Vegetation is the basic component of the terrestrial ecosystem and plays an important role in energy exchange, carbon cycling and hydrological cycling process on the earth surface [1
]. Therefore, timely and accurate land surface vegetation information is of great significance for earth system science, ecological environment assessment and climate change related studies [4
]. Leaf area index (LAI), generally defined as one half of the total green leaf area per unit of horizontal ground surface area [5
], is an important parameter for characterizing land surface vegetation conditions [6
]. LAI is an essential parameter that characterizes the density of leaves and canopy structure, which can reflect the vegetation’s ability for biophysical processes such as photosynthesis, respiration and transpiration, and is also an important input variable for the carbon cycle models, crop growth models and water cycle models [9
]. Therefore, accurate estimation of LAI on regional and global scales is of great significance for ecosystem modelling, biogeochemical cycle modelling, agriculture monitoring and related applications.
Remote sensing provides the only effective means to estimate LAI at the regional and global scales for its ability of continuous observation and to provide broad and impartial earth observation data [7
]. Currently, several global LAI products have been generated using medium to low spatial resolution satellite data, such as the MODIS LAI products [6
], GEOV1 LAI Products [7
], MERIS LAI products [16
], GLASS LAI products [17
] and GLOBCARBON LAI products [18
]. However, the currently existed LAI products generally have low spatial resolutions at the kilometric level, whereas decametric spatial resolution LAI data will be better suited for applications related to agriculture monitoring, as compared to kilometric resolution LAI data, which are usually larger than the typical scales of most croplands. The wide field view (WFV) sensor on board GF-1, the first satellite of the China High-resolution Earth Observation System, can acquire multi-spectral data with high spatial and temporal resolutions [19
], which are valuable data sources for dynamic LAI monitoring at regional scale. Therefore, exploring the application potential of GF-1 WFV data on land surface LAI estimation and developing the specific LAI estimation algorithm for GF-1 WFV data are urgently needed.
The key of LAI estimation using remote sensing data is how to establish the relationship between LAI and land surface reflectance data according to the radiative transfer process of photon in the vegetation canopy and its spectral response characteristics [20
]. Many algorithms have been developed to retrieve LAI using satellite remote sensing data and generally two types of algorithms can be distinguished including empirical methods and physical model based methods [6
]. The empirical LAI estimation methods are based on the statistical relationships between LAI and vegetation indices (VIs), which are calibrated using field LAI measurements and remote sensing data or simulated data from canopy radiative transfer models [28
]. The empirical methods simplify the complex radiative transfer process in the canopy, which are simple and can obtain satisfactory LAI estimation accuracy in small regions. Furthermore, when using VIs as independent variables, the VIs can highlight vegetation information and weaken the influences of canopy shadows, soil backgrounds, atmospheric conditions and angle effects [29
]. However, the empirical methods only use the VIs calculated from several band reflectances, which cannot make full use of the multi-band spectral information of remote sensing data. Furthermore, the reduction of the VIs from multi-band reflectances to an index also reduces the LAI estimation constraints and increases the uncertainty of LAI estimation results, resulting in the empirical relationship between LAI and VIs changing with sensor types, vegetation types, time and geographical areas. Therefore, the application of empirical LAI estimation methods to a large scale is very difficult, due to the complexity of land surface.
The physical model based methods are mainly based on the simulation of radiative transfer in vegetation canopy and establishing the physical relationship between canopy reflectances and LAI, and then the canopy radiative transfer models are inversed [30
]. The direct inversion of radiative transfer models is very difficult due to the complexity of the models, and iterative optimization (OPT) method [32
], lookup table (LUT) method [6
] and machine learning methods [26
] are usually used for indirect inversion of physical models to achieve LAI estimation. The OPT method is based on the iterative minimization of a cost function which requires hundreds of runs of the canopy radiative transfer model for each pixel and therefore computationally too demanding. For practical applications, LUT method and machine learning methods are popular LAI estimation methods, which are based on the database simulated by physical models. The LUT method is conceptually the simplest technique by finding the solution for a given set of reflectance measurements, which consisted of selecting the closest cases in the database according to a cost function, and then based on extracting the corresponding set of LAI [6
]. However, the LUT method usually requires a fixed number of input variables, unless there are very large lookup tables, which are difficult to manipulate. Machine learning methods can efficiently and accurately approximate the complex nonlinear functions, and train the algorithm parameters through training samples to realize the efficient and accurate estimation of LAI using remote sensing data [26
]. Commonly used machine learning methods mainly include neural networks (NNs), as well as support vector machines and decision trees [35
]. NNs trained over radiative transfer model simulations have been applied, with success, to estimate LAI from several sensors’ data, leading to several operational LAI production algorithms, such as the CYCLOPES and MERIS LAI products [26
]. Therefore, based on the reality of work in the field of LAI estimation using remote sensing data, the NNs inversion of physical models is a potential and reasonable choice for LAI estimation from GF-1WFV data.
Therefore, the objective of this study is to develop a general LAI estimation algorithm for GF-1 WFV reflectance data under various land surface conditions using NNs trained over radiative transfer model simulations. Meanwhile, the field LAI measurements in an agriculture region with maize as the dominated crop type are used to assess the performance of GF-1 WFV data on LAI estimation using the proposed algorithm.
The LAI estimation results using the proposed algorithm and the pre-processed GF-1 WFV land surface reflectance data are shown in Figure 5
. In the visual aspect, the spatial and temporal distributions of LAI estimates are reasonable. As for the spatial distribution, LAI estimates using GF-1 WFV data acquired on 27 June generally show small values, because maize at this period is at the early emergence stage with small LAI, while larger LAI estimates only present at an orchard on the upper left corner of the study area. After July, with the rapid growth of maize, the large LAI estimates are mainly distributed in the area of farmland, and the LAI estimates at other areas are smaller. Meanwhile, LAI estimates in the cloud and cloud shadow regions are all zero. Regarding the temporal variations, LAI estimates of maize show low values in late June and continuous increases in July and mid-August, reaching the peak LAI values in the late August and a downward trend in mid-September. The temporal variation characteristics of LAI estimates exactly reflect the growth processes of maize from germination, rapid growth to peak value, and then gradually decline, which is in accord with the growth characteristics of maize. Therefore, the visual observation of the spatial and temporal characteristics of LAI estimates using GF-1 WFV data could preliminarily indicate the reasonability of the proposed LAI estimation algorithm.
The field LAI measurements located in the cloud and cloud shadow regions were deleted based on the visual observation, and the remaining field LAI measurements, which contained 43 LAI data, were used to directly validate the LAI estimates from GF-1 WFV data using the proposed algorithm (Figure 6
). It could be seen that there was a good linear relationship between field LAI measurements and LAI estimates from GF-1 WFV data. The LAI estimation accuracy was satisfactory with values of R2
, RMSE, RRMSE, CRM, and EF equaling 0.818%, 0.5%, 19.0%, 0.04% and 0.797%, respectively. In addition, some larger differences between the field survey LAI and the predicted LAI were observed in the region of large LAI values. The main reason might be that the field LAI values measured on 5 September were matched with LAI estimates interpolated from LAI estimates on 24 August and 18 September, whereas maize in this period was close to the mature stage with great LAI change rate, thus the interpolated LAI results might have some uncertainties. The other reason might be that the clouds influenced the atmospheric conditions of the study area which further influenced the accuracy of obtaining GF-1 WFV land surface reflectances and caused the uncertainties of LAI estimates.
In summary, the difference between the field LAI measurements and LAI estimates from GF-1 WFV data was small and the LAI estimation results were satisfactory. Therefore, the results indicated that the proposed LAI estimation algorithm for GF-1 WFV data was reliable and GF-1 WFV data could achieve acceptable performance on LAI estimation, which had the potential for providing high temporal and spatial resolution LAI dataset for related applications, such as agriculture monitoring.
6. Discussion and Conclusions
This study proposed a LAI estimation algorithm for GF-1 WFV land surface reflectance data based on BPNNs with training samples generated from the radiative transfer model simulations under different soil and vegetation conditions, thus the algorithm could adapt to a variety of different underlying conditions. The validation results using field LAI measurements in an agriculture region with maize as the dominated crop type showed that the proposed algorithm could achieve satisfactory LAI estimation accuracy (e.g., RMSE = 0.50), which indicated GF-1 WFV data had good performance on LAI estimation and the algorithm had potential to operationally estimate LAI from GF-1 WFV land surface reflectances. The LAI estimation algorithm was automatically operated without prior knowledge on land cover, and no human interaction and no empirical model parameters were needed. Therefore, the proposed LAI estimation algorithm for GF-1 WFV data could overcome the difficulties in determination of model parameters for empirical LAI estimation methods, which were generally changed with time, region and vegetation types. In addition, the studies of LAI estimation using remote sensing data based on NNs trained over radiative transfer model simulations in the past mainly focused on the kilometric spatial resolution remote sensing data, and several global LAI products were generated from SPOT VEGETATION, MERIS and MODIS data. Therefore, the developed LAI estimation algorithm for decametric spatial resolution remote sensing data using NNs based on physical model was a new try in this study, and LAI with decametric spatial resolution would be more useful for agriculture, ecosystem and environment management than kilometric spatial resolution data which were usually larger than the typical scales of most landscapes.
There were also some limitations about this study, even though it achieved satisfactory LAI estimation performance. Firstly, the field LAI measurements for validation were collected from an agriculture region covering the main growing season of maize, though they had temporal–spatial representativeness of cropland and the agriculture monitoring was the main application field of LAI, exploratory investigations should be strengthened using much more validation data from various vegetation types in the future work. Moreover, the pre-processes of GF-1 WFV data were conducted manually. If an automated data preprocessing approach for radiation correction and geometric correction could be developed, a streamlined LAI estimation workflow would be formed from the original DN values to LAI estimates, which would make the LAI estimation from GF-1 WFV data much simpler and quicker. Finally, the accuracy assessment of LAI estimation algorithm was difficult to achieve due to the difficulties in obtaining real LAI values from field measurements to match with the pixel level LAI estimates. Currently, using plant canopy analyzer to obtain field LAI measurements was the commonly used method, which could obtain relatively accurate LAI data. However, some errors might be made when matching the field LAI measurements and pixel scale LAI estimates due to the non-absolute uniformity of the land surface, though averaging the multi-point measurements could reduce this error to a certain extent. Perhaps using an unmanned aircraft system flying in the lower altitude to take photos from the nadir and developing corresponding LAI extraction algorithm from photos was a considerable way for real LAI data collection. Although the field LAI measurements had many uncertainties, they still played an important role in assessing LAI estimation algorithms using remote sensing data, and more effective real LAI obtaining approaches would be expected in the future.
In conclusion, the proposed LAI estimation algorithm for GF-1 WFV data was reliable and GF-1 WFV data were confirmed as having a good performance on LAI estimation, which could provide reliable high spatial and temporal resolution LAI data for related applications. Further work would focus on validating the proposed LAI estimation algorithm using field LAI measurements with less uncertainties under various land cover type conditions.