values of 98 features corresponding to each hexagonal grid cell were extracted to reduce the impact of outliers or noise. The “AGB DataSet” is constructed using the following data from all grids: the biomass of Haloxylon ammodendron, the 98-features data, and the center point coordinates (xCentroid and yCentroid).

The statistical grid aggregates data from all units to obtain the distribution percentages for cover (Fig. 4A), density (Fig. 4B), height (Fig. 4C), and biomass (Fig. 4D).

Statistical analysis revealed that the indicator values for most sampling units were generally low. For example, the percentage of units with less than 10% cover was 94.47%, the percentage with a density fewer than 20 plants was 93.22%, the percentage with an average height less than two m was 94.16%, and the percentage with an AGB less than 50 kg was 95.76%. This indicated that the basic nature of the ecosystem distribution of Haloxylon ammodendron in the area under examination had an overall low plant density, predominantly sparse vegetation, and only isolated areas of high density.

CV results

Fifty percent of the data was randomly sampled from the “AGB DataSet”, and the createFolds function from the R CAST package was then used to generate a 10-fold random CV dataset (Fig. 5A). These sampling points were relatively scattered, without any obvious spatial clustering patterns or regularities. For comparison, the kNNDM function from the R CAST package used the xCentroid and yCentroid of the “AGB DataSet” to create a 10-fold spatial cross-validation dataset (Fig. 5B). In comparison with the results of random sampling, the distribution of these sampling points exhibited a certain spatial structure or pattern.

The GEODIST function was used to calculate the nearest neighbor distances between prediction locations and training locations under different sampling scenarios to further assess the reliability of predictions. The calculations performed by the GEODIST function included the assessment of the nearest neighbor distances between prediction and training points (prediction-to-sample), the distances between test and training points in a leave-one-out cross-validation (LOO CV) scenario (sample-to-sample), and the distances between test and training points in a specific CV setup (CV-distances). In an ideal scenario, the distances observed throughout CV would closely mirror those seen in the prediction phase for the predictive conditions to be accurately reflected in the cross-validation process.

Under random sampling conditions, the distance between training points in terms of geographical nearest neighbor distance (NND) is significantly shorter than the distance between prediction points and training points (Fig. 5C). In this instance, due to the aggregation of sample locations, the distance between predictions and samples becomes significantly greater than that observed in cross-validation, resulting in notable overfitting of the model. This results in strong training performance with random sampling, but poor performance when predicting in unknown areas (Domingos, 2012). The model’s predictive capability is limited by random CV within the boundaries of the training data. Although the geographic distance distributions between test points and training points and between prediction points and training points do not perfectly match under spatial CV (Fig. 5D), there is a considerable overlap between them. This indicated that the prediction results from spatial CV were closer to real-world values.

Results of feature selection

Fifty percent of the data was randomly sampled from the “AGB DataSet”, and the kNNDM function was used to perform a 10-fold spatial cross-operation to obtain 10 datasets. Next, nine subsets of each dataset were selected as the training set, and the remaining one subset was used as the validation set. The Boruta algorithm was applied to evaluate the feature importance of the data, with the number of iterations set to 150 and the confidence level set to 0.01. The importance scores of different features in multiple iterations were calculated using the median. The feature importance scores of multiple folds were integrated and averaged, and the feature importance was sorted. Significant features were identified based on their mean significance levels. With the exception of seven GLCM indices identified in winter, all other 91 features were associated with AGB. Among the various vegetation indices, the NDTI index was identified as the most crucial, succeeded by RSR, EVI, MSAVI, and NDVI. The importance of spectral bands ranked from highest to lowest was B12, B11, B2, and B3 reflectance (Fig. 6A). Elevation also significantly influenced the AGB of Haloxylon ammodendron forest in the study area, with VH having a greater effect than VV. Among the GLCM texture factors, contrast, inverse difference moment, and sum of average were notably important. When accumulating the importance of seasonal factors, autumn factors had the highest cumulative importance, and although winter factors were fewer in number, they ranked high in importance (Fig. 6B).

To further enhance the interpretability of the model while reducing its size and computational load, the top 30 variables from the Boruta importance scores were selected. Highly correlated variables (r >  ± 0.75) were identified, and the less important factors were excluded, resulting in 12 predictor variables (Fig. 6C). These 12 selected predictive variables and the corresponding biomass data were combined into the “Predictor DataSet”.

Results of DI and AOA

Continuous hexagonal grids of 600 square meters were generated for the entire research area and 12 features of each grid from the “Original Dataset” were calculated to constitute the “Predicted Dataset”. Using the AOA function in the CAST package, a dissimilarity index (DI) was calculated based on distances to the training data in the multidimensional predictor variable space. If a certain predictive data point had similar attributes to the training data, its distance in the predictor variable space would be relatively low (DI approaching 0), while the DI at locations with a significant difference in attributes would be relatively high. The computation of a threshold value determined the area of applicability (AOA). The dissimilarity measure of a data point was calculated by dividing its minimum distance to the training data by the average distance. The final threshold value was determined by multiplying the IQR by 1.5 and adding this amount to the 75th percentile value. This specific threshold served as a benchmark for determining if new data points aligned with the training data, ultimately establishing whether they fell within the model’s AOA. The binary map generated by the AOA analysis distinguished regions as either within the AOA (designated as ‘1’), where model predictions are reliable, or outside the AOA (designated as ‘0’), where model predictions become uncertain, as illustrated in Fig. 7.

The AOA analysis indicated that approximately 35% of the study area differed from the training samples. This included regions such as the southern oasis, the downstream wetland of the Manas River (Mayi Lake), the Dabasan Nuur Salt Lake, the Small Salt Lake and its surrounding areas, the northwestern piedmont Gobi, and parts of the eastern Gurbantünggüt Desert. In these regions, land cover significantly deviated from the sampled areas, thus falling outside the model’s applicability. The remaining 65% of the region was suitable for model prediction.

AGB uncertainty map

Fifty percent of the data was randomly sampled from the “Predictor DataSet”, and 10-fold data was generated using both random and spatial CV methods, which served as the training control data for the quantregForest algorithm. The evaluation metric was set as R-squared to evaluate performance. As a result, AGB Models were constructed using both random and spatial CV methods. The criteria chosen for assessment included coefficient of determination (R²; Eqs. (2) and (3)), root mean squared error (RMSE; Eq. (4)), mean absolute error (MAE; Eq. (5)), symmetric mean absolute percentage error (SMAPE; Eq. (6)), and relative absolute error (RAE; Eq. (7)) (Chicco, Warrens & Jurman, 2021; Correndo, 2024). The calculation methods for each metric were as follows: (In these equations, the predicted ith value was denoted as X_i, while the actual ith value was represented by the Y_i element).

(2)${\displaystyle \overline{Y}=\frac{1}{m}\sum _{i=1}^m{Y}_i}$ (3)${\displaystyle R^2=1-\frac{\sum _{i=1}^m{\left(X_i-Y_i\right)}^2}{\sum _{i=1}^m{\left(\overline{Y}-Y_i\right)}^2}}$ (4)${\displaystyle \text{RMSE}=\sqrt{\frac{1}{m}\sum _{i=1}^m{\left(X_i-Y_i\right)}^2}}$ (5)${\displaystyle \mathrm{MAE}=\frac{1}{m}\sum _{i=1}^m|X_i-Y_i|}$ (6)${\displaystyle \text{SMAPE}=\frac{100\text{\%}}{m}\sum _{i=1}^m\frac{|X_i-Y_i|}{\left(\left|X_i\right|+\left|Y_i\right|\right)/2}}$ (7)${\displaystyle \mathrm{RAE}=\frac{\sum |X_i-Y_i|}{\sum |Y_i-\overline{Y}|}.}$

The range of positive coefficient of determination (R²) values is from 0–1, where a value of 1 indicates an accurate prediction. SMAPE (%) is standardized, dimensionless, and bounded, with values ranging from 0–2, where 0 represents a perfect prediction and 2 represents the worst possible prediction. This evaluation metric, RAE, involves a calculation that compares the absolute errors of predictions to the absolute deviations of observations from their mean, and a RAE of < 1 indicates that the model’s predictive ability is better than the baseline model. MAE and RMSE are suitable for comparison across different regression models, while R², SMAPE, and RAE can be used to evaluate regression quality in absolute terms (Chicco, Warrens & Jurman, 2021).

Although the random CV model results appeared ideal based on the metrics (Fig. 8A), this approach resulted in the concealment of overfitting and generated an excessively positive evaluation of predictive capability. In contrast, the spatial CV model focused on anticipating new clusters within the feature space in each round of training, which led to a reduction in spatial dependence between training and testing observations, resulting in the spatial CV model demonstrating significantly higher MAE, RMSE, RAE, and SMAPE, along with a notably lower R² compared to the random CV model (Fig. 8B). Nonetheless, the spatial CV model approach aids in averting over-extrapolation and offers a more precise evaluation of the model’s dependability in authentic scenarios. Therefore, despite the spatial CV metrics appearing inferior to those of random CV, the results these metrics provide are more valuable for reference.

The data in the “Predicted Dataset” with an AOA value of 1 were added into the AGB Model, which was developed using the spatial CV method, in order to predict the spatial distribution of Haloxylon ammodendron forest AGB throughout the entire study area (Fig. 9A).The results showed that the biomass of Haloxylon ammodendron forest exhibited significant spatial heterogeneity within the study area, with a trend of increasing from northwest to southeast. Higher biomass concentrations were discovered in the southeastern section of the study area, near the Gurbantünggüt Desert boundary, as well as in the central region, while the western and northern parts showed relatively lower biomass levels. Certain areas in the southern part also exhibited high biomass concentrations. The range of values predicted by the QRF model’s 0.05 and 0.95 quantiles was used to represent the uncertainty of the prediction results (Fig. 9B). The areas surrounding the cultivated land in the southern oasis had the highest uncertainty. Specifically, higher uncertainty is observed in the region located to the southeast of Dabasan Nuur Salt Lake. This significant uncertainty can be attributable to the insufficient training data available for these areas and the potential existence of substantial AGB in these territories. It is imperative to collect more samples in future research for enhanced prediction accuracy.

Considering the complex heterogeneity of the ecological environment in the study area, and despite some degree of uncertainty in the predictions, the AGB map still effectively reflects the overall spatial pattern and local characteristics of Haloxylon ammodendron forest AGB within the study area.

Analysis of variables

This study used all visible, near-infrared, and shortwave infrared bands at a 10–20-meter resolution from Sentinel-2, five derived vegetation indices, seven texture metrics, and the dual-polarization (VH and VV) backscatter measurements from Sentinel-1. Spatial CV used the Boruta algorithm to analyze the significance of each element. A high rank in importance was achieved by bands B11 and B12 in the shortwave infrared (SWIR) spectrum, along with the NDTI and RSR indices, which incorporate SWIR wavelengths. The high importance of these bands and indices may be linked to the distinctive biophysical characteristics of Haloxylon ammodendron trees found in the study area.

Due to the extremely arid climatic conditions of the study area, the leaves of Haloxylon ammodendron trees in this area have degenerated into scales, and their branches have grown as assimilating branches to perform photosynthesis. The chlorophyll area mass gradually decreases with the lignification process, where the mass of lignin and cellulose increases. Moreover, a certain proportion of the community consists of dead, standing plants (NPV). The infrared and near-infrared bands, along with common vegetation indices (e.g., NDVI, MSAVI), cannot effectively distinguish NPV from soil (Li, 2017), thus losing the capacity to estimate AGB. In environments with sparse canopy cover, the presence of bare soil significantly reduces the distinctive spectral features of photosynthetic vegetation (green plants). Additionally, the near-infrared and visible spectral bands display a noteworthy resemblance between soil and non-photosynthetic vegetation (e.g., dead plants and litter), with differences mainly observed in reflectance levels at specific wavelengths (Ren & Zhou, 2019), thus restricting the significance of the near-infrared and red bands.

The 2,000–2,400 nm shortwave infrared bands are sensitive to lignin and cellulose and are characteristic spectra for distinguishing dry vegetation, bare soil, and exposed bedrock (Cao et al., 2020). Studies have shown that the SWIR bands also exhibit some sensitivity to vegetation height and biomass (Cunliffe et al., 2020), while the SWIR spectrum shows sensitivity to the water content of canopy leaves within forest formations (Zhao et al., 2022). The NDTI, calculated using two different shortwave infrared bands (SWIR1 and SWIR2), effectively reduces the interference of soil background on vegetation indices, thereby more accurately reflecting vegetation conditions. This sensitivity makes the NDTI highly effective in distinguishing surface cover (such as crop residues and vegetation cover) from bare ground (Dai et al., 2018).

The role of the reduced simple ratio (RSR) is also significant. Vegetation indices such as EVI and MSAVI, which reduce the impact of soil background, are also prominent. The importance of elevation reflects the primary trend of Haloxylon ammodendron forest AGB increasing from the northwest to the southeast of the study area. This geomorphological transition—from low-lying areas (silty plains) and small dunes (less than five meters in height) to large dunes (greater than 10 m in height)—aligns with previous studies that have observed large-scale decline of Haloxylon ammodendron populations in flat terrain and small dunes, whereas those in larger dunes tend to thrive relatively well (Sun, 2018).

The role and importance of multi-temporal data

Included in the 12 factors ultimately used in the regression model of this study were two spring factors, two summer factors, four autumn factors, three winter factors, and elevation. In previous studies, data from the autumn and winter seasons were not customarily used. The prominent role of seasonal factors in this study might be attributable to the unique floristic composition of ephemeral plants in the Junggar Basin desert. In years with substantial winter and spring precipitation, the overall vegetation cover can exceed 70%, creating a lush green spring “meadow” appearance, which is significantly different from other deserts (Pan & Zhang, 1996). This spring seasonality of sand vegetation is unique to this region. Spring and summer are the growing seasons for these ephemeral plants, during which their vigorous growth reduces the difference between Haloxylon ammodendron trees and the soil background, altering the ecosystem’s structural composition and potentially causing signal confusion (Yuan & Tang, 2010; Chen, Zhang & Hu, 1983). In autumn and winter, ephemeral plants in the Junggar Basin gradually wither and die, revealing the desert background. The withered plant remains and exposed soil forms a unique autumn landscape, rendering the contours of perennial plants like the Haloxylon ammodendron tree more distinct and in stark contrast with the soil background. In winter, the soil is covered by ice and snow, further accentuating the difference between Haloxylon ammodendron and the soil background, thereby reducing signal interference in remote sensing data. Consequently, data from the autumn and winter seasons are particularly important for predicting Haloxylon ammodendron tree biomass. This finding enriches the understanding of remote sensing applications in desert ecosystems.

Analysis of the sources of uncertainty

Conventional random cross-validation is appropriate for datasets exhibiting a relatively uniform distribution (De Bruin et al., 2022). When machine learning models are used in locations that are geographically distant from the initial training location, these models may face challenges, especially if the new region being predicted exhibits a distinct feature space. In the presence of spatial autocorrelation, the conventionally used random 10-fold CV may fail to provide an accurate assessment of model performance. In this study, the UAV sampling points were of an aggregated sampling type from a spatial distribution perspective, and they did not achieve uniform coverage of the study area. As a result, there were discrepancies in the feature space between the training instances and the forecasted targets. Such differences in feature space representation inevitably leads to model extrapolation, which can reduce the performance and transferability of regression modeling (Schmidt et al., 2014; Wadoux, Brus & Heuvelink, 2019). For strongly clustered sampling designs that predict most of the map through extrapolation, the use of block spatial cross-validation can ensure that the extrapolation results are closest to the accuracy metrics of the reference map. Limiting the prediction area can help avoid unreasonable extrapolation. This study employed the K-fold kNNDM spatial cross-validation method to establish a QRF quantile model. Through AOA analysis and dissimilarity index (DI) analysis, the prediction area was restricted to a range as similar to the training samples as possible, excluding areas with significantly different spatial features such as lakes, cultivated lands, wetlands, and salt lakes, thereby avoiding unreasonable model extrapolation.

Although the model’s relative error is high (SMAPE of 101%), combining the SMAPE and RAE from the final QRF model (spatial cross-validation) resulted in an absolute error smaller than the baseline model (RAE of 0.89; Fig. 8). Considering the low or zero values of biomass in the study area, any small prediction error would be amplified, resulting in a very high relative percentage error and a significant increase in SMAPE value. In this context, it may indicate that the model’s predictions are relatively accurate in regions with higher biomass values. To verify this hypothesis, the training data were divided into low-value and high-value categories based on the 50th percentile of biomass values. An analysis of the observed and predicted values for these two groups yielded the results in Fig. 10.

In the low-value region, although the absolute errors (MAE, RMSE) were smaller, the relative errors (RAE, SMAPE) were larger. This indicates that the model’s predictions in these regions were relatively accurate; however, due to the small actual values, even minor prediction errors appeared relatively large. In the high-value region, the absolute errors (MAE, RMSE) were larger, but the relative errors (RAE, SMAPE) were smaller. This indicates that, although the model’s prediction errors were larger in these regions, the error proportion was smaller relative to the actual values. The larger errors in the high-value region were usually accompanied by higher prediction uncertainty, as the model’s predictions in these areas exhibited not only larger errors but also higher volatility. This explains why the RMSE was higher in the high-value region and reflected greater prediction uncertainty.

Potential method improvements

By incorporating a blend of data from various satellites and on-site measurements, machine learning algorithms were used to create models for estimating biomass in regions characterized by extremely sparse vegetation. The findings suggest that the assessment of biomass in such arid regions using the methodology developed in this study presents potential opportunities despite notable uncertainties. Further endeavors are necessary to develop more resilient, pragmatic, and economical methodologies for estimating shrub AGB through satellite remote sensing. One potential avenue for further research involves the exploration of diverse machine learning models or the fusion of remote sensing physical models. Using ensemble machine learning models that combine predictions from various individual models may lead to more accurate forecasts (Hao et al., 2020; Robert et al., 2016). Additionally, incorporating the latest deep learning techniques could be beneficial as these techniques can handle complex, non-linear relationships in remote sensing data and reduce uncertainties in biomass assessment (Theron et al., 2022) . Models need to establish a balance between accuracy and generalization performance to achieve better outcomes (Schratz et al., 2019).

Selecting suitable remote sensing parameters is essential for the estimation of AGB (Lu et al., 2014). Different combinations of variables can influence how AGB relates to remote sensing data. These factors include the limitations of remote sensing data—such as spectral bands, spatial resolution, and availability—along with climatic, topographical, and soil conditions that are relevant to the plant growth environment (Zhao et al., 2016). Selecting appropriate variables is essential for enhancing the accuracy of predictions. Research indicates that simpler machine learning models, which are built with only a limited number of predictive variables, demonstrate better adaptability than intricate models (Lauria et al., 2015). The results of this study suggest that the training inputs should be limited to essential model features.

Ultimately, the model’s precision relies on the distribution and representative nature of the sampling sites. To enhance the model’s adaptability and effectiveness, as well as its applicability to surveys in different or larger regions, it is crucial to ensure a uniform distribution of training data across all areas. Examining specific regions where the model shows weaknesses could indicate potential areas for future sampling.

The impact and novelty of the findings

This study presents a novel mapping approach by integrating UAV LiDAR data with high-resolution, multi-temporal satellite images (Sentinel-1 and Sentinel-2) to estimate the aboveground biomass (AGB) of sparse shrub communities in arid regions. This combination capitalizes on the complementary strengths of different data sources. UAV LiDAR offers detailed structural information on a small scale, while satellite images provide large-scale coverage and multi-temporal perspectives. The application of a spatial cross-validation method and the quantile regression forest (QRF) prediction model further improved the accuracy and reliability of AGB estimates. This integrated methodology fills an important gap in previous research on desert shrub AGB estimation.

The results of this study offer insights into the ecological characteristics of the Haloxylon ammodendron forest in the Junggar Basin. Mapping the AGB distribution over 14,000 square kilometers revealed significant spatial heterogeneity, with biomass increasing from northwest to southeast. This significant spatial heterogeneity was closely related to environmental factors like elevation and soil conditions. Notably, the discovery of the crucial role of seasonal factors, especially autumn and winter data, in predicting AGB is novel,as it uncovers the unique ecological processes in the desert ecosystem, where the phenology of ephemeral plants interacts with the perennial Haloxylon ammodendron, influencing spectral characteristics and AGB estimation accuracy. This deepens the understanding of the complex dynamics and factors shaping vegetation growth and distribution in desert ecosystems.

Accurate estimation of Haloxylon ammodendron AGB and its spatial distribution is highly significant for ecosystem conservation in arid regions. The results of this study serve as a scientific foundation for formulating targeted conservation strategies. Areas with higher biomass can be designated as key protection zones for implementing measures like restricting human activities and enhancing ecological monitoring. In contrast, regions with lower biomass and higher uncertainty may require more intensive restoration efforts such as reforestation and soil improvement. This study also provides a quantitative tool for evaluating the effectiveness of conservation measures over time by monitoring AGB changes. Overall, this research contributes to the sustainable management and conservation of desert ecosystems, which are vulnerable to climate change and human impacts.