Modeling MODIS LAI time series using three statistical methods

Development of the GLASS 250-m leaf area index product (version 6) from MODIS data using the bidirectional LSTM deep learning model

The Global Land Surface Satellite (GLASS) leaf area index (LAI) product is one of the most widely used global LAI products in the scientific community. The latest (version 5) GLASS LAI product has been generated from Moderate Resolution Imaging Spectroradiometer (MODIS) surface reflectance data. The purpose of this paper is to evaluate the quality of the version 5 GLASS LAI product. The GLASS LAI product was compared with the latest MODIS LAI product (MCD15A2 H, Collection 6) and the second version of Geoland2 (GEOV2) LAI product to evaluate their temporal and spatial discrepancies. A direct validation was conducted to compare these LAI products to the LAI values derived from the high-resolution reference maps from the Validation of Land European Remote Sensing Instruments (VALERI) and Implementing Multi-Scale Agricultural Indicators Exploiting Sentinels (IMAGINES) sites. The results show that the GLASS and GEOV2 LAI products have great spatial integrity. However, the MODIS LAI product contains many missing pixels in tropical areas. These LAI products follow fairly consistent seasonal characteristics. The spatial discrepancies of these LAI products mainly exist in forest areas, especially evergreen broadleaf forests where the GLASS LAI values are generally lower than the GEOV2 LAI values by approximately 1.0 LAI units and lower than the MODIS LAI values by 0.5 to 1.0 LAI units. The spatial distribution of these LAI products has slight discrepancies in savannahs, broadleaf crops, grasses/cereal crops and shrubs. The GLASS and GEOV2 LAI products capture a complete and reasonable temporal profile, contrasting with the MODIS LAI product, which shows dramatic fluctuations, particularly during the growing seasons. These LAI products show similar temporal trajectories and interannual variations for all biome types except evergreen broadleaf forests. The direct validation shows that the accuracy of the GLASS LAI product is better than the accuracy of the MODIS and GEOV2 LAI products. The coefficient of determination (R2) of the GLASS, MODIS and GEOV2 LAI products versus the LAI values derived from the high-resolution reference maps are 0.68, 0.47 and 0.55, respectively, and the root mean square error (RMSE) of these products are 0.86, 1.22 and 1.21, respectively.

Multiple satellite data can provide rich information for time series leaf area index (LAI) inversion. Long Short-Term Memory (LSTM) has been widely applied to capture long time dependencies from sequential data and solve the problem of vanishing (or exploding) gradients. It's feasible to retrieve time series LAI based on the LSTM algorithm. In this study, an LSTM model was built based on MODIS reflectance and a fused LAI from three satellite LAI products, including GLASS, MODIS and GEOV2 LAI. Then, time series LAI was obtained from the LSTM model coupled with reflectance data. Finally, the retrieval results were validated against fine resolution LAI reference maps and compared with the three LAI products. The accuracy of the proposed LSTMfusion LAI was best (R <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> =0.85, RMSE=0.69) among all the LAI estimations. In addition, the comparison among the temporal profiles of LSTMfusion LAI and the three global LAI products illustrated that the proposed model was able to efficiently generate time series LAI, which was more continuous and smoother than MODIS and VIIRS LAI.

The Multi-Angle Imaging Spectroradiometer (MISR) is a powerful sensor for leaf area index (LAI) mapping with its simultaneous multi-angle observations. However, the LAI product derived from MISR observations has low temporal resolution, which is unsatisfactory for many applications. This paper presents an algorithm that expands the MISR LAI product to high temporal resolution with the aid of Moderate Resolution Imaging Spectroradiometer (MODIS) data. The algorithm establishes relationships between the MISR LAI and the MODIS red/near-infrared band ratio (simple ratio (SR)) pixel by pixel using coincident data of these two sensors for the past nine years. Using these pixel-based SR-LAI relationships, a new LAI product with the merits of the original MISR product and high temporal resolution is obtained from MODIS surface reflectance. The expanded LAI series was compared with the original MISR and MODIS LAI products, as well as field LAI measurements made at the Baohe and Maoershan forest sites and the Hulunbeier grassland site, to assess the algorithm's performance. The results show that the temporal coverage of the MISR LAI improved from 15.5% to 65.2% in an 8-day composite, and the mean root-mean-square error is 0.74 for the vegetated pixels. This LAI product has similar temporal consistency and seasonal dynamics to the existing MODIS LAI product generated from the main algorithm, but is more robust against the low quality of reflectance inputs. The expanded LAI product differs with field measurements by about 11.5%, with agreement to field observations at all three sites within an accuracy of 0.8 LAI.

The modeling and predicting of vegetation Leaf area index (LAI) is an important component of land surface model and assimilation of remote sensing data. The MODIS LAI product (i.e. MOD15A2) is one of the most widely used LAI data sources. However, the time series of MODIS LAI contains some data of low quality. For example, because of the influence of the cloud, aerosol, etc., the MODIS LAI presents the characteristics of the discontinuous in time and space. In fact, the time series of MODIS LAI include both linear and nonlinear components, which cannot be accurately modeled and predicted by either linear method or nonlinear method alone. In this paper, the original LAI time series data were first smoothed with Savitzky-Golay (SG) filtration and linear interpolation; SARIMA, BP neural network and a hybrid method of SARIMA-BP neural network were then used for modeling and predicting MODIS LAI time series. The SARIMA-BP neural network combined both SARIMA and BP neural network, which could model the linear and the nonlinear component of MODIS LAI time series respectively. That is, the final result of SARIMA-BP neural network was the sum of results of the two methods. Experiments showed that the time series of MODIS LAI that were smoothed with the SG filtration and linear interpolation were more smooth than original time series, with a determination coefficient up to 0.981, closer to 1 than that of SARIMA (0.941) and BP neural network (0.884); the correlation coefficient between SARIMA-BP neural network and the observation is 0.991, higher than that of between SARIMA (0.971) or BP neural network (0.942) SARIMA and the observation. Thus, it can be concluded that, the proposed SARIMA-BP neural network method can better adapt to the LAI time series, and it outperforms the SARIMA and BP neural network methods.

Estimating bamboo forest aboveground biomass using EnKF-assimilated MODIS LAI spatiotemporal data and machine learning algorithms

Leaf area index (LAI) products retrieved from observations acquired on one occasion have obvious discontinuity in the time series owing to cloud coverage and other factors, and the accuracy may not meet the needs of many applications. Effectively utilizing data assimilation techniques to retrieve biophysical parameters from the time series of remote-sensing data has attracted special interest. The data assimilation technique is based on a reasonable consideration of dynamic change rules of biophysical parameters and time series observational quantities, thereby improving the quality of retrieved profiles. In this article, a variational assimilation procedure for retrieving LAI from the time series of remote-sensing data is developed. The procedure is based on the formulation of an objective function. A dynamic model is constructed based on the climatology from multi-year Moderate Resolution Imaging Spectroradiometer (MODIS) LAI data to evolve LAI in time, and a radiative transfer model is coupled with the dynamic model to simulate a time series of surface reflectances. A shuffled complex evolution method (developed at the University of Arizona; SCE-UA) optimization algorithm is then used to minimize the objective function and estimate the dynamic model states and the parameters of the coupled model from the MODIS reflectance data with a higher quality in a given time window. The variational assimilation method is applied to the MODIS surface reflectance data for the whole of 2008 at the Heihe river basin to produce regional LAI mapping results. The ground LAI data measured in situ are used to develop algorithms to estimate LAI from Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) surface reflectance, and ASTER LAI maps are produced for each ASTER scene using the algorithms developed. Then the ASTER LAI maps are aggregated to compare with the new LAI products. It is found that the variational assimilation method is able to produce temporal continuous LAI products and that accuracy has been improved over the MODIS LAI standard product.

On intra-annual EVI variability in the dry season of tropical forest: A case study with MODIS and hyperspectral data

Long-term high-quality global leaf area index (LAI) and fraction of absorbed photosynthetically active radiation (FAPAR) products are urgently needed for the study of global change, climate modeling, and many other problems. As the successor of the Moderate Resolution Imaging Spectroradiometer (MODIS) sensor, the Visible Infrared Imaging Radiometer Suite (VIIRS) will continue to provide global environmental measurements. This paper aims to generate longer time series Global LAnd Surface Satellite (GLASS) LAI and FAPAR products after the era of the MODIS sensor. To ensure spatial and temporal consistencies between GLASS LAI/FAPAR values retrieved from different satellite observations, the GLASS LAI/FAPAR retrieval algorithms were adapted in this study to retrieve LAI and FAPAR values from VIIRS surface reflectance time-series data. After reprocessing of the VIIRS surface reflectance to remove remaining effects of cloud contamination and other factors, a database generated from the GLASS LAI product and the reprocessed VIIRS surface reflectance for all Benchmark Land Multisite Analysis and Intercomparison of Products (BELMANIP) sites was used to train general regression neural networks (GRNNs). The reprocessed VIIRS surface reflectance data from an entire year were entered into the trained GRNNs to estimate the one-year LAI values, which were then used to calculate FAPAR values. A cross-comparison indicates that the LAI and FAPAR values retrieved from VIIRS surface reflectance were generally consistent with the GLASS, MODIS and Geoland2/BioPar version 1 (GEOV1) LAI/FAPAR values in their spatial patterns. The LAI/FAPAR values retrieved from VIIRS surface reflectance achieved good agreement with the GLASS LAI/FAPAR values (R2 = 0.8972 and RMSE = 0.3054; and R2 = 0.9067 and RMSE = 0.0529, respectively). However, validation of the LAI and FAPAR values derived from VIIRS reflectance data is now limited by the scarcity of LAI/FAPAR ground measurements.

A new method, based on pixel-to-pixel correlations, was proposed to replace unreliable Moderate Resolution Imaging Spectroradiometer (MODIS) leaf area index (LAI) time series (termed the “pixel-to-pixel correction (PPC) method”). MODIS LAI products improved by the PPC method were compared to field-measured LAI data. Results show that the PPC method could reduce the error associated with MODIS LAI products in relation to the original MODIS LAI product, and provide a level of accuracy that meets the most reliable quality control level criterion (quality control = 0). The PPC method performed well at correcting unrealistically high and unreliable LAI values, and performed slightly better than the “upper envelope” smoothing method. The corrected data were also more continuous and smoother than the raw MODIS LAI products. Thus, the PPC method can successfully reduce the uncertainty and improve the accuracy of raw MODIS LAI products.

Ecosystem process modeling is vital for a broad range of applications. It requires a key metric to characterize vegetation canopies: the leaf area index (LAI). Many land surface models utilize satellite-derived LAI for modeling ecosystem processes. The use of satellite-derived LAI constrains the models for predicting ecosystem processes when observational data are unavailable. Enriching prognostic models for predicting the LAI is favorable to land surface studies for forecasting vegetation-related processes. An attention-enhanced long and short memory (AELSTM) model was developed to predict the vegetation LAI time series based on climatic data. The AELSTM outperformed comparative machine learning models when assessed using satellite and flux tower data. The LAI predicted by AELSTM are temporally and spatially consistent with satellite-derived LAI. The R2 values attained 0.86–0.93 across biomes. AELSTM can be coupled with the land surface models to predict gross primary productivity. Our proposed model was demonstrated to be effective in predicting the LAI time series under different Shared Socioeconomic Pathways in CMIP6. Our study demonstrated that deep learning approaches are capable of modeling and characterizing the spatial and temporal patterns of key vegetation variables such as LAI. Moreover, the study has provided references for vegetation processes in land surface studies.

Leaf area index (LAI) is a key indicator of vegetation structure and function, and its products have a wide range of applications in vegetation condition assessment and usually act as important input parameters for ecosystem modeling. Grassland plays an important role in regional climate change and the global carbon cycle and numerous studies have focused on the product-based analysis of grassland vegetation changes. However, the performance of various LAI products and their discrepancies across different grassland types in drylands remain unclear. Therefore, it is critical to assess these products prior to application. We evaluated the accuracy of four commonly used LAI products (GEOV2, GLASS, GLOBMAP, and MODIS) using LAI reference maps based on both bridging and cross-validation approaches. Under different grassland types, the GLASS LAI performed better in meadow steppe (R2 = 0.26, RMSE = 0.41 m2/m2) and typical steppe (R2 = 0.32, RMSE = 0.38 m2/m2); the GEOV2 LAI performed better in desert steppe (R2 = 0.39, RMSE = 0.30 m2/m2). When we assessed their spatial and temporal discrepancies during the period from 2010 to 2019, the four LAI products overall showed a high spatial and temporal consistency across the region. Compared with GLASS LAI, the most consistent to least consistent correlations can be ordered by GEOV2 LAI (R2 = 0.94), MODIS LAI (R2 = 0.92), and GLOBMAP LAI (R2 = 0.87). The largest differences in LAI throughout the year occurred in July for all grassland types. Limited by the location and number of sample plots, we mainly focused on spatial and temporal variations. The spatial heterogeneity of land surface is pervasive, especially in vast grassland areas with rich grassland types, and the results of this study can provide a basis for the application of the product in different grassland types. Furthermore, it is essential to develop highly accurate and reliable satellite-based LAI products focused on grassland from the regional to the global scale according to these popular approaches, which is the next step in our work plan.

Time series of Leaf Area Index (LAI) is of utmost importance for various disciplines of bio- and geosciences where satellite remote sensing makes LAI estimation possible for large areas. Remote sensing LAI, validated against in situ LAI (LAIinsitu), is used as base for calculating LAI for large areas e.g. on catchment scale. Various vegetation indices (NDVI, SAVI, both with and without substituting the red band with red-edge band) were applied for better estimates of LAIrapideye. SAVI (Soil Adjusted Vegetation Index) and NDVI (Normalized Difference Vegetation Index) present same correlation between remote sensing based predicted LAI (LAIrapideye) against LAIinsitu in winter wheat fields. Both NDVI and SAVI with red-edge band showed improved correlation of remote sensing based VI and in situ measurements. Prior to vegetation indices calculation, radiometric normalization was applied to the time series of RapidEye data. To test the impact of radiometric normalization for calculating vegetation indices on a time series of satellite images, pre- and post-radiometric normalization LAIrapideye was compared. More precise and high resolution estimation of LAI for large areas is of vital importance for improving evapotranspiration and soil moisture.

The aim of this study was to compile the leaf area index (LAI) map of a 3×3 km VALERI test site in Järvselja, Estonia. Canopy transmittance measurements of LAI by LAI-2000 Plant Canopy Analyzers and digital cameras supplied with fisheye converters were carried out on 42 elementary sampling units at ground level and at breast height level. The vegetation LAI was estimated as a sum of the tree canopy true green LAI obtained from the inversion of canopy gap fraction data and of the ground vegetation effective LAI. Red channel from SPOT-4 HRV-IR image and airborne lidar data based canopy transmittance were used for up-scaling. The LAI map was compared with standard LAI products from Terra MODIS and ENVISAT MERIS. We found that lidar data based LAI estimate on up-scaled map saturated at high values (LAI &gt; 4.5) compared to the LAI estimates based on SPOT-4 HRV-IR red channel. Validation of MODIS LAI product revealed substantial underestimates of LAI compared to the up-scaled field measurements and rather large random noise. ENVISAT MERIS LAI product was more similar to up-scaled field measurements; however, rather large unexpected random variations exist in its time series.

Climate, nitrogen (N) and leaf area index (LAI) are key determinants of crop yield. N additions can enhance yield but must be managed efficiently to reduce pollution. Complex process models estimate N status by simulating soil-crop N interactions, but such models require extensive inputs that are seldom available. Through model-data fusion (MDF), we combine climate and LAI time-series with an intermediate-complexity model to infer leaf N and yield. The DALEC-Crop model was calibrated for wheat leaf N and yields across field experiments covering N applications ranging from 0 to 200 kg N ha−1 in Scotland, UK. Requiring daily meteorological inputs, this model simulates crop C cycle responses to LAI, N and climate. The model, which includes a leaf N-dilution function, was calibrated across N treatments based on LAI observations, and tested at validation plots. We showed that a single parameterization varying only in leaf N could simulate LAI development and yield across all treatments—the mean normalized root-mean-square-error (NRMSE) for yield was 10%. Leaf N was accurately retrieved by the model (NRMSE = 6%). Yield could also be reasonably estimated (NRMSE = 14%) if LAI data are available for assimilation during periods of typical N application (April and May). Our MDF approach generated robust leaf N content estimates and timely yield predictions that could complement existing agricultural technologies. Moreover, EO-derived LAI products at high spatial and temporal resolutions provides a means to apply our approach regionally. Testing yield predictions from this approach over agricultural fields is a critical next step to determine broader utility.