T. Nikaein
Please Note
6 records found
1
From InSAR Time-Series to Crop Growth
Machine Learning and Physics-Guided Models for Radar-Based Vegetation Analysis
The first part of this research evaluates the added value of Sentinel-1 InSAR coherence time series for land cover classification, using an agricultural region in São Paulo, Brazil, as a case study. This region is characterized by a mixture of crops, pastures, and sugarcane plantations, all managed asynchronously. The findings demonstrate that incorporating InSAR coherence alongside SAR backscatter improves classification accuracy, particularly during the dry season when the distinctions between vegetation and bare soil are more pronounced. The research employed machine learning approaches to analyze pixel-level and field-level classifications using different sampling schemes. It highlights how multi-looking strategies can be adjusted to improve the accuracy of the classification outcomes in agricultural settings. This research highlights the usefulness of coherence data for the detection of events such as harvesting, offering valuable insights for more dynamic agricultural monitoring. The sensitivity of the coherence to agricultural changes leads to the observed improvement in Land Use Land Cover (LULC) mapping.
Forward models, or observation operators, are essential for the interpretation of radar observations and for the development of assimilation frameworks. In particular, in this research, we are interested in forward modeling the relation between crop bio-geophysical parameters, such as Leaf Area Index (LAI), Above Ground Biomass (AGB), and soil moisture, the inputs to our data-driven model, and radar observables, the outputs.
In the second part of this research, we integrate an existing crop growth model, the Decision Support System for Agrotechnology Transfer (DSSAT), with machine learning techniques to train a forward model to predict SAR observables over silage maize fields in The Netherlands across multiple years. Using crop growth models circumvents the dependency on limitedly available field measurements. When we use training and validation data from the same growth season, we obtain accurate predictions, with a mean absolute error (MAE) of less than 1.23 dB. Some of the field-to-field variability is accounted for by including the mean backscatter intensity during a few acquisitions before crop emergence. The obtained performance suggests the potential of using this approach to generate observation operators for data assimilation frameworks or for anomaly detection, supporting large-scale agricultural monitoring. However, the results also highlight one of the main challenges: the resulting data-driven model fails to generalize when presented with input bio-geophysical parameters that fall outside the regions of the parameter space spanned by the training dataset, as can happen, for example, during a drought period.
In the final part, we develop a physics-guided machine learning approach to address the limitations of data-driven models: lack of generalizability, tendency to overfitting, and reliance on extensive training data sets. We introduce physical constraints in an artificial neural network (ANN) in two ways. First, by modifying the loss function, used to train the ANN, by including a penalty for unphysical behavior, in particular by penalizing negative values of the partial derivative of the predicted backscatter intensity with respect to the surface soil moisture, since we assume this should always be positive. Second, by mirroring the architecture of the widely used Water Cloud Model (WCM) in the network topology. The added physical term to the loss function improves the ANN performance in all cases considered, with an R2 increase of 3 percentage points (p.p). The WCM-inspired model performs slightly worse when trained and tested with data from the same year, but it generalizes better, producing significantly better results for unseen conditions. In addition, the WCM-inspired model also produces individual contributions to the observed intensity, such as the surface-scattering component and the vegetation backscatter component. ...
The first part of this research evaluates the added value of Sentinel-1 InSAR coherence time series for land cover classification, using an agricultural region in São Paulo, Brazil, as a case study. This region is characterized by a mixture of crops, pastures, and sugarcane plantations, all managed asynchronously. The findings demonstrate that incorporating InSAR coherence alongside SAR backscatter improves classification accuracy, particularly during the dry season when the distinctions between vegetation and bare soil are more pronounced. The research employed machine learning approaches to analyze pixel-level and field-level classifications using different sampling schemes. It highlights how multi-looking strategies can be adjusted to improve the accuracy of the classification outcomes in agricultural settings. This research highlights the usefulness of coherence data for the detection of events such as harvesting, offering valuable insights for more dynamic agricultural monitoring. The sensitivity of the coherence to agricultural changes leads to the observed improvement in Land Use Land Cover (LULC) mapping.
Forward models, or observation operators, are essential for the interpretation of radar observations and for the development of assimilation frameworks. In particular, in this research, we are interested in forward modeling the relation between crop bio-geophysical parameters, such as Leaf Area Index (LAI), Above Ground Biomass (AGB), and soil moisture, the inputs to our data-driven model, and radar observables, the outputs.
In the second part of this research, we integrate an existing crop growth model, the Decision Support System for Agrotechnology Transfer (DSSAT), with machine learning techniques to train a forward model to predict SAR observables over silage maize fields in The Netherlands across multiple years. Using crop growth models circumvents the dependency on limitedly available field measurements. When we use training and validation data from the same growth season, we obtain accurate predictions, with a mean absolute error (MAE) of less than 1.23 dB. Some of the field-to-field variability is accounted for by including the mean backscatter intensity during a few acquisitions before crop emergence. The obtained performance suggests the potential of using this approach to generate observation operators for data assimilation frameworks or for anomaly detection, supporting large-scale agricultural monitoring. However, the results also highlight one of the main challenges: the resulting data-driven model fails to generalize when presented with input bio-geophysical parameters that fall outside the regions of the parameter space spanned by the training dataset, as can happen, for example, during a drought period.
In the final part, we develop a physics-guided machine learning approach to address the limitations of data-driven models: lack of generalizability, tendency to overfitting, and reliance on extensive training data sets. We introduce physical constraints in an artificial neural network (ANN) in two ways. First, by modifying the loss function, used to train the ANN, by including a penalty for unphysical behavior, in particular by penalizing negative values of the partial derivative of the predicted backscatter intensity with respect to the surface soil moisture, since we assume this should always be positive. Second, by mirroring the architecture of the widely used Water Cloud Model (WCM) in the network topology. The added physical term to the loss function improves the ANN performance in all cases considered, with an R2 increase of 3 percentage points (p.p). The WCM-inspired model performs slightly worse when trained and tested with data from the same year, but it generalizes better, producing significantly better results for unseen conditions. In addition, the WCM-inspired model also produces individual contributions to the observed intensity, such as the surface-scattering component and the vegetation backscatter component.
Physics-Guided Machine Learning Based Forward-Modeling of Radar Observables
A Case Study on Sentinel-1 Observations of Corn-Fields
Artificial neural networks have the potential to model the interaction of radar signals with vegetation but often do not follow the physical rules. This article aims to develop a new physics-guided machine learning approach that combines neural networks and physics-based models to leverage their complementary strengths and improve the modeling of physical processes. We propose a data-driven framework to model synthetic aperture radar observables by incorporating physical knowledge in two ways: through the network architecture and the loss function. A key aspect of our approach is its ability to integrate knowledge encoded in physics-based models. The results show that by using scientific knowledge to guide the construction and learning of the neural network, we can provide a framework with better generalizability and stability.
In this article, our aim is to estimate synthetic aperture radar (SAR) observables, such as backscatter in VV and VH polarizations, as well as the VH/VV ratio, cross ratio, and interferometric coherence in VV, from agricultural fields. In this study, we use the decision support system for agrotechnology transfer (DSSAT) crop-growth simulation model to simulate parcel-level phenological and growth parameters for over 1500 parcels of silage maize in the Netherlands. The crop model was calibrated using field data, including silage maize phenological phases, leaf area index, and above-ground dry biomass (AGB). The simulations incorporate fine-resolution gridded precipitation data and soil parameters to model the interaction between soil-plant-atmosphere and genotype in DSSAT. The crop variables produced by DSSAT are then used as inputs to a support vector regression model. This model is trained to simulate SAR observables in 2017, 2018, and 2019, and its performance is evaluated using independent fields in each of these years. The results show a close fit between modeled and observed SAR C-band observables. The importance of vegetation variables in the estimation of SAR observables is assessed. The AGB showed significant importance in the estimation of backscatter. This study demonstrates the potential value of combining crop-growth simulation models and machine learning to simulate SAR observables. For example, the SVR model developed here could be used as an observation operator in an assimilation context to constrain vegetation and soil water dynamics in a crop-growth model.
The study is aimed at understanding the value of interferometric coherence in mapping regions characterized by a mixture of crops and grasses. The results highlight that a 5% improvement in the classification accuracy can be achieved by using the coherence in addition to the backscatter intensity and by combining VV and VH. It is shown that the largest contribute in class discrimination is brought in winter, when dry vegetation and bare soils can be expected. It was also notably observed that coherence information can enhance the identification of harvesting events in a small but significant number of cases.
Synthetic aperture radar (SAR) acquisitions are mainly deemed suitable for mapping dynamic land-cover and land-use scenarios due to their timeliness and reliability. This particularly applies to Sentinel-1 imagery. Nevertheless, the accurate mapping of regions characterized by a mixture of crops and grasses can still represent a challenge. Radar time-series have to date mainly been exploited through backscatter intensities, whereas only fewer contributions have focused on analyzing the potential of interferometric information, intuitively enhanced by the short revisit. In this paper, we evaluate, as primary objective, the added value of short-temporal baseline coherences over a complex agricultural area in the São Paulo state, cultivated with heterogeneously (asynchronously) managed annual crops, grasses for pasture and sugarcane plantations. We also investigated the sensitivity of the radar information to the classification methods as well as to the data preparation and sampling practices. Two supervised machine learning methods—namely support vector machine (SVM) and random forest (RF)—were applied to the Sentinel-1 time-series at the pixel and field levels. The results highlight that an improvement of 10 percentage points (p.p.) in the classification accuracy can be achieved by using the coherence in addition to the backscatter intensity and by combining co-polarized (VV) and cross-polarized (VH) information. It is shown that the largest contribution in class discrimination is brought during winter, when dry vegetation and bare soils can be expected. One of the added values of coherence was indeed identified in the enhanced sensitivity to harvest events in a small but significant number of cases.