Table of Contents
Introduction
Especially in the field of agriculture, where time-critical information is significant and difficult to be obtained by optical imagery that is prone to cloud cover, SAR data shows its benefits. A lack of ready-to-use products and automated ways of deriving at exploitable SAR features together with the difficulty in interpreting the information may represent two main reasons for the community’s reluctance to use SAR data. Therefore – in the context of the current post – processing pipelines for the calculation of backscatter intensity, polarimetric decomposition parameters and coherence values based on S-1 SLC images are presented. Subsequently, analyses regarding the informative value of these SAR features are exemplified considering a seven-month timeseries of S-1 data to analyse a total of a few hundred fields that cover six different crop types in an area of interest located in Upper Austria. A valorisation of the timeseries information performing random forest-based classifications is presented finally.
Processing of SAR data towards usable features
- Backscatter coefficient – The backscatter coefficient is proportional to the power density carried by the wave, which is a function of radar system variables and the radar cross section σ. The latter reveals information on the properties of the illuminated surface. The set of relevant surface properties concerns the surface geometry, roughness and dielectricity. They interact with the parameters of the radar system such as wavelength, angle of incidence and polarisation in determining the backscatter coefficient.
- Polarimetric features – These features are based on the polarisation of the electromagnetic wave as expressed by the phase differences of the backscattered wave in the plane perpendicular to the direction of propagation. Scattering polarimetry and corresponding decomposition approaches explicitly study the changes of the polarisation states between emitted and backscattered waves induced by the properties of the analysed target. Considering the incident and scattered wave, polarimetric information can be described by a 2 x 2 complex scattering matrix relating the polarisation state descriptors of the involved electromagnetic waves. Decomposition approaches are subsequently used to derive at a vector representation of this matrix to facilitate its interpretation. The underlying idea is to decompose the matrix, which given real SAR data acquisition settings may embody information from distributed scatteres with different and complex responses, into a sum of responses corresponding to canonical mechanisms. One differentiates coherent and incoherent decomposition approaches. Focusing on incoherent decompositions, one popular approach is the eigenvector-eigenvalue based H/A/α-decomposition. To ease the interpretation of the physical information given by the eigenvalues and -vectors, secondary parameters are defined subsequently. These are the eponymous entropy (H), anisotropy (A) and mean alpha angle (α). H quantifies the degree of randomness of the scattering process, varying between 0 expressing the predominance of a single pure scatterer and 1 expressing the combination and mixture of different targets. α describes the scattering mechanism itself with values close to 0° in case of rough surface single-bounce scattering, values close to 45° in case of volume scattering and values close to 90° for double-bounce scattering. Finally, the anisotropy contains complementary information to H but is often considered secondary as it can only be interpreted reliably if the H exceeds 0.7.
- Interferometric coherence – Interferometry exploits the phase information of the backscattered signal. In order to use the cyclic phase information, two co-registered SAR images are required so that the phase information of both images can be related. A complex interferogram is achieved by cross-multiplying the first SAR image by the complex conjugate of the second image. Interferometric coherence can then be calculated as the normalised complex cross-correlation coefficient of the images. It is a measure that quantifies the degree of the pixel-to-pixel resemblance between the two images where higher values close to 1 express a greater degree of agreement. The total coherence is composed of multiple parts. Different decorrelation sources induce phase noise in the interferogram. The most important factor in repeat-pass settings is temporal decorrelation. Apart from changes in the atmosphere, this can largely be attributed to physical surface changes between the two points in time at which the images were taken.
All of the features described above can be computed relying on ESAs SNAP software. Using the programmatic interface snappy, the subsequently depicted processing chains are implemented in python scripts to automise the process of product creation. Note that due to some operators relying on spatial neighbourhoods (e.g. multilooking or also the matrix decomposition itself), the final resolution of the derived products varies. It is highest for the backscatter coefficient (approx. 20m x 20m), whereas the coherence products produced by the following pipeline have an effective resolution of 37.5m x 37.5m. The resolution for the polarimetric decomposition products is even coarser (50m x 50m).
Study Area & Field Data
To investigate the informative value of a timeseries of S-1 data processed using the presented pipelines, a study area (AoI) located in Upper Austria subject to intensive agricultural use was chosen. The vegetation period lasts from the end of March to the beginning of November. The timespan for S-1 data was set correspondingly. Field boundaries and crop type information for the year of interest (2021) were derived from the INVEKOS data set compiled by Agrarmarkt Austria, the competent body governed by public law. The six most frequent field usages were selected to obtain a set of crop types that covers a sufficiently large but still manageable selection of plants to be analysed subsequently. Also due to the discussed spatial resolution of the products, small fields with sizes less than 0.5 ha were excluded. Thereby, the the variability in the timeseries across fields of the same crop type is limited and the identification of dominating patterns eased. As part of the pre-processing field borders were excluded by introducing an inner buffer of 20m to avoid boundary effects.
Timeseries of SAR features
The following figures show the temporal evolution of the analysed SAR features averaged by fruit type. Similar time series are grouped together in terms of illustration, e.g. winter wheat and barley next to each other. The similarities and differences that are visible can be correlated fairly closely with the corresponding development cycles of the plants. The timeseries of barley and wheat, for example, are similiar as the two cereals have similar appearances during most phenological stages. They go through the vegetation cycle in an analogous manner – only with a slight time delay. A more profound analyses of the correlation between phenological developments of the plants and the evolution of SAR features requires to review the meaning of the derived SAR features, which will not be provided here. In case of further interest, please refer to my bachelor thesis and/or the full report at the end of this blog post.
Crop classification
To test the valorisation of the timeseries information for crop classifications, a set of models with different incoming explanatory variables was formed. This was done to assess the influence of the different SAR features obtained for different orbits and polarisations in terms of their influence on classification accuracies. Random forest was chosen as a classification algorithm for all models. As a de-facto standard in the remote sensing field this ensemble model stands out due to its robustness against overfitting in combination with good performance and high accuracies that are achieved. For the training and evaluation of the random forest results, two points have to be taken into account in the given setting: first, the small sample sizes, and, second, the class imbalance.
- The small sample sizes induce a considerable variability in the results of the classification in repeated trainings with different underlying random bootstraps. To counteract this, each model was trained 1000 times and the results averaged over all runs, so that narrow confidence intervals for the accuracies can be given despite the low sample size.
- The class imbalance, which leads to biases in training and results of the classifier, can basically be handled in two ways. Usually adjusted accuracy metrics are used post-hoc and metrics are reported per class to enable differentiated evaluation. Due to the very pronounced imbalance in the given setting, however, an upstream adjustment is used here. For the training run of a model, the existing set of fields is subsampled according to the number of fields in the weakest class (green fallow) and only the now equally distributed set of fields is presented to the classifier. This eases the interpretation of various accuracy metrics and facilitates an honest evaluation of the actual imbalance-free performance of the classifier differentiated by classes using a common confusion matrix. Applying a test-heavy 60/40 train/test sample split, this procedure leads to 75 fields, which are used as test samples in each run.
The best performing model – relying on backscatter & coherence information from both polarisations and orbits under consideration – achieves an overall test accuracy of 92.9 %.The subsequent confusion matrix allows a more detailed differentiation of the performance per crop type. The main source of error is the misclassification of green fallow and other crops as green fallow. Accordingly, the user (UA) and observer accuracy (OA) for green fallows are the lowest at 88.2 % and 83.1 %, respectively. This problem can basically be attributed to the high variability across different green fallow plots. The fact that green fallows are also the least populated class makes it even more difficult to adequately represent this variability in a 60/40 split with less than 20 training fields. Winter barley (UA: 99.4 %, OA: 96.6 %), winter wheat (UA: 98.2 %, OA: 94.9 %) and oat (UA: 95.2 %, OA: 94.2 %) are classified most reliably. As it may be particularly difficult to distinguish between these cereals, which have a similar phenological cycle only slightly delayed, by means of optical data, this emphasises the complementary added value of SAR data for field crop classification.
(figures averaged across all test runs)
At least as interesting as looking at the best model is comparing the different models tested. Here it should be notred first that all models perform reasonable on the given classification task considering that the accuracy for the worst model is still 84.9 %. The difference of 8 % to the best model at the same time shows that the specific parameter constellation has a significant influence on a model’s performance. Meeting the general expectations, models with a larger variety of features, polarisations and orbits tend to perform better. The top seven models are based on more than 100 explanatory variables, meaning that the consider multiple features and/or polarisations and/or orbits Specifically, at least two of these conditions hold true for all of the seven eight models. Contrary, the two models performing worst are based on less than 50 variables expressing single feature, polarisation and orbit information. Assessing the influence of the individual parameters systematically (see figure below), it seems that…
…the orbit configuration causes the largest variability among models. Especially noteworthy is the consistency in the better performance of models relying on orbit 44 information compared to those which use variables derived from orbit 95. Among the models that perform worse than the comparable 44 models when using the Orbit 95 are first those that are based on the backscatter alone, with -3.5 % (VH), -4.0 % (VV & VH) and -5.0 % (VV). For models that additionally include polarimetric and/or coherence information, the differences with values between -3.6 % and -0.9 % are still considerable although somewhat weaker. This suggests that the poorer performance of the orbit 95 models is mainly due to the stronger incidence angle of the corresponding measurements. The lower penetration of the signal manifests itself in a stronger component of the soil relative to the vegetation information in the backscatter intensity. This makes the backscatter intensity less suitable for classification, especially in the phases of weak vegetation cover. Apart from the difference between the single orbits, the benefit of using dual-orbit backscatter information is evident. Comparing the dual-orbit models against the models relying solely on orbit 44 leads to improvements up to 2.3 %, for orbit 95 the benefits are correspondingly larger and reach up to 5.8 %.
…the choice of a single polarisation has a less clear impact on model performance. However – similar to the dual-orbit configurations – looking at the performance of the models that include both polarisation information, better accuracies are observed across all other parameter constellations.
…that include coherence in addition to backscatter consistently achieve higher accuracies than those that rely solely on backscatter information. The improvements of about 1.6 % on average are quite considerable. The addition of polarimetric information, however, does not seem to improve the model performance any further. Five models perform slightly worse with differences of up to -0.7 %, while four models outperform their counterparts based on intensity and coherence. Thus, the fact that polarimetric variables should contain additional information not represented by the backscatter intensity and coherence does not manifest itself in the empirical results. Therea sons for this could be mixed pixel effects at the field borders due to the lower spatial resolution of the polarimetric grids as well as the general limitations in the meaningfulness of polarimetric variables in dual-pol constellations. Finally, the high redundancy with correlations greater than 0.98 between the individual polarimetric variables may affect the performance of the random forest classifier. Including only a single polarimetric variable, one may achieve better results.
Summary
Appendix - Scripts & Full report
The complete analyses – somewhat more detailed and extensive than those presented above – can be reproduced by means of the scripts made available below. The results are explained in the provided report.
