A Review of Hybrid Approaches for Quantitative Assessment of Crop Traits Using Optical Remote Sensing: Research Trends and Future Directions

Remote sensing technology allows to provide information about biochemical and biophysical crop traits and monitor their spatiotemporal dynamics of agriculture ecosystems. Among multiple retrieval techniques, hybrid approaches have been found to provide outstanding accuracy, for instance, for the inference of leaf area index (LAI), fractional vegetation cover (fCover), and leaf and canopy chlorophyll content (LCC and CCC). The combination of radiative transfer models (RTMs) and data-driven models creates an advantage in the use of hybrid methods. Through this review paper, we aim to provide state-of-the-art hybrid retrieval schemes and theoretical frameworks. To achieve this, we reviewed and systematically analyzed publications over the past 22 years. We identified two hybrid-based parametric and hybrid-based nonparametric regression models and evaluated their performance for each variable of interest. From the results of our extensive literature survey, most research directions are now moving towards combining RTM and machine learning (ML) methods in a symbiotic manner. In particular, the development of ML will open up new ways to integrate innovative approaches such as integrating shallow or deep neural networks with RTM using remote sensing data to reduce errors in crop trait estimations and improve control of crop growth conditions in very large areas serving precision agriculture applications.

Impact of Modeling Abstractions When Estimating Leaf Mass per Area and Equivalent Water Thickness over Sparse Forests Using a Hybrid Method

Equivalent water thickness (EWT) and leaf mass per area (LMA) are important indicators of plant processes, such as photosynthetic and potential growth rates and health status, and are also important variables for fire risk assessment. Retrieving these traits through remote sensing is challenging and often requires calibration with in situ measurements to provide acceptable results. However, calibration data cannot be expected to be available at the operational level when estimating EWT and LMA over large regions. In this study, we assessed the ability of a hybrid retrieval method, consisting of training a random forest regressor (RFR) over the outputs of the discrete anisotropic radiative transfer (DART) model, to yield accurate EWT and LMA estimates depending on the scene modeling within DART and the spectral interval considered. We show that canopy abstractions mostly affect crown reflectance over the 0.75–1.3 μm range. It was observed that excluding these wavelengths when training the RFR resulted in the abstraction level having no effect on the subsequent LMA estimates (RMSE of 0.0019 g/cm2 for both the detailed and abstract models), and EWT estimates were not affected by the level of abstraction. Over AVIRIS-Next Generation images, we showed that the hybrid method trained with a simplified scene obtained accuracies (RMSE of 0.0029 and 0.0028 g/cm2 for LMA and EWT) consistent with what had been obtained from the test dataset of the calibration phase (RMSE of 0.0031 and 0.0032 g/cm2 for LMA and EWT), and the result yielded spatially coherent maps. The results demonstrate that, provided an appropriate spectral domain is used, the uncertainties inherent to the abstract modeling of tree crowns within an RTM do not significantly affect EWT and LMA accuracy estimates when tree crowns can be identified in the images.

Joint Use of PROSAIL and DART for Fast LUT Building: Application to Gap Fraction and Leaf Biochemistry Estimations over Sparse Oak Stands

Gap Fraction, leaf pigment contents (content of chlorophylls a and b (Cab) and carotenoids content (Car)), Leaf Mass per Area (LMA), and Equivalent Water Thickness (EWT) are considered relevant indicators of forests’ health status, influencing many biological and physical processes. Various methods exist to estimate these variables, often relying on the extensive use of Radiation Transfer Models (RTMs). While 3D RTMs are more realistic to model open canopies, their complexity leads to important computation times that limit the number of simulations that can be considered; 1D RTMs, although less realistic, are also less computationally expensive. We investigated the possibility to approximate the outputs of a 3D RTM (DART) from a 1D RTM (PROSAIL) to generate in very short time numerous extensive Look-Up Tables (LUTs). The intrinsic error of the approximation model was evaluated through comparison with DART reference values. The model was then used to generate LUTs used to estimate Gap Fraction, Cab, Car, EWT, and LMA of Blue Oak-dominant stands in a woodland savanna from AVIRIS-C data. Performances of the approximation model for estimation purposes compared to DART were evaluated using Wilmott’s index of agreement (dr), and estimation accuracy was measured with coefficients of determination (R2) and Root Mean Squared Error (RMSE). The low approximation error of the proposed model demonstrated that the model could be considered for canopy covers as low as 30%. Gap Fraction estimations presented similar performances with either DART or the approximation (dr 0.78 and 0.77, respectively), while Cab and Car showed improved performances (dr increasing from 0.65 to 0.77 and 0.34 to 0.65, respectively). No satisfying estimation methods were found for LMA and EWT using either models, probably due to the high sensitivity of the scene’s reflectance to Gap Fraction and soil modeling at such low LAI. Overall, estimations using the approximated reflectances presented either similar or improved accuracy. Our findings show that it is possible to approximate DART reflectances from PROSAIL using a minimal number of DART outputs for calibration purposes, drastically reducing computation times to generate reflectance databases: 300,000 entries could be generated in 1.5 h, compared to the 12,666 total CPU hours necessary to generate the 21,840 calibration entries with DART.

A Refined Four-Stream Radiative Transfer Model for Row-Planted Crops

In modeling the canopy reflectance of row-planted crops, neglecting horizontal radiative transfer may lead to an inaccurate representation of vegetation energy balance and further cause uncertainty in the simulation of canopy reflectance at larger viewing zenith angles. To reduce this systematic deviation, here we refined the four-stream radiative transfer equations by considering horizontal radiation through the lateral “walls”, considered the radiative transfer between rows, then proposed a modified four-stream (MFS) radiative transfer model using single and multiple scattering. We validated the MFS model using both computer simulations and in situ measurements, and found that the MFS model can be used to simulate crop canopy reflectance at different growth stages with an accuracy comparable to the computer simulations (RMSE < 0.002 in the red band, RMSE < 0.019 in NIR band). Moreover, the MFS model can be successfully used to simulate the reflectance of continuous (RMSE = 0.012) and row crop canopies (RMSE < 0.023), and therefore addressed the large viewing zenith angle problems in the previous row model based on four-stream radiative transfer equations. Our results demonstrate that horizontal radiation is an important factor that needs to be considered in modeling the canopy reflectance of row-planted crops. Hence, the refined four-stream radiative transfer model is applicable to the real world.

Sensitivity Analysis of the DART Model for Forest Mensuration with Airborne Laser Scanning

Airborne Laser Scanning (ALS) measurements are increasingly vital in forest management and national forest inventories. Despite the growing reliance on ALS data, comparatively little research has examined the sensitivity of ALS measurements to varying survey conditions over commercially important forests. This study investigated: (i) how accurately the Discrete Anisotropic Radiative Transfer (DART) model was able to replicate small-footprint ALS measurements collected over Irish conifer plantations, and (ii) how survey characteristics influenced the precision of discrete-return metrics. A variance-based global sensitivity analysis demonstrated that discrete-return height distributions were accurately and consistently simulated across 100 forest inventory plots with few perturbations induced by varying acquisition parameters or ground topography. In contrast, discrete return density, canopy cover and the proportion of multiple returns were sensitive to fluctuations in sensor altitude, scanning angle, pulse repetition frequency and pulse duration. Our findings corroborate previous studies indicating that discrete-return heights are robust to varying acquisition parameters and may be reliable predictors for the indirect retrieval of forest inventory measurements. However, canopy cover and density metrics are only comparable for ALS data collected under similar acquisition conditions, precluding their universal use across different ALS surveys. Our study demonstrates that DART is a robust model for simulating discrete-return measurements over structurally complex forests; however, the replication of foliage morphology, density and orientation are important considerations for radiative transfer simulations using synthetic trees with explicitly defined crown architectures.

Evaluation of the Snow Albedo Retrieved from the Snow Kernel Improved the Ross-Roujean BRDF Model

The original kernel-driven bidirectional reflectance distribution function (BRDF) models were developed based on soil-vegetation systems. To further improve the ability of the models to characterize the snow surface scattering properties, a snow kernel was derived from the asymptotic radiative transfer (ART) model and used in the kernel-driven BRDF model framework. However, there is a need to further evaluate the influence of using this snow kernel to improve the original kernel-driven models in snow albedo retrieval applications. The aim of this study is to perform such an evaluation using a variety of snow BRDF data. The RossThick-Roujean (RTR) model is used as a framework for taking in the new snow kernel (hereafter named the RTS model) since the Roujean geometric-optical (GO) kernel captures a neglectable hotspot effect and represents a more prominent dome-shaped BRDF, especially at a small solar zenith angle (SZA). We obtained the following results: (1) The RTR model has difficulties in reconstructing the snow BRDF shape, especially at large SZAs, which tends to underestimate the reflectance in the forward direction and overestimate reflectance in the backward direction for various data sources. In comparison, the RTS model performs very well in fitting snow BRDF data and shows high accuracy for all data. (2) The RTR model retrieved snow albedos at SZAs = 30°–70° are underestimated by 0.71% and 0.69% in the red and near-infrared (NIR) bands, respectively, compared with the simulation results of the bicontinuous photon tracking (bic-PT) model, which serve as “real” values. However, the albedo retrieved by the RTS model is significantly improved and generally agrees well with the simulation results of the bic-PT model, although the improved model still somewhat underestimates the albedo by 0.01% in the red band and overestimates the albedo by 0.05% in the NIR band, respectively, at SZAs = 30°–70°, which may be negligible. (3) The albedo derived by these two models shows a high correlation (R2 > 0.9) between the field-measured and Polarization and Directionality of the Earth's Reflectances (POLDER) data, especially for the black-sky albedo. However, the albedo derived using the RTR model is significantly underestimated compared with the RTS model. The RTR model underestimates the black-sky albedo (white-sky albedo) retrievals by 0.62% (1.51%) and 0.93% (2.08%) in the red and NIR bands, respectively, for the field-measured data. The shortwave black-sky and white-sky albedos derived using the RTR model for the POLDER data are underestimated by 1.43% and 1.54%, respectively, compared with the RTS model. These results indicate that the snow kernel in the kernel-driven BRDF model frame is more accurate in snow albedo retrievals and has the potential for application in the field of the regional and global energy budget.

Spatiotemporal Image Fusion in Remote Sensing

In this paper, we discuss spatiotemporal data fusion methods in remote sensing. These methods fuse temporally sparse fine-resolution images with temporally dense coarse-resolution images. This review reveals that existing spatiotemporal data fusion methods are mainly dedicated to blending optical images. There is a limited number of studies focusing on fusing microwave data, or on fusing microwave and optical images in order to address the problem of gaps in the optical data caused by the presence of clouds. Therefore, future efforts are required to develop spatiotemporal data fusion methods flexible enough to accomplish different data fusion tasks under different environmental conditions and using different sensors data as input. The review shows that additional investigations are required to account for temporal changes occurring during the observation period when predicting spectral reflectance values at a fine scale in space and time. More sophisticated machine learning methods such as convolutional neural network (CNN) represent a promising solution for spatiotemporal fusion, especially due to their capability to fuse images with different spectral values.

Biological processes dominate seasonality of remotely sensed canopy greenness in an Amazon evergreen forest

Satellite observations of Amazon forests show seasonal and interannual variations, but the underlying biological processes remain debated. Here we combined radiative transfer models (RTMs) with field observations of Amazon forest leaf and canopy characteristics to test three hypotheses for satellite-observed canopy reflectance seasonality: seasonal changes in leaf area index, in canopy-surface leafless crown fraction and/or in leaf demography. Canopy RTMs (PROSAIL and FLiES), driven by these three factors combined, simulated satellite-observed seasonal patterns well, explaining c. 70% of the variability in a key reflectance-based vegetation index (MAIAC EVI, which removes artifacts that would otherwise arise from clouds/aerosols and sun-sensor geometry). Leaf area index, leafless crown fraction and leaf demography independently accounted for 1, 33 and 66% of FLiES-simulated EVI seasonality, respectively. These factors also strongly influenced modeled near-infrared (NIR) reflectance, explaining why both modeled and observed EVI, which is especially sensitive to NIR, captures canopy seasonal dynamics well. Our improved analysis of canopy-scale biophysics rules out satellite artifacts as significant causes of satellite-observed seasonal patterns at this site, implying that aggregated phenology explains the larger scale remotely observed patterns. This work significantly reconciles current controversies about satellite-detected Amazon phenology, and improves our use of satellite observations to study climate-phenology relationships in the tropics.

Sand dune ridge alignment effects on surface BRF over the Libya-4 CEOS calibration site

The Libya-4 desert area, located in the Great Sand Sea, is one of the most important bright desert CEOS pseudo-invariant calibration sites by its size and radiometric stability. This site is intensively used for radiometer drift monitoring, sensor intercalibration and as an absolute calibration reference based on simulated radiances traceable to the SI standard. The Libya-4 morphology is composed of oriented sand dunes shaped by dominant winds. The effects of sand dune spatial organization on the surface bidirectional reflectance factor is analyzed in this paper using Raytran, a 3D radiative transfer model. The topography is characterized with the 30 m resolution ASTER digital elevation model. Four different regions-of-interest sizes, ranging from 10 km up to 100 km, are analyzed. Results show that sand dunes generate more backscattering than forward scattering at the surface. The mean surface reflectance averaged over different viewing and illumination angles is pretty much independent of the size of the selected area, though the standard deviation differs. Sun azimuth position has an effect on the surface reflectance field, which is more pronounced for high Sun zenith angles. Such 3D azimuthal effects should be taken into account to decrease the simulated radiance uncertainty over Libya-4 below 3% for wavelengths larger than 600 nm.