A Survey of Active Learning for Quantifying Vegetation Traits from Terrestrial Earth Observation Data

The leaf chlorophyll content (LCC) is a critical index to characterize crop growth conditions, photosynthetic capacity, and physiological status. Its dynamic change characteristics are of great significance for monitoring crop growth conditions and understanding the process of material and energy exchange between crops and the environment. Extensive research has focused on LCC retrieval with hyperspectral data onboard various sensor platforms. Nevertheless, limited attention has been paid to LCC inversion from multispectral data, such as the data from Landsat-8, and the potentials and capabilities of the data for crop LCC estimation have not been fully explored. The present study made use of Landsat-8 Operational Land Imager (OLI) imagery and the corresponding field experimental data to evaluate their capabilities and potentials for LCC modeling using four different retrieval methods: vegetation indices (VIs), machine learning regression algorithms (MLRAs), lookup-table (LUT)-based inversion, and hybrid regression approaches. The results showed that the modified triangular vegetation index (MTVI2) exhibited the best estimate accuracy for LCC retrieval with a root mean square error (RMSE) of 5.99 μg/cm2 and a relative RMSE (RRMSE) of 10.49%. Several other vegetation indices that were established from red and near-infrared (NIR) bands also exhibited good accuracy. Models established from Gaussian process regression (GPR) achieved the highest accuracy for LCC retrieval (RMSE = 5.50 μg/cm2, RRMSE = 9.62%) compared with other MLRAs. Moreover, red and NIR bands outweighed other bands in terms of GPR modelling. LUT-based inversion methods with the “K(x) = −log (x) + x” cost function that belongs to the “minimum contrast estimates” family showed the best estimation results (RMSE = 8.08 μg/cm2, RRMSE = 14.14%), and the addition of multiple solution regularization strategies effectively improved the inversion accuracy. For hybrid regression methods, the use of active learning (AL) techniques together with GPR for LCC modelling significantly increased the estimation accuracy, and the combination of entropy query by bagging (EQB) AL and GPR had the best accuracy for LCC estimation (RMSE = 12.43 μg/cm2, RRMSE = 21.77%). Overall, our study suggest that Landsat-8 OLI data are suitable for crop LCC retrieval and could provide a basis for LCC estimation with similar multispectral datasets.

<p>Optical Earth observation satellites provide spatially-explicit data that are necessary to study trends in vegetation dynamics. However, more of often than not optical data are discontinuous in time, due to persistent cloud cover and instrumental noises. Hence, the operating constraints of these data require several essential pre-processing steps, especially when aiming to reach towards monitoring of vegetation seasonal trends.  To facilitate this task, here we present an end-to-end processing software framework applied to Sentinel-2 images.</p><p>To do so, first biophysical retrieval models were generated by means of a trained machine learning regression algorithm (MLRA) using simulated data coming from radiative transfer models. Among various tested MLRAs, the variational heteroscedastic Gaussian process regression (VHGPR) was evaluated as best performing. to train the retrieval model.  The training and retrieval were conducted in the Automated Radiative Transfer Models Operator (ARTMO) software framework.</p><p>Subsequently, in view of retrieving the phenological parameters from the obtained vegetation products, a novel times series toolbox as part of the ARTMO framework was used, called:  Decomposition and Analysis of Time Series software (DATimeS). DATimeS provides temporal interpolation among other functionalities with several advanced MLRAs for gap filling, smoothing functions and subsequent calculation of phenology indicators. Various MLRAs were tested for gap filling to reconstruct cloud-free maps of biophysical variables at a step of 10 days.</p><p>A demonstration case is presented involving the retrieval of Leaf area index (LAI), fraction of Absorbed Photosynthetically Active Radiation (FAPAR) from sentinel-2 time series.  A large agricultural Algerian site of 143, 75 km² including Oued Rhiou, Ouarizane, Djidioua (1,345,075 pixels) was chosen for this study.  A reference image was excluded from the time series in order to evaluate the reconstruction accuracy over a 40-day artificial gap.</p><p>  The reference vs.  Reconstructed maps produced by the gap-filling methods were compared with statistical goodness-of-fit metrics.  Considering both accuracy and processing speed, the fitting algorithms Gaussian process regression (GPR) and Next neighbour interpolation (R²= 0.90 / 0.081 sec per pixel and R²=0.88 / 0.001 sec per pixel respectively) interpolations proved to reconstruct the vegetation products the most efficient, with GPR as more accurate but Next faster by a factor of 70.</p><p>Finally, we evaluated of the phenology indicators such as start-of-season and end-of-season based on LAI and FAPAR. The obtained maps provide valid information of the vegetation dynamics.  Altogether, the ARTMO-DATimeS software framework enabled seamless processing of all essential steps:  (1) from L2A sentinel-2 images converted to vegetation products, (2) to cloud-free composite products, and finally (3) converted into vegetation phenology indicators.</p><p> </p>

In this study, we reviewed methods, sensors and study areas in estimation of vegetation biophysical (VBph) and biochemical (VBch) parameters using satellite optical remote sensing, in terms of (1) their performance and (2) extent of their use. Performance of parametric regression methods is dependent on the type of sensor, vegetation parameter, vegetation cover and biome of study area. Artificial neural network, partial least-squares regression, random forest, support vector machine and Gaussian process regression (GPR) had acceptable performances in the estimations. Performance of physically based methods depends on selecting appropriate radiative transfer model, inversion method and cost function, optimal solution (e.g. mean of some of the best lookup-table solutions) and adding appropriate Gaussian noise. In recent years, the use of active learning techniques (especially Euclidean distance-based diversity), to optimize training datasets against field datasets, with machine learning regression algorithms (MLRAs) (especially GPR) has led to acceptable performances by hybrid methods in estimation of VBph and VBch parameters using real and/or simulated (Re/Si) satellite hyperspectral (and multispectral) data with dimensionality reduction techniques in the spectral domain (especially principal component analysis with 20 principal components). In overall, Re/Si satellite optical sensors data containing red-edge (mostly for VBch parameters) and/or short-wave infrared (mostly for VBph parameters) spectral band(s) performed well. Emerging trend observed in recent years in the estimations by using Re/Si medium spatial resolution sensors data with relatively good spectral resolutions (e.g. precursore iperspettrale della missione applicativa, environmental mapping and analysis program and Sentinel-2 multispectral instrument) and non-parametric regression (especially MLRAs) and hybrid methods.

The unprecedented availability of optical satellite data in cloud-based computing platforms, such as Google Earth Engine (GEE), opens new possibilities to develop crop trait retrieval models from the local to the planetary scale. Hybrid retrieval models are of interest to run in these platforms as they combine the advantages of physically-based radiative transfer models (RTM) with the flexibility of machine learning regression algorithms. Previous research with GEE primarily relied on processing bottom-of-atmosphere (BOA) reflectance data, which requires atmospheric correction. In the present study, we implemented hybrid models directly into GEE for processing Sentinel-2 (S2) Level-1C (L1C) top-of-atmosphere (TOA) reflectance data into crop traits. To achieve this, a training dataset was generated using the leaf-canopy RTM PROSAIL in combination with the atmospheric model 6SV. Gaussian process regression (GPR) retrieval models were then established for eight essential crop traits namely leaf chlorophyll content, leaf water content, leaf dry matter content, fractional vegetation cover, leaf area index (LAI), and upscaled leaf variables (i.e., canopy chlorophyll content, canopy water content and canopy dry matter content). An important pre-requisite for implementation into GEE is that the models are sufficiently light in order to facilitate efficient and fast processing. Successful reduction of the training dataset by 78% was achieved using the active learning technique Euclidean distance-based diversity (EBD). With the EBD-GPR models, highly accurate validation results of LAI and upscaled leaf variables were obtained against in situ field data from the validation study site Munich-North-Isar (MNI), with normalized root mean square errors (NRMSE) from 6% to 13%. Using an independent validation dataset of similar crop types (Italian Grosseto test site), the retrieval models showed moderate to good performances for canopy-level variables, with NRMSE ranging from 14% to 50%, but failed for the leaf-level estimates. Obtained maps over the MNI site were further compared against Sentinel-2 Level 2 Prototype Processor (SL2P) vegetation estimates generated from the ESA Sentinels' Application Platform (SNAP) Biophysical Processor, proving high consistency of both retrievals (R2 from 0.80 to 0.94). Finally, thanks to the seamless GEE processing capability, the TOA-based mapping was applied over the entirety of Germany at 20 m spatial resolution including information about prediction uncertainty. The obtained maps provided confidence of the developed EBD-GPR retrieval models for integration in the GEE framework and national scale mapping from S2-L1C imagery. In summary, the proposed retrieval workflow demonstrates the possibility of routine processing of S2 TOA data into crop traits maps at any place on Earth as required for operational agricultural applications.

Active learning for modeling and prediction of dynamical fluid processes

Physical radiative transfer models (RTMs) of leaf and canopies with sufficient realism enable the retrieval of biophysical variables from imaging spectroscopy through numerical inversion. However, advanced RTMs are computationally intensive, which hampers practical applicability of inversion schemes against remote sensing images. To bypass the computational load such RTMs, it has been proposed to approximate these models by means of statistical learning, i.e. emulation. Here we tested three machine learning regression algorithms, i.e. neural networks, kernel ridge regression and Gaussian processes regression, on their ability to emulate the advanced RTM SCOPE (Soil-Canopy-Observation of Photosynthesis and the Energy balance) for limited set of input variables. The best performing emulator was implemented into a numerical inversion scheme to process a subset of an hyperspectral image into a multitude of vegetation properties. Obtained maps are not only consistent, but also processing time was in the order of minutes - in comparison, by using SCOPE the processing would have taken days.

Regional seismic fragility assessment of bridge portfolios must address the embedded uncertainties and variations stemming from both the earthquake hazard and bridge attributes (e.g., geometry, material, design detail). To achieve bridge‐specific fragility assessment, multivariate probabilistic seismic demand models (PSDM) have recently been developed that use both the ground motion intensity measure and bridge parameters as inputs. However, explicitly utilizing bridge parameters as inputs requires numerous nonlinear response history analyses (NRHAs). In this situation, the associated computational cost increases exponentially for high‐fidelity bridge models with complex component connectivity and sophisticated material constitutive laws. Moreover, it remains unclear how many analyses are sufficient for the response data and the resulting demand model to cover the entire solution space without overfitting. To deal with these issues, this study integrates Gaussian process regression (GPR) and active learning (AL) into a multistep workflow to achieve efficient regional seismic fragility assessment of bridge portfolios. The GPR relaxes the probability distribution assumptions made in typical cloud analysis‐based PSDMs to enable heteroskedastic nonparametric seismic demand modeling. The AL leverages the varying standard deviation to select the least but most representative bridge‐model‐ground‐motion sample pairs to conduct NRHA with much‐improved efficiency. Both independent and correlated multi‐output GPRs are proposed to deal with bridge portfolios with seismic demand correlations among multiple components (column, bearing, shear key, abutment, unseating, and joint seal). Considering a single benchmark highway bridge class in California as the case study, the AL‐GPR framework and the associated component‐level fragility results are investigated in terms of their efficiency, accuracy, and robustness. The fragility results show that 70 AL‐selected samples would enable the GPR to derive bridge‐specific fragility models comparable to the ones using the multiple stripes analysis approach with 1950 ground motions considered for each individual bridge. The AL‐GPR model also successfully captures the physics of how bridge span length, deck area, column slenderness, and steel reinforcement ratio would change the damage state exceedance probabilities of different bridge components. The efficiency of AL stems from the fact that, with the multi‐output independent GPR, a stable and reliable fragility model can be achieved using 50 AL‐selected samples compared to at least 270 randomly chosen samples. The proposed methodology advances the state of the art in enabling more efficient and reliable regional seismic fragility assessment of multi‐component bridge portfolios.

Active Learning (AL) is a methodology from machine learning in which the learner interacts with the data source. In this paper, we investigate application of AL techniques to a new domain: regression problems in performance analysis. For computational systems with many factors, each of which can take on many levels, fixed experiment designs can require many experiments, and can explore the problem space inefficiently. We address these problems with a dynamic, adaptive experiment design, using AL in conjunction with Gaussian Process Regression (GPR). The performance analysis process is seeded with a small number of initial experiments, then GPR provides estimates of regression confidence across the full input space. AL is used to suggest follow-up experiments to run, in general, it will suggest experiments in areas where the GRP model indicates low confidence, and through repeated experiments, the process eventually achieves high confidence throughout the input space. We apply this approach to the problem of estimating performance and energy usage of HPGMG-FE, and create good-quality predictive models for the quantities of interest, with low error and reduced cost, using only a modest number of experiments. Our analysis shows that the error reduction achieved from replacing the basic AL algorithm with a cost-aware algorithm can be significant, reaching up to 38% for the same computational cost of experiments.

We proposed a framework to generate high spatio-temporal resolution leaf area index (LAI) and uncertainty maps based on the integration of LAINet observation system, Consistent Adjustment of the Climatology to Actual Observations (CACAO) method and Gaussian process regression (GPR). LAINet, which is a wireless sensor network based automatic LAI observation instrument, was used to provide temporally continuous field measurements; CACAO, a data blending method, was used to blend the high and low spatial resolution remote sensing observations to obtain high spatio-temporal resolution remote sensing observations synchronous with the field measurements. GPR, a machine learning regression algorithm, was used to upscale the spatially discrete field measurements to spatially explicit LAI maps, and get the concomitant uncertainty maps. The performance of the proposed method was evaluated over a crop site, where seven LAI maps and their accompanying uncertainty maps all with 30 m and 8 days resolutions were generated. Results show that the framework can provide accurate LAI retrievals. In addition, the concomitant uncertainty maps provide insight into the reliability of the LAI retrievals. This paper contributes to precision agriculture and validation activities for coarse resolution LAI products.

Physically-based approaches for estimating Leaf Area Index (LAI) using remote sensing data rely on radiative transfer (RT) models. Currently, many RT models are freely available, but determining the appropriate RT model for LAI retrieval is still problematic. This study aims to evaluate the necessity of RT model selection for LAI retrieval and to propose a retrieval methodology using different RT models for different vegetation types. Both actual experimental observations and RT model simulations were used to conduct the evaluation. Each of them includes needleleaf forests and croplands, which have contrasting structural attributes. The scattering from arbitrarily inclined leaves (SAIL) model and the four-scale model, which are 1D and 3D RT models, respectively, were used to simulate the synthetic test datasets. The experimental test dataset was established through two field campaigns conducted in the Heihe River Basin. The results show that the realistic representation of canopy structure in RT models is very important for LAI retrieval. If an unsuitable RT model is used, then the root mean squared error (RMSE) will increase from 0.43 to 0.60 in croplands and from 0.52 to 0.63 in forests. In addition, an RT model’s potential to retrieve LAI is limited by the availability of a priori information on RT model parameters. 3D RT models require more a priori information, which makes them have poorer generalization capability than 1D models. Therefore, physically-based retrieval algorithms should embed more than one RT model to account for the availability of a priori information and variations in structural attributes among different vegetation types.

Earth observation offers an unprecedented opportunity to monitor intensively cultivated areas providing key support to assess fertilizer needs and crop water uptake. Routinely, vegetation traits mapping can help farmers to monitor plant development along the crop’s phenological cycle, which is particularly relevant for irrigated agricultural areas. The high spatial and temporal resolution of the Sentinel-2 (S2) multispectral instrument leverages the possibility to estimate leaf area index (LAI), canopy chlorophyll content (CCC), and vegetation water content (VWC) from space. Therefore, our study presents a hybrid retrieval workflow combining a physically-based strategy with a machine learning regression algorithm, i.e., Gaussian processes regression, and an active learning technique to estimate LAI, CCC and VWC of irrigated winter wheat. The established hybrid models of the three traits were validated against in-situ data of a wheat campaign in the Bonaerense valley, South of the Buenos Aires Province, Argentina, in the year 2020. We obtained good to highly accurate validation results with LAI: R2 = 0.92, RMSE = 0.43 m2 m−2, CCC: R2 = 0.80, RMSE = 0.27 g m−2 and VWC: R2 = 0.75, RMSE = 416 g m−2. The retrieval models were also applied to a series of S2 images, producing time series along the seasonal cycle, which reflected the effects of fertilizer and irrigation on crop growth. The associated uncertainties along with the obtained maps underlined the robustness of the hybrid retrieval workflow. We conclude that processing S2 imagery with optimised hybrid models allows accurate space-based crop traits mapping over large irrigated areas and thus can support agricultural management decisions.

Limited experimental data remains a key challenge in applying machine learning to drug discovery, particularly for cancer-related targets. In this study, we present a data-efficient active meta-deep learning framework to predict mitogen-activated protein kinase 1 (MAPK1) inhibitors, which are promising candidates for cancer-related therapies. Our approach integrates active learning (AL) with a meta-model that combines four deep architectures: a convolutional neural network, an attention, a graph convolutional network, and a graph neural network-attention, trained on molecular descriptors and graph-based representations. These models generate four probability-based features that feed into an attention-based meta-learner, improving predictive performance by 5.12% in the area under the precision-recall curve (AUPRC) and 5.48% in the Matthews correlation coefficient (MCC) using only 10% of the training data. Among the AL sampling strategies evaluated, entropy sampling showed competitive performance in selecting informative molecules for model improvement. Overall, our framework achieves an AUPRC of 0.835 ± 0.017 and MCC of 0.817 ± 0.017, on par with a traditional training method despite using only 26.7% of the training data. Compared to a conventional random forest model trained on brute-force, a 100% full training set, our approach shows a 10.6% improvement in AUPRC and modest gains in MCC, confirming the effectiveness of the proposed framework. Under severe class imbalance, balanced accuracy steadily increased across AL iterations, reaching values greater than 0.85 at the final iteration for all uncertainty-driven strategies. Molecular docking confirmed successful prioritization of the top four predicted compounds. Evaluation on an external MAPK1 data set demonstrated generalizability, with our approach achieving an AUPRC of 0.818 and an MCC of 0.403, comparable to the independent test set. These results highlight the potential of combining intelligent data selection with deep learning architectures through the meta-model to accelerate predictive performance in data-scarce drug discovery. Scientific contribution: This study contributes a novel, data-efficient active meta-deep learning framework for predicting MAPK1 inhibitors, addressing the challenge of limited experimental data in a cancer-specific target. By integrating AL with a meta-model composed of four deep architectures, the approach significantly enhances the predictive performance using only a fraction of the training data. The framework achieves superior metrics compared to traditional training methods, highlighting its potential to accelerate drug discovery in data-scarce settings.

The assessment of food and industrial crops during harvesting is important to determine the quality and downstream processing requirements, which in turn affect their market value. While machine learning models have been developed for this purpose, their deployment is hindered by the high cost of labelling the crop images to provide data for model training. This study examines the capabilities of semi-supervised and active learning to minimise effort when labelling cotton lint samples while maintaining high classification accuracy. Random forest classification models were developed using supervised learning, semi-supervised learning, and active learning to determine Egyptian cotton grade. Compared to supervised learning (80.20–82.66%) and semi-supervised learning (81.39–85.26%), active learning models were able to achieve higher accuracy (82.85–85.33%) with up to 46.4% reduction in the volume of labelled data required. The primary obstacle when using machine learning for Egyptian cotton grading is the time required for labelling cotton lint samples. However, by applying active learning, this study successfully decreased the time needed from 422.5 to 177.5 min. The findings of this study demonstrate that active learning is a promising approach for developing accurate and efficient machine learning models for grading food and industrial crops.

In preparation for new-generation imaging spectrometer missions and the accompanying unprecedented inflow of hyperspectral data, optimized models are needed to generate vegetation traits routinely. Hybrid models, combining radiative transfer models with machine learning algorithms, are preferred, however, dealing with spectral collinearity imposes an additional challenge. In this study, we analyzed two spectral dimensionality reduction methods: principal component analysis (PCA) and band ranking (BR), embedded in a hybrid workflow for the retrieval of specific leaf area (SLA), leaf area index (LAI), canopy water content (CWC), canopy chlorophyll content (CCC), the fraction of absorbed photosynthetic active radiation (FAPAR), and fractional vegetation cover (FVC). The SCOPE model was used to simulate training data sets, which were optimized with active learning. Gaussian process regression (GPR) algorithms were trained over the simulations to obtain trait-specific models. The inclusion of PCA and BR with 20 features led to the so-called GPR-20PCA and GPR-20BR models. The 20PCA models encompassed over 99.95% cumulative variance of the full spectral data, while the GPR-20BR models were based on the 20 most sensitive bands. Validation against in situ data obtained moderate to optimal results with normalized root mean squared error (NRMSE) from 13.9% (CWC) to 22.3% (CCC) for GPR-20PCA models, and NRMSE from 19.6% (CWC) to 29.1% (SLA) for GPR-20BR models. Overall, the GPR-20PCA slightly outperformed the GPR-20BR models for all six variables. To demonstrate mapping capabilities, both models were tested on a PRecursore IperSpettrale della Missione Applicativa (PRISMA) scene, spectrally resampled to Copernicus Hyperspectral Imaging Mission for the Environment (CHIME), over an agricultural test site (Jolanda di Savoia, Italy). The two strategies obtained plausible spatial patterns, and consistency between the two models was highest for FVC and LAI (R2 = 0.91, R2 = 0.86) and lowest for SLA mapping (R2 = 0.53). From these findings, we recommend implementing GPR-20PCA models as the most efficient strategy for the retrieval of multiple crop traits from hyperspectral data streams. Hence, this workflow will support and facilitate the preparations of traits retrieval models from the next-generation operational CHIME.

Foot placement early prediction is important for designing compliant controllers for wearable robotic systems. There have been many researches on human walking gait analysis, but most of them focus on historic foot placement measurement and estimation, the work on foot placement early prediction has been rarely seen. This paper investigated three machine learning regression algorithms for foot placement prediction: Linear Regression, Support Vector Machine Regression and Gaussian Process Regression. The regression models were trained on the collected walking data set, and tested on the test data set, in which the subject and the walking speeds were different from those in the training data set. The results indicated that Gaussian Process Regression showed the best performance in foot placement prediction, and the prediction error decreased with the window size of the input data increasing. The experimental results demonstrated that, given the foot position information during the early 0.2 s in the swing phase, Gaussian Process Regression can predict the next foot placement. The Root Mean Squared Error was 0.0440 m and 0.0424 m along the walking direction and cross-walking direction, respectively, which was less than 5% of the average stride length. The results of this paper are expected to help researchers select a suitable regression model for gait prediction and inspire the following works.