The biome classification map serves as an auxiliary dataset for MODIS LAI, primarily aimed at reducing uncertainty in the retrieval algorithm. In the MODIS retrieval algorithm, parameters are configured based on the biome classification map to establish an accurate relationship between satellite observations and ground parameters (Knyazikhin, 1999). The MODIS Land Cover Type product (MCD12Q1) provides global land cover maps at annual time steps and a 500 m spatial resolution from 2001 to the present. The product was created using supervised classification of MODIS reflectance data for a total of 13 scientific datasets. MCD12Q1 adopts an LAI legacy classification scheme, including B1 (grass and cereal crops), B2 (shrubs), B3 (broadleaf crops), B4 (savanna), B5 (evergreen broadleaf forest, EBF), B6 (deciduous broadleaf forest, DBF), B7: (evergreen coniferous forest, ENF), and B8 (deciduous coniferous forest, DNF) (Yan et al., 2016a; Sulla-Menashe and Friedl, 2018). Figure 1 illustrates the approximate global distribution of biome types based on the land cover in 2021.

The standard MODIS LAI C6.1 product suite (MOD15A2H) offers global coverage with an 8 d temporal resolution and a 500 m spatial resolution (Yan et al., 2016a) spanning from February 2000 to the present. Typically, there are 46 composites per year, but some composites may be missing due to factors such as sensor issues or other anomalies (e.g., in 2001, 2016, and 2022). This LAI product is projected onto a sinusoidal grid and distributed as a standard Hierarchical Data Format (HDF) file. Each file contains six science datasets (SDSs): fraction of photosynthetically active radiation absorbed by vegetation (FPAR), LAI, FparLai_QC, FparExtra_QC, FparStdDev, and LaiStdDev. The LAI, FparLai_QC, and LaiStdDev layers store the LAI retrieval, quality control information, and retrieval uncertainty, respectively (Knyazikhin et al., 1998; Yan et al., 2016a; Myneni, 2020). Therefore, for this study, we utilized these three SDSs.

With the increasing availability of Earth observation products, increasing attention has been paid to the product uncertainty assessed by verification based on ground measurements (Baret et al., 2006; Fang et al., 2012; Wenze et al., 2006). To evaluate the performance of the HiQ-LAI product, we utilized the Copernicus Ground-Based Observations for Validation (GBOV) LAI, DIRECT 2.1 LAI, and the BEnchmark Land Multisite ANalysis and Intercomparison of Products 2.1 (BELMANIP 2.1) as ground site references (Bai et al., 2019; Baret et al., 2006; Morisette et al., 2006).

The BELMANIP site network was designed to represent the global variability in vegetation types and climatological conditions, independent of ground-based experimental measurements (Baret et al., 2006). The network primarily incorporates sites from existing experimental networks (FLUXNET, AERONET, VALERI, BigFoot, etc.) supplemented by sites from the Global Land Cover 2000 (GLC2000) land cover map. The BELMANIP V2.1 that we used was constructed using the GlobCover vegetation land cover map derived from Medium Resolution Imaging Spectrometer (MERIS) images in 2009. Site selection was conducted for each band of latitude (10° width), ensuring that the selected sites represented the same proportion of biome types as the entire latitude band. These sites are characterized by a uniform and nearly flat topography within a 10 km × 10 km area, with a minimum proportion of urban areas and permanent waterbodies. The updated BELMANIP V2.1 dataset (Fig. 1) added 25 sites corresponding to bare land (desert) and tropical forests, resulting in a total of 445 sites (https://calvalportal.ceos.org/web/olive/site-description, last access: 21 March 2024). It is important to note that most of these sites do not have ground reference measurements. Consequently, the network is primarily used for comparative analysis among sites rather than direct validation.

We proposed a spatiotemporal information composition algorithm (STICA) aimed at reducing noise fluctuations and improving the overall quality of the MODIS LAI product. This algorithm directly incorporates the prior spatiotemporal correlation information and multiple quality assessment (MQA) information into the existing MODIS LAI product. The detailed algorithmic process can be found in the article published by Wang et al. (2023). The algorithm consists of four main steps: multiple quality assessment, employing spatial correlation information, employing temporal correlation information, and multiple information compositing.

Satellite remote-sensing observations are often subject to uncertainties arising from climatic factors, sensor malfunctions, and other sources, resulting in varying levels of uncertainty for individual pixels. To address this issue, this approach employed multiple indicators to evaluate the uncertainty for each pixel (referred to as MQA hereafter). These indicators encompass the algorithm path, LAI standard deviation (SD), and relative time-series stability (TSS). The algorithm path (AP) is a crucial quality index, distinguishing between the main and backup algorithms. The main algorithm offers superior quality and precision retrieval, and the weight ratio of the main algorithm and backup algorithm has been determined as 6:4 in a previous study (Wang et al., 2023). The LAI SD reflects the retrieval uncertainty. The AP and LAI SD are derived from the FparLai_QC and LaiStdDev layers of the original MODIS data. The third indicator, relative TSS (RE-TSS), indicates the fluctuation in a time series (Zou et al., 2022). Following the principle of assigning a higher weight to smaller values, the LAI SD and RE-TSS are incorporated into the retrieval with the main algorithm, resulting in the generation of a new quality classification indicator, the MQA. Subsequently, the inverse distance weighting (IDW) method is utilized on the spatial scale to calculate the weighted average of all eligible pixels (belonging to the same land cover type) within the half-width of four pixels and the power exponent of 2 (Wang et al., 2023) of the target pixel. In this algorithm, the contribution of a pixel is determined not only by its spatial distance but also by its MQA value. In a word, pixels with closer proximity and a higher MQA value make a more significant contribution to the target pixel. On the temporal scale, the simple exponential smoothing (SES) method is employed to calculate the weighted average of all eligible pixels within the smoothing parameter of 0.5 and the half-length of 3 (Wang et al., 2023). Pixels that are closer in time to the target pixel and possess higher MQA values are assigned greater weights. Utilizing spatial/temporal correlation is based on spatial and temporal autocorrelation, i.e., everything is related to everything else, but near things are more related than distant things. The final step of the algorithm is to take a weighted average of the original MODIS LAI and the LAI calculated using spatial/temporal correlation, with their respective weights quantified using an indicator (TSS) that represents the temporal fluctuation in the time series (Zou et al., 2022). All processes of the method are implemented using the GEE cloud computing platform. The reprocessed LAI dataset, namely the HiQ-LAI product, has been generated with the help of the powerful cloud computing capability of GEE, covering the period from 2000 to 2022.

In this study, we utilized the GBOV LAI measurements from a total of 29 sites spanning from 2013 to 2021 as our ground reference LAI (Bai et al., 2019; Brown et al., 2020). A 3 km × 3 km square centered on the site location was selected as the study area (Fig. 1) so that the corresponding LAI product of each site was 36 (6 × 6) pixels. To enhance the credibility of the ground truth LAI, we filtered the ground LAI reference of these 29 sites based on the criterion that the “effective pixel” exceeded 90 % and the input and output of the land product value in the data aggregation process were within the specified range. This filtering process yielded a total of 818 reliable verification data points. Contrary to previous studies (Wang et al., 2023) that utilized only 2018 data from the GBOV site as a reference, this study expanded the timeline from 2013 to 2021, increased the number of sites from 24 to 29, and raised the criterion for effective pixels from 80 % to 90 %. These modifications were aimed at enhancing the reliability of the ground LAI data. Additionally, previous research focused on proposing and testing algorithms mainly at the tile scale, but this study migrated the algorithm to GEE to generate global long-term data series. Furthermore, the scope of analysis was also broadened to a global spatial scale and long-term time series. A comparative analysis was conducted at the spatial scale to examine the global spatial distribution of the LAI in February and July 2021. The mean LAI values for latitude bands were then calculated at 1° intervals during these specific months. Furthermore, we compared the global consistency of MODIS LAI and HiQ-LAI in 2021 using the BELMANIP V2.1 sites (445 in total) (Baret et al., 2006). Employing these sites not only reduced the computational burden at a global scale but also mitigated additional uncertainties arising from geometric registration bias and land cover misclassification. Similar to the GBOV, we selected a study area of 6 pixels × 6 pixels centered on each site location (Fig. 1). The MCD12Q1 data in 2021 were utilized to determine the biome type of each site, which was further classified into pure pixels and hybrid pixels based on B1–B7. The total number of data available for comparison was 16 420. Additionally, we used DIRECT V2.1 ground measurements in this research (Morisette et al., 2006; Garrigues et al., 2008). However, these data were not utilized for direct validation due to the discontinuity in the observed time series at these sites. Instead, the DIRECT V2.1 sites provided valuable reference values in Sect. 5.2. Similarly, a research area of 6 pixels × 6 pixels was selected for each site, and we compared the R2 and RMSE of the two products with sites across different quality grades. The analysis involved determining the RMSE reduction percentage and R2 increase in the percentage of HiQ-LAI relative to MODIS under various quality grades.

In this study, the Theil–Sen slope (TS) method and Mann–Kendall (MK) test (Suhartati, 2013; Theil, 1992) were employed to extract LAI trends from the two products. The TS method computes pairwise slopes across the study period, with the median slope representing the sign and magnitude of the long-term trend. Unlike ordinary least-squares linear regression, the TS trend is less susceptible to the influence of outliers. Meanwhile, the MK test is utilized to determine the significance of the trend (Kendall, 1948). The combination of TS and MK forms a robust approach for identifying trends in long-term sequential data. TS and MK are calculated as follows: 1TS=medianXj-Xij-i,2000≤i<j≤2022, where Xj and Xi represent the LAI value of year j and year i, respectively. Here, TS>0 indicates an increasing trend, whereas TS≤0 indicates a decreasing trend. Following this, the MK test was applied to assess the annual mean trends for MODIS LAI and HiQ-LAI from 2000 to 2022, ensuring the statistical significance of the identified trends. 2S=∑i=1n-1∑j=i+1nsgnxj-xi,3VarS=nn-12n+5-∑i=1mtiti-12ti+518,4Zs=S-1VarS,ifS>00,ifS=0S+1VarS,ifS<0. Here, S represents the sum of step function values obtained from the differences between any two distinct points within the time series, n signifies the total number of data points, m indicates the count of continuous groups in the data (duplicate dataset), and ti refers to the associated count (the number of repetitions in the ith range). Ultimately, we calculate the test statistic Zs, when Zs>Z1-α/2 means reject the null hypothesis (i.e., the absence of a trend), with α representing the significance level. In our analysis, we set α=0.05, with Z1-α/2=1.96 (indicating significance at 90 % and 95 % confidence levels when equal to 1.65 or 1.96, respectively).

The definition of TSS as follows: 5TSSt=Xtn+1-Xtn-1×tn-Xtn×tn+1-tn-1-Xtn+1-Xtn-1×tn-1+Xtn-1×tn+1-tn-1Xtn+1-Xtn-12-tn+1-tn-12, where X(tn), Xtn+1, and Xtn-1 represent the LAI value at target moment t, the adjacent time-series data obtained at the previous moment, and the next moment, respectively. TSS denotes the deviation of a value at a given point in time from the linear interpolation line. In this study, higher TSS values indicate greater variability over time.

Figure 2 depicts the validation results obtained from GBOV ground reference LAI, highlighting the superior performance of HiQ-LAI compared with MODIS LAI. From MODIS LAI to HiQ-LAI, the R2 increases from 0.69 to 0.71, the RMSE decreases from 0.87 to 0.78, the relative RMSE (RRMSE) decreases from 26.63 % to 24.04 %, and the bias shifts from -0.17 to -0.06. Notably, the fitted line of MODIS LAI deviated more prominently from the 1:1 line compared with HiQ-LAI, indicating that HiQ-LAI exhibits higher accuracy with the ground reference LAI. The two methods had negligible differences in LAI performance in terms of the pure-grassland and mixed-grass biome types. However, MODIS LAI exhibited a substantial overestimation for the pure-forest type (Fig. 2, triangles). Incorporating data quality information (MQA), we observed a correspondence between high LAI values and low MQA values, and the majority of pure-forest and mixed-forest sites show the phenomenon of low MQA values. In contrast, the mixed savannas and forest biomes (Fig. 2, pentagrams) displayed an underestimation in MODIS LAI. Furthermore, MODIS LAI exhibited noticeable abnormal retrieval values (Fig. 2, red pentagrams in the upper left-hand corner), whereas HiQ-LAI effectively mitigated these issues. Comparing the LAI difference distribution among various vegetation types (Fig. 3) revealed that HiQ-LAI exhibited a tighter concentration around the zero value, resulting in a decreased RMSE across most categories, except for the third biome type (grasses and shrubs). Mixed savannas and forests emerged as the vegetation types with the widest MODIS LAI difference range. The enhanced HiQ-LAI notably narrowed this distribution range, although the median and mean deviated further from zero. Notably, the two biome types exhibiting the most conspicuous changes in difference distribution were pure forest and mixed crops and savannas. The verification analysis (Fig. 4), comparing both products against GBOV LAI references across different seasons, demonstrated that HiQ-LAI had superior performance over MODIS LAI throughout all four seasons and also outperformed the ground references. Analyzing the LAI density distribution revealed that MODIS LAI (Fig. 4, green) skewed towards higher values on the right-hand side compared with HiQ-LAI (Fig. 4, black). This indicated that MODIS LAI predominantly occupied high-value areas. Furthermore, the RMSE and RRMSE of HiQ-LAI are always smaller than those of MODIS LAI.

The LAI curve should exhibit stable annual and seasonal fluctuations in response to seasonal changes. Nevertheless, various factors such as harsh atmospheric conditions, sensor hardware issues, and other technical challenges introduce significant discrepancies between adjacent time windows of MODIS LAI observations (Garrigues et al., 2008), leading to abnormal fluctuations in the LAI time-series profiles. As depicted in the LAI time-series comparison diagram (Fig. 5 and Fig. S1 in the Supplement), HiQ-LAI demonstrates improved quality for different vegetation types. The time-series curves generated by HiQ-LAI exhibit reduced abnormal fluctuations and better alignment with expected phenological patterns compared with the original MODIS LAI. Notably, the RMSE values of all sites (totaling 29, as shown in Table 1 and Figs. 5 and S1) exhibit varying degrees of reduction, except for the DSNY, JERC, JORN, and SERC sites. Four sites (MOAB, STER, VASN, and WOMB) were excluded from calculating correlation coefficients due to the unavailability of validation data after filtering. Note that, as in situ measurements may be sensitive to all elements of the canopy, the resulting estimate should technically be referred to as the plant area index (PAI) (Brown et al., 2021). Studies have demonstrated that the PAI may overestimate LAI by as much as 61 % in certain scenarios (Brown et al., 2021). Considering this, Figs. 2, 5, and S1 illustrate that MODIS overestimates the LAI of forests to some extent, which is a finding consistent with prior research (Yan et al., 2016b, 2021a). However, HiQ-LAI corrects this overestimation to a certain degree.

Figure 6 displays the 5 km spatial distribution of global MODIS LAI and HiQ-LAI for February and July. Overall, the two products demonstrate comparable spatial patterns across global regions. However, notable distinctions are shown in February, particularly in the Amazon rainforest region of South America. The mean LAI latitudinal profile for both products shows remarkable similarity, indicating consistent overall performance. Nevertheless, it is worth noting that the standard deviation of HiQ-LAI (green shading in Fig. 6c1 and c2) consistently falls within the range of the standard deviation of MODIS LAI (red shading in Fig. 6c1 and c2). This suggests that HiQ-LAI exhibits more stability at a global scale, indirectly implying superior data quality and enhanced stability compared with MODIS LAI.

To further assess the performance of HiQ-LAI across different biome types, we extracted MODIS LAI and HiQ-LAI values from the BELMANIP V2.1 sites. A scatterplot (Fig. 7) was constructed based on the biome type to evaluate the consistency between the two products. The results demonstrate that, except for B5, the R2 for other monocultural vegetation types exceeds 0.88, and B1 and B3 surpassed 0.95. The consistency of mixed pixels is also relatively high, as indicated by an RMSE of 0.42 and an R2 of 0.86. However, B5 exhibits a significant disparity, with an R2 value of 0.15. The biome of B5 is primarily grown in the Amazon rainforest region of South America, Indonesia, and central Africa (Fig. 1). These regions are affected by the long-term influence of thick cloud cover, a high concentration of aerosol, and a saturation of red NIR, resulting in the limited availability of high-quality observation data and the poor accuracy of the LAI retrieval (Xu et al., 2018; Yan et al., 2016b). Consequently, it can be inferred that HiQ-LAI and MODIS LAI are consistent in most areas (different vegetation types) but inconsistent in local areas that suffer from quality issues.

Furthermore, we conducted a histogram of the numerical distribution (Fig. 8) of the two products under different vegetation types globally in 2021 by season. A similar phenomenon can be found in Fig. 7, with the most notable discrepancies observed in the numerical distribution of type B5. Moreover, substantial differences were noted across various seasons for B5, and the red vertical line representing the mean of MODIS LAI is consistently positioned to the left of the green vertical line representing the mean of HiQ-LAI. The distribution of the B1–B4 biomes remained highly consistent throughout different seasons with minimal variation. Conversely, the B6–B8 biomes displayed differences in the numerical distribution between spring and summer, demonstrating an opposite trend. In spring, the mean value of MODIS LAI (red vertical line) was consistently higher than that of HiQ-LAI (green vertical line), whereas the red vertical line was always located on the right of the green line in summer.

To compare the vegetation change trend between MODIS LAI and HiQ-LAI products on a global scale, we computed the LAI mean from 2000 to 2022. By determining the slope of the time series through the fitting and then utilizing the Mann–Kendall (MK) test to discern the significance and monotonicity of trends, we obtained the spatial trend change map for this period (Fig. 9), revealing regions exhibiting greening trends (positive slope) or browning trends (negative slope) in vegetation. The spatial trend change map provides valuable insights into the dynamics of vegetation across the globe. Both products exhibit similar spatial patterns in greening trends, particularly in global hotspots such as China and India. Similarly, the browning trend is observed in similar locations in both products. From MODIS to HiQ-LAI, the proportion of insignificant change increased from 37.03 % to 39.95 %. The greening trend and browning trend of the two products were 60.24 % (MODIS) and 56.69 % (HiQ) and 2.73 % (MODIS) and 3.36 % (HiQ), respectively. The slight difference between the two products is mainly concentrated in high-latitude areas.

Typically, the LAI exhibits continuous variations throughout the year without significant fluctuations (Zou et al., 2022). However, the annual seasonal variation curve of MODIS LAI shows pronounced abnormal fluctuations (e.g., sudden increases and decreases). This phenomenon is attributed to the independent pixel-by-pixel daily inversion process of MODIS LAI. When affected by atmospheric conditions, sensor malfunctions, and retrieval algorithm uncertainties, MODIS LAI experiences poor spatiotemporal consistency and an accompanying increase in the noise level (Yan et al., 2021b; Fang et al., 2019). Consequently, it fails to accurately capture the long-term trend in LAI variations, thereby restricting its application with respect to crop modeling and prediction (Fang et al., 2011; Ines et al., 2013; Zhuo et al., 2022), climate change and vegetation dynamics research (Zheng et al., 2021; Chen et al., 2021), and long-term ecological environment monitoring and assessment (Dhorde and Patel, 2016; Mariano et al., 2018). From the ground-based direct validation results (Sect. 4.1), it is evident that HiQ-LAI effectively improves the data quality of the original MODIS LAI. It reduces apparent error in the raw LAI retrievals and generates smoother LAI time-series profiles that align better with expected phenological patterns. However, the GBOV site data are limited by spatial scale and resources, making them insufficient for comparing temporal changes on a global scale. Therefore, we utilize the TSS (Zou et al., 2022), a quantifiable metric for time-series stability, to further understand the temporal stability performance of the two products. HiQ-LAI encompasses abundant smaller TSS pixels (Fig. 10), particularly in regions near the Equator that are characterized by the EBF vegetation cover type. The proportion of low TSS values globally increases from 31.8 % (MODIS) to 78.8 % (HiQ). This confirms that our proposed HiQ-LAI product substantially improves the stability of time series compared with the original MODIS LAI, resulting in an overall enhancement in data quality. This will provide more reliable data support for the research and applications of ecology, climatology, land-use planning, and other fields.

Ground verification also reveals a consistent pattern between low MQA values and high LAI retrievals. Regions with high LAI values are mainly observed in tropical regions where there is greater cloud cover and aerosol load, causing signal saturation (Yan et al., 2016b; Samanta et al., 2012a). In the first step of our algorithm, pixel quality is assessed based on the algorithm path and LAI standard deviation of the main algorithm, where the saturation phenomenon reduces the MQA value. Thus, this phenomenon can be attributed to the overestimation of MODIS LAI retrievals due to signal saturation (Heiskanen et al., 2012). Moreover, other studies have confirmed that the main algorithm's saturation tends to yield higher MODIS LAI (Yuan et al., 2011; Lin et al., 2023). We graded pixels according to the MQA value (Fig. 11) and observed that the improvement effect on LAI retrievals became more pronounced as the quality level decreased. In the poor-quality level, HiQ-LAI exhibited a 17.81 % increase in the R2 and an 18.99 % reduction in the RMSE compared with MODIS LAI. However, the RMSE only decreased by 2.11 % in the good-quality level. HiQ-LAI substantially enhances the quality of LAI retrievals affected by observational conditions and inversion algorithms while maintaining the same pixels with good quality in the original MODIS LAI. This highlights the effectiveness of our proposed algorithm with respect to improving the spatiotemporal consistency of LAI products.

Due to the long-term effects of extensive cloud cover contamination, high aerosols, and dense vegetation saturation, the availability of high-quality observation data from optical remote sensing in tropical regions is severely limited (Huete et al., 2002; Yuan et al., 2011; Xu et al., 2018). Among these regions, the Amazon forests, in particular, have a significant impact on global climate (Cox et al., 2013; Jimenez et al., 2018; Guimberteau et al., 2017) and carbon and water cycles (Marengo and Espinoza, 2016; Poulter et al., 2014; Yang et al., 2018). MODIS LAI retrievals are calculated independently pixel-by-pixel and daily. The varying observational conditions between adjacent time windows introduce more uncertainty, leading to relatively poor consistency in both temporal and spatial dimensions. Spatial pattern changes in MODIS LAI in short-term time series (Fig. 13) show that some higher values suddenly decrease in the next period, while some lower values suddenly increase. This phenomenon is unusual in heavily vegetated tropical regions, as previous studies have concluded that the LAI in Amazon forest areas is seasonal (Hashimoto et al., 2021; Samanta et al., 2012b; Myneni et al., 2007). Correspondingly, regions with sudden LAI changes exhibit lower MQA values, indicating that our method effectively identifies retrievals with quality issues.

The differences between the two products are mainly observed in the Amazon forest regions of South America, Indonesia, and central Africa. These areas are primarily characterized by the EBF vegetation type (Fig. 1), which aligns with the differences observed in the biome comparisons (Figs. 7, 8). The spatial pattern difference between MODIS LAI (Fig. 12a1) and HiQ-LAI (Fig. 12b1) at an 8 d resolution is the most pronounced. However, the spatial distributions of MODIS LAI and HiQ-LAI gradually become consistent as the synthesis period increases (Fig. 12). The 8 d relative difference (RD) density distribution shows a distinct peak towards the leftmost side. Nevertheless, the peak flattens out as the synthesis period increases, and the RD becomes more concentrated around zero. No significant spatial distribution change was observed in HiQ-LAI with an increasing synthesis period, and the 32 d synthesis data of MODIS LAI exhibited a similar spatial pattern to the 8 d product of HiQ-LAI. This evidence indirectly indicates that, compared with the original MODIS LAI, our products exhibit fewer abnormal fluctuations in the time series and greater stability in product quality. In terms of time-series stability and anomalies (Fig. 14), HiQ-LAI also outperforms MODIS LAI. By comprehensively considering pixel quality information and prior spatiotemporal correlation information, our algorithm effectively compensates for the abnormal fluctuations caused by differences in observational conditions. Through the optimization of MODIS LAI retrievals with poor quality, our algorithm substantially improves the quality of original MODIS LAI products and enhances their applicability in tropical regions. The increased stability and reliability of the HiQ-LAI product provide more reliable data support for research in tropical regions.