This dataset is based on the backpropagation neural network (BPNN) model and pixel integration method, and a new generation of GIMMS LAI product (GIMMS LAI4g, 1982-2020) has been developed. The feature of GIMMS LAI4g is the use of Peking University GIMMS NDVI products and a large number of high-quality Landsat LAI samples. The recently released PKU GIMMS NDVI effectively eliminates the effects of NOAA orbital drift and AVHRR sensor degradation, which have been key issues with existing LAI products. The total number of high-quality global land satellite LAI samples reaches 3.6 million, covering the period from 1984 to 2015, which provides convenience for creating spatiotemporal consistent BPNN models. The GIMMS LAI4g product, which is consistent in time and space, covers the time span from 1982 to 2020 with a time resolution of 15 days. It can provide strong data support for long-term vegetation monitoring and high-precision, high reliability model development.
This study used eight global datasets, namely Peking University GIMMS NDVI, Landsat LAI sample dataset, MODIS land cover types, reprocessed MODIS LAI, GLASS LAI, GLOBMAP LAI, GIMMS LAI3g, and field LAI measurements. PKU GIMMS NDVI is the main data source for generating GIMMS LAI4g. The land satellite LAI sample dataset is used as a LAI reference for machine learning model building and product evaluation. Field LAI measurement is also used for product evaluation. MODIS land cover type products provide vegetation biotic community types for LAI modeling. The reprocessed MODIS LAI is used to extend the time coverage of GIMMS LAI4g. GLASS LAI, GLOBMAP LAI, and GIMMS LAI3g are three mainstream global LAI products that have been included for mutual comparison.
Using GIMMS NDVI and Landsat LAI samples from Peking University to address the uncertainty issue of remote sensing and LAI reference data. Using this data, a community specific backpropagation neural network (BPNN) model was developed, with explanatory variables added (longitude and latitude, NDVI month, and number and year after NOAA emission). Then generate GIMMS LAI4g products based on the BPNN model. Finally, the GIMMS LAI4g was merged with the reprocessed MODIS NDVI product through pixel fusion method, expanding the time coverage to 2020
According to the validation of Landsat LAI samples, R2is 0.96, with a root mean square error of 0.32m2m-2, an average absolute error of 0.162m-2, and an average absolute percentage error of 13.6%, achieving the accuracy target proposed by the global climate observation system. In most vegetation communities on land, its LAI performance is superior to other products. It effectively eliminates the effects of satellite orbit drift and sensor degradation, and exhibits better temporal consistency around 2000. After merging with the reprocessed MODIS LAI, the time coverage of GIMMS LAI4g has expanded from 2015 to the near future (2020), and the generated LAI trend remains highly consistent around 2000 and is consistent with the reprocessed MODIS LAI trend of the MODIS era. The GIMMS LAI4g product may help reduce the divergence between global vegetation long-term change studies and also facilitate model development in Earth and environmental science
This work is licensed under a
Creative
Commons Attribution 4.0 International License.
CSTR:11738.11.NCDC.ZENODO.DB6427.2024
10.5281/zenodo.8281930