(Liu et al., 2017)：A synthetic learning methodology was developed in this study to estimate GAI from ground-based LiDAR observations over wheat .…… This approach allows circumventing the limits of training the neural network over actual experimental measurements that are prone to uncertainties and are limited by the number and range of cases that can be sampled. Another important advantage of the method proposed is that it takes explicitly into account the sensor specifications to simulate the measurements asit is actually acquired. …………
(Yang et al., 2022)：In this paper, we combined high spatial resolution UAS imagery with high spectral resolution imaging spectroscopy (IS) data from AVIRIS-NG and developed robust machine learning models to upscale and estimate the fractional cover (FCover) of key Arctic PFTs in the Seward Peninsula, western Alaska. We show that using this multi-platform scaling approach, the FCover of 12 dominant Arctic tundra PFTs can be accurately mapped with IS data from AVIRIS-NG (RMSE 0.14–0.22) and outperformed traditional spectral unmixing analysis (RMSE: 0.22–0.31). In particular, high-resolution UAS data provided an important avenue to obtain the quantitative training samples necessary to develop accurate scaling models. …………
(Macfarlane et al., 2011)：The main conclusion of this study is that the method used to classify mixed pixels in canopy images is of little importance, once homogeneous regions of sky and canopy have been identified. Previous studies have often tested a single image classification metho. but even studies that have compared multiple methods have relied heavily on subjective visual inspection of images to rate the performance of methods although in the latter study an attempt was also made to ‘score’ methods based on more objective criteria. …………
(Calders et al., 2018)：Within this paper, we compared three different ground-based instruments and quantified aspects of their measurement uncertainty. Similar to Woodgate et al. (2015b), Ryu et al. (2010b), our results indicated that the agreement between these instruments does not meet the 5% accuracy specified by WMO (2012). Ryu et al. (2010b) also reports higher estimates from LAI-2x00 measurements compared to DHP in savanna ecosystems. …………
(Hadi et al., 2017)：Despite our best efforts in providing ecologically realistic range of the inputs—on which results of sensitivity analysis are dependent—used in the BRF models, the difficulty, and consequently lack of available spectral measurements in humid tropical forest of South East Asia caused unavoidable uncertainties in our modelling experiments. …………
(Tang et al., 2019)：4.2. Estimation error and algorithm scales. In some cases, the estimation errors of remote sensing systems can contribute more than the definitional difference to the observed cross product comparisons. For example, errors in the Landsat-based tree cover datasets were clearly larger than those in the lidar datasets, and all three Landsat products had only moderate agreement with each other. The GLCF TC product used the multi-year MODIS VCF data as training data. …………
(Liu et al., 2017)：The success of the method will highly bear on the realism and plasticity of the 3D canopy model used. The inter-plant variability is simulated stochastically in the current version of ADEL-Wheat model, neglecting the possible interactions between neighboring individuals. ……. Finally, the training database used to relate the LiDAR 3D point cloud features to GAI should represent well the possible variability of canopy structure, including the range of genotypic and environmental conditions.…… Nevertheless, the results of the available experiments should be organized to facilitate their compilation for the generation of the prior distributions and co-distributions of the derived model parameters.
(Zhang et al., 2023)：4.3. Future improvements of SIFtotal and GPPSIF. It has been shown that the relationship between SIF and GPP can become non-linear at the satellite and canopy levels under a wide range of conditions but in particular at finer scales. Clearly, consideration of the spatial and temporal scale dependencies of different SIF data is critical to assimilate the information embedded in SIF and infer GPP. …………
(Brüllhardt et al., 2020)：4.4. Potential areas of application. Applications across a wide range of questions benefit from spatially distinct data on radiation regimes. These span microclimate mapping over larger scales to the unravelling of patterns and interrelations of plant growth and to the improvement of ecological niche models. …………Based on our findings, the applicability of light estimation from synthetic hemispherical pictures opens the possibility to use low cost aerial photography instead of often more expensive LiDAR. ……
当然，这四个要素并不一定全部包含，有的论文中仅对(2)和(3)进行讨论，这取决于作者的习惯与研究内容的可讨论性。 Ecarnot et al., (2015) 也对讨论的写法进行了总结，我觉得还是挺不错的，可以借鉴一下。
(Richardson et al., 2009)：The present study investigated the applicability of various models to estimate effective LAI from aerial discrete-return LIDAR using a unique data set of ground-based LAI collected from a heterogeneous forest with a large range of LAI.
(Hofle et al., 2007)：In this paper two independent methods for correcting airborne laser scanning intensities were presented. The first approach – data-driven method – performs a least squares adjustment for a given empirical model including intensity and range, …………The second approach – model-driven method – is derived from the radar equation, which describes the loss of emitted pulse power.
(Beland et al., 2011)：The level of agreement between the leaf area estimates obtained from leaf harvesting and the TLS measurements shows the great potential value of TLSs as stand-alone tools for measuring spatially explicit individual tree leaf area with a high level of detail. The differences found suggest that for some crown shapes a 2-scan configuration may cause errors related to grazing of the leaves by the laser beams (which then generate very low intensity returns), and that the main limitation in the case of high LAI trees is occlusion effects (i.e. limited beam penetration inside the crown).
(Kayad et al., 2022)：The main findings can be summarized as follows: 1. PROSAIL model inversion was capable to retrieve maize LAI, presenting reasonable R2 value (0.5) and considerably low RMSE value (0.8); 2. maize yield estimation showed better performance at the late vegetative growth stages (V16), around 73 DAS; 3. maize biomass and grain yield estimations through NDRE showed higher performance in comparison with retrieved LAI, GNDVI and NDVI, when applied to an independent dataset.
(Aasen et al., 2015)：The described method allows the gathering of complementary hyperspectral and 3D information with only one single lightweight UAV imaging system. ………… Additionally, we hope that the described way of embedded quality assurance information will help to establish best practice procedures for small lightweight sensors to make UAV sensing systems a reliable source for remote sensing data.
(Chianucci et al., 2016)：Among the canopy attributes, the method allows objective evaluation of forest canopy cover, without requiring additional parameters (e.g., foliage projection coefficient, leaf angle distribution); while this option holds great potential for forest inventory purposes, the high-resolution canopy cover estimates obtainable from UAV platforms could also be used for calibrating metrics obtained from coarser-scale remote sensing products and/or analyses that use morphological processing (rather than relying only on vegetation indices), avoiding the need of ground measurements.
