Conference Agenda

Multi-temporal Interferometric Synthetic Aperture Radar (MT-InSAR) is widely used in earth observation, but there are challenges when using it for monitoring urban linear infrastructure, such as the estimation of accurate deformation in the presence of noisy observations and the identification of deformation temporal patterns, particularly for large-scale and long-term time series analysis (Wu et al., 2020; Ma et al., 2022). To improve interpretation efficiency and accuracy, classification of the temporal evolution of deformation can be helpful. However, the majority of the existing classification methods are based on models of the temporal evolution, hence relying on prior knowledge and professional expertise, leading to relatively low application efficiency. To address this issue, we develop a data-driven post-processing method that provides a new solution for fine monitoring of linear infrastructure using MT-InSAR analyses.

Our method is based on a variational autoencoder network composed of Long Short-Term Memory layers and a clustering layer to identify different deformation evolution patterns. We also add spatial information of observation points to obtain more reliable clustering results, as measured by clustering assessment indices such as the Davies-Bouldin score and Silhouette score. We evaluate the effectiveness of our method on simulated datasets, which include several typical deformation temporal patterns, as well as on real datasets from different regions with varying degrees of deformation: Tianjin (China) railway lines, the Shanghai (China) maglev, and Wakefield (UK) railway lines. The corresponding datasets are TerraSAR-X images from 2015 to 2019 covering Tianjin, TerraSAR-X images from 2013 to 2020 covering Shanghai, and Sentinel-1 images from 2016 to 2022 covering Wakefield. These datasets have varying temporal epochs, wavelengths, and resolutions. The MT-InSAR process is carried out using StaMPS (Hooper et al., 2007), and unwrapping errors are separately identified and corrected. In addition, our processing includes linking two overlapping frames of time series for Shanghai.

Our deep learning-based clustering method is independent of predefined deformation models, which sets it apart from previous classification methods (Cigna et al., 2011; Berti et al., 2013; Chang and Hanssen, 2016; Mirmazloumi et al., 2022). We find that our method outperforms the baseline methods, including K-Means, in identifying deformation evolution patterns along linear infrastructures. We will interpret our clustering results with external measures such as geological background, structural knowledge, and surrounding groundwater level change. In this way, we will gain deeper insight into the mechanism of deformation. Our preliminary results indicate that deep clustering on MT-InSAR analyses can improve monitoring efficiency and automation.

Reference:

Berti, M., Corsini, A., Franceschini, S., Iannacone, J., 2013. Automated classification of persistent scatterers interferometry time series. Natural Hazards and Earth System Sciences, 13, 1945-1958.

Chang, L., Hanssen, R., 2016. A probabilistic approach for InSAR time-series postprocessing. IEEE Transactions on Geoscience and Remote Sensing, 54, 421-430.

Cigna, F., Del Ventisette, C., Liguori, V., Casagli, N., 2011. Advanced radar-interpretation of InSAR time series for mapping and characterization of geological processes. Natural Hazards and Earth System Sciences, 11, 865-881.

Hooper, A., Segall, P., Zebker, H., 2007. Persistent scatterer interferometric synthetic aperture radar for crustal deformation analysis, with application to Volcán Alcedo, Galápagos. Journal of Geophysical Research, 112, B07407.

Ma, P., Lin, H., Wang, W., Yu, H., Chen, F., Jiang, L., Zhou, L., Zhang, Z., Shi, G., Wang, J., 2022. Toward Fine Surveillance: A review of multitemporal interferometric synthetic aperture radar for infrastructure health monitoring. IEEE Geoscience and Remote Sensing Magazine, 10, 207-230.

Mirmazloumi, S., Wassie, Y., Navarro, J., Palamà, R., Krishnakumar, V., Barra, A., Cuevas-González, M., Crosetto, M., Monserrat, O., 2022. Classification of ground deformation using sentinel-1 persistent scatterer interferometry time series. GIScience & Remote Sensing, 59, 374-392.

Wu, S., Le, Y., Zhang, L., Ding, X., 2020. Multi-temporal InSAR for urban deformation monitoring: progress and challenges. Journal of Radars, 9, 277-294.

Monitoring ground displacements identifies potential geohazard risks early before they cause critical damage. Interferometric synthetic aperture radar (InSAR) is one of the techniques that can monitor these displacements with an accuracy of sub-millimeters. However, using the InSAR technique is quite challenging because of the requirements of a high experience level, a massive amount of data, and many other complications. In this aspect, numerous machine learning algorithms have been integrated with the domain of InSAR to develop indications and predictions about the information of the displacements that the InSAR technique provides. Mainly, we refer to predicting the ground displacements in volcanic regions using Convolutional Neural Networks (CNN), examining an association between infrastructure displacements utilizing Machine Learning Algorithms (MLAs) and International Roughness Index (IRI) values depending on Persistent Scatterer Interferometric Synthetic Aperture Radar (PS-InSAR), and training labeled wrapped and unwrapped interferograms to detect the fringes using CNN. Assuredly we can mention other recognized articles regarding the integration between InSAR and Artificial Intelligence (AI), but this is far from the subject of our research work.

Here, we implemented numerous machine learning algorithms to inspect the possibility of predicting an indication about ground displacements directly from Sentinel-1's wrapped interferograms. In other words, the objective is to use these intelligent algorithms in developing a methodology that can automatically analyze large InSAR data packets to identify areas where the ground is at risk of displacement.

We used the Parallel Small BAseline Subset (P-SBAS) service to achieve the mentioned objective, depending on the advanced cloud computing platform, the GeoHazards Exploitation Platform (G-TEP). The P-SBAS service creates the required interferograms, coherence, and time series maps for the case studies. The work on this research was sponsored by ESA’s valuable Network of Resources (NoR) initiative that gave us access to the G-TEP and allowed a complete connection to the Sentinels-1 repositories.

However, the Atmospheric Phase Screen (APS) difference is one of the most problematic effects that limit the accuracy of derived displacements by InSAR. Subsequently, the atmospheric artifacts cannot be ignored. On the contrary, they should be removed from the wrapped interferograms to reduce the global and local phase errors. Accordingly, we applied a high pass filter on the interferograms to improve the accuracy of the predicted signals, considering that the main power of the low-frequency signal comes from the atmospheric artifacts (especially for the interferograms of very short temporal baselines). The applied high-pass filter cuts off all low-frequency components at a specified distance D0 from the origin of the transform.

The inputs of the implemented machine learning models are groups of pixels clipped from the filtered wrapped interferograms using a coherence threshold of 0.7. Then, the outputs are the same groups of pixels labeled as either Stability or Movement. The labels have been created depending on the velocity values of the Measurement Points (MPs) located in the pixels.

It is worth noting that the coherence threshold has been extracted for each pixel from the average coherence map. The value of 0.7 is determined by CNR-IREA as the minimum coherence value for selecting the pixels to produce the time series of the MPs.

The first step before training the model was to set a 0 cm/year threshold to differentiate between uplifting and subsiding. If the velocity value of the MP is less than 0 cm/year, the label will be Subsiding, and if the velocity value of the MP is higher than 0 cm/year, the label is Uplifting.

Then we selected a -0.7 cm/year threshold of stability/movement. In other words, when the velocity value of the MP is less than -0.7 cm/year, the label will be Movement. When the velocity value of the MP is between 0 and -0.7 cm/year, the label is Stability. Hence, the algorithm can predict if the pixel is subsiding or uplifting and whether the subsidence is low or high.

Above and beyond all other considerations, we had to observe the difference between the signals of subsidence/ uplifting and stability/movement before feeding the datasets to the machine learning models. Successfully, we could detect the patterns of Heaving/Subsiding by figuring out the distinct distributions between the histograms that represent each dataset. Moreover, we analyzed the signals of Subsidence Stability/Movement and detected a unique chart for each of them.

The experiments included three case studies from Milan (Italy), Lisbon (Portugal), and Washington (United States) regarding their high sensitivity to landslides. Two datasets have been developed for each case study. The first dataset involved single pixels and had one measurement point (MP) inside each pixel. The second dataset involved patches of 3*3 pixels and had one MP in the center of each pixel's group.

All the trained models for 1×1 pixel datasets achieved high validation accuracy (5 folds). On the contrary, not all of them performed high test accuracy, considering that the test sets included pixels from a different adjacent area to the train sets areas. In other words, the test sets were not hidden parts of the training set to the same geographic extent. Instead, they were samples from an adjoining area.

Experiencing several machine learning methods, we found that the Subspace KNN (Ensemble) and Cosine KNN (KNN) models achieved the best test accuracy.

The validation accuracy for Uplifting/Subsiding was 94.9%, 97.3%, and 93.4% for Milan, Washington, and Lisbon, respectively, while the test accuracy scored 79.1%, 85.2%, and 79.5% for Milan, Washington, and Lisbon, respectively.

The validation accuracy for Subsiding Stability/Movement was 98.3%, 89.7%, and 99.4% for Milan, Washington, and Lisbon, respectively, while the test accuracy reached 87.2%, 82%, and 74.8% for Milan, Washington, and Lisbon, respectively.

Both models are controlled by the KNN search technique, which is a simple yet effective non-parametric supervised learning classifier. It uses the approximation and the distance functions to predict the grouping of an individual data point.

The KNN subspace model arbitrarily chooses a set of predictors from the possible values (without replacement) to train a weak learner. It repeats this step to develop a group of weak learners. Then, it classifies the category depending on the highest average score of the learners.

The cosine similarity KNN model is the measure of similarity between two data points in multidimensional space. It detects whether the two samples are pointing in the same direction and operates entirely on the cosine principles, i.e., where there is a decrease in the angle, the similarity of data points increases.

The reason behind creating other datasets of 3×3 pixels was to explore if the algorithm can recognize the pattern of displacements in the test sets more effectively. But by training these datasets, we could get neither higher validation accuracy nor testing accuracy.

Furthermore, we used QGIS to compare the ground truth with the predictions to visualize the results better.

Finally, it is worth noting that we trained the algorithm depending on three classes (Slight, Moderate, and High) too. The models achieved high validation accuracy, but we could not get high test accuracy. The reason is that the training samples became fewer after equally dividing the dataset into three categories. The upcoming work is to check the implementation of the PSI technique through the same workflow and to check the capacity of the algorithm to define the stability/movement classes of an uplifting phenomenon.

Keywords: Sentinel-1, Ground Displacements, P-SBAS, Machine Learning

Land subsidence is a potentially catastrophic geohazard affecting many areas in the world, often caused by overexploitation of aquifers and requiring regular monitoring. In recent years, the dramatic increase in open source Interferometric Synthetic Aperture (InSAR) data enabled scientists to monitor land subsidence over large areas (100’s km) and at high spatial resolution (~30 m pixel). However, loss of coherence over time in areas prone to rapid physical changes of the Earth’s surface is still a limitation, causing non-spatially homogeneous ground-motion samples, even when using SAR data from missions with frequent revisit times i.e. Sentinel-1. In this study, we try to overcome this limitation by introducing a machine learning method for increasing the density of reliable ground-motion pixels in InSAR velocity maps. Our study area is the Carpi town (Modena province, northern Italy) where subsidence has been reported. We formed a series of ascending and descending interferograms from Sentinel-1 acquisitions spanning the 2017-2021 period over the town of Carpi. We calculated the average velocity map and time-series of incremental LOS (Line-of-Sight) displacements using two known methods P-SBAS (Parallel Small BAseline Subset) and Persistent Scatterer Interferometry (PSI) for both ascending and descending data. The InSAR velocity maps show a clear pattern consistent with subsidence in Carpi of up to 20 mm/y along the Line Of Sight (LOS) but with non-spatially homogeneous ground-motion sampling in both P-SBAS and PSI solutions. We then analyze the InSAR time-series using a machine learning method based on Deep Transformers in order to extract an average velocity map with complete spatially distributed ground-motion pixels. Deep Transformers is a machine learning architecture with dominant performance, low calibration cost and agnostic method. We generated the machine learning dataset consisting of the collection of vectorial records derived from the InSAR time-series. The training and the testing sets were made by the 80% and 20% of randomly extracted vectorial records, respectively. Our InSAR velocity map derived from Deep Transformers has nearly completed the whole study area, showing complementary ground-motion pixels, and it is also consistent with the InSAR velocity maps obtained with conventional SBAS and PSI techniques in those areas that remain coherent. The final results of this work clearly showed the potential of Transformer approach to solve InSAR loss of coherence limitation.

Synthetic aperture radar (SAR) represents nowadays a well-recognized technique for a broad variety of remote sensing applications, being able to acquire high-resolution images of the Earth’s surface, independently of daylight and weather conditions. Next-generation SAR systems will bring significant improvements in performance through the exploitation of digital beam forming (DBF) techniques in combination with multiple acquisition channels, bi- and multi-static sensor configurations, together with the use of large bandwidths. This will allow, among others, to overcome the limitations imposed by conventional SAR imaging for the acquisition of wide swaths and, at the same time, of finer resolutions. These paradigms are currently being widely applied in innovative studies, technology developments and mission concepts by several space institutions and industries. The significant improvements that can be achieved in terms of acquisition capabilities are associated with the generation of huge volumes of data, which, in turn, set harder requirements for the onboard memory and downlink capacity of the system. Indeed, present global SAR mapping missions, such as Sentinel-1 or TanDEM-X, or future missions, such as NISAR and especially ROSE-L and Sentinel-1 Next Generation, will acquire data over selected areas with a temporal sampling down to one week, resulting in large data volumes which need to be handled by the SAR sensor.

In this scenario, the efficient digitization of the SAR raw data represents an aspect of crucial importance as, on the one hand, it directly defines the amount of on-board data volume but also, on the other hand, it affects the quality of the generated SAR products. These two aspects must be traded off due to the limited acquisition and downlink capacity and onboard resources of the system. One of the most widely used compression schemes for SAR raw data digitization is the Block-Adaptive Quantization (BAQ) [1]. In the last years, building on the principle of BAQ, novel algorithms have been proposed, allowing for a finer performance and resource optimization. These methods are based on acquisition-dependent compression schemes, as in the case of the FDBAQ [2], or combined with the implementation of non-integer data rates [3]. In the context of the Performance-Optimized BAQ (PO-BAQ), the basic concept of the original BAQ is further extended in [4], which represents a first attempt for an optimization of the resource allocation depending on the performance requirement defined for SAR and InSAR product. Such an optimization can be achieved by exploiting the a-priori knowledge of the SAR backscatter statistics of the acquired scene, an information which must therefore be made available to the sensor, in form of look-up-tables (LUTs) or backscatter maps, hence leading to an increased required computational effort and resources. Overall, the above-mentioned SAR compression approaches are not fully adaptive with respect to the acquired raw data, since the quantization settings are derived from prior considerations and do not account for the actual radar backscatter statistics at the time of the SAR survey. Indeed, the quantization performance depends on the local characteristics of the illuminated scene on ground, which are, in turn, linked to the local topography, targets characteristics and illumination geometry, resulting in different backscatter absolute levels and degrees of heterogeneity.

In order to overcome these limitations, Artificial Intelligence (AI) represents a promising approach in the remote sensing community, since it enables scalable exploration of big data and bringing new insights on information retrieval solutions [5]. In this work, we investigate the use of AI, and in particular of deep learning (DL), for on-board SAR raw data compression, with the aim of deriving an optimized and adaptive data rate allocation, depending on a pre-defined performance requirement. The latter can be defined on typical SAR and InSAR quality metrics which are affected by raw data quantization, such as the Signal-to-Quantization Noise Ratio (SQNR), the interferometric coherence loss or the SAR/InSAR phase error.

We propose to approach the described problem as a deep supervised semantic segmentation task, where the number of bits to be allocated for quantization is derived on board and is identified as the target output class. To generate the reference bitrate maps required during the training of the proposed DL architecture, we exploit experimental TanDEM-X data acquired in BAQ-bypass conditions (i.e. the data are quantized with an 8-bit ADC), which are then re-compressed on ground using different BAQ rates (i.e., 2, 3, 4, 5 and 6 bits/sample). In this way, we assess the impact of SAR quantization on the desired performance parameter, and starting from this a reference bit rate map (BRM) is derived by considering the minimum local bitrate which satisfies the given requirement. A large variety of different landcover classes as well as acquisition geometries need to be considered for the generation of the complete training data set, in order to be representative and so to properly generalize the proposed DL model.

After the generation of the two-dimensional bit rate map during the DL architecture inference stage, a state-of-the-art quantizer, such as the BAQ, is locally applied to the input raw data samples, leading to the generation of a quantized raw data matrix with local variable bitrate.

A clear advantage of this approach relies in the fact that the quantization bitrate maps are dynamically generated from the raw data matrix that is given as input feature map to the DL architecture. This means, e.g., that when considering the same area of acquisition, different bitrate maps can be generated depending on the local characteristics of the backscatter at the time of the SAR survey.

Aim of this research work will be to provide a complete analysis including the detailed description of the uncompressed SAR raw data, the adopted DL architecture(s) used for efficient data compression, as well as the overall performance assessment with respect a state-of-the-art quantizer.

References

[1] R. Kwok and W.T.K. Johnson, “Block adaptive quantization of Magellan SAR data,” IEEE Transactions on Geoscience and Remote Sensing, vol. 27, no. 4, pp. 375–383, 1989.

[2] Paul Snoeij, Evert Attema, Andrea Monti Guarnieri, and Fabio Rocca, “FDBAQ a novel encoding scheme for Sentinel-1,” in 2009 IEEE International Geoscience and Remote Sensing Symposium (IGARSS), 2009, vol. 1, pp. I–44–I–47.

[3] Michele Martone, Benjamin Bräutigam, Paola Rizzoli, and Gerhard Krieger, “Azimuth-Switched Quantization for SAR Systems and Performance Analysis on TanDEM-X Data,” IEEE Geoscience and Remote Sensing Letters, vol. 11, no. 1, pp. 181-185, 2013.

[4] Michele Martone, Nicola Gollin, Paola Rizzoli, and Gerhard Krieger, “Performance-optimized quantization for SAR and InSAR applications,” IEEE Transactions on Geoscience and Remote Sensing, vol. 60, 2022.

[5] Xiao Xiang Zhu, Sina Montazeri, Mohsin Ali, Yuansheng Hua, Yuanyuan Wang, Lichao Mou, Yilei Shi, Feng Xu, and Richard Bamler, “Deep learning meets SAR: Concepts, models, pitfalls, and perspectives,” IEEE Geoscience and Remote Sensing Magazine, vol. 9, no. 4, pp. 143–172, 2021.

Rainforests are of utmost importance in the dynamics of the global ecological balance, being known as Earth's thermostat because of its role in stabilizing climate. They are crucial in mitigating global warming by producing around 20% of our planet's oxygen, storing billions of tons of carbon dioxide every year and regulating its water cycle not only locally but also dictating precipitation patterns at a global scale. Tropical forests, in particular, are extremely complex ecosystems, hosting more than half of the known flora and fauna. However, the equilibrium of these rich yet fragile ecosystems is menaced by forest degradation and deforestation caused by human activity. The Amazon biome -- home to both the largest worldwide tropical rainforest and river basin -- is threatened for instance by agricultural advancements, illegal mining and logging, as well as urban expansion. Thus, the availability of reliable and up-to-date data describing land cover features, and more specifically those related to forest parameters, is critical in the joint effort to mitigate habitat loss and the disruption of these environments.
A potential data source up to this task arises from spaceborne Synthetic Aperture Radar (SAR) systems, whose imaging capabilities surpass those from optical sensors in the sense that they are operational around-the-clock even over cloudy atmospheric conditions, which is typically the case over tropical rainforests. However, the main limitation of employing SAR imagery for land cover monitoring is their difficult interpretability through visual inspection, further aggravated by noise contributions such as speckle. In order to address such challenges, we first propose to investigate temporal information from repeat-pass interferometric SAR (InSAR) data in addition to the most traditional SAR backscatter feature by considering short time series with a temporal baseline smaller than a month, i.e., we use these time series of data to generate a forest map every 30 days or less. More specifically, we concentrate on the Amazon rainforest basin and, in particular, on its arc of deforestation by assuming the stationariety of the illuminated scene during a period of only 24 days. At the present moment, Sentinel-1 InSAR short time series have been successfully used as input to a supervised deep learning model in a proof of concept over the Brazilian state of Rondonia between April and May 2019, showing the potential of convolutional networks for the systematic mapping of forests when compared with traditional machine learning techniques for the same reference maps.
In this work, we will further develop the current findings by investigating ways of generalizing the current model to an operational framework for year-round forest mapping at a larger scale. Preliminary results show that a model trained during the Amazon's dry season only might be characterized by a lower performance when predicting land cover classes in the wet season (as an example, it is expected that the mean precipitation between December and February is up to 15 times higher than between June and August in the state of Rondonia). Moreover, a feature analysis of backscatter and interferometric coherence stacks at temporal baselines of 12 and 24 days for both co- and cross-polarization channels throughout the year suggests a certain degree of correlation between the amount of precipitation in a region and the difficulty in discriminating the forest from other land cover classes. To this end, we expect that the use of ancillary data such as precipitation and land surface temperature statistics might bridge the performance gap between the dry and wet seasons, contributing to the achievement of a more robust classification scheme. Another potential agent in these discrepancies between seasons might lie in the fact that the ground-truth data over the Amazon rainforest is typically generated over the dry season only (when most forest disturbances happen through logging and fires and selected cloud-free optical images are used for producing the references through visual inspection). To this end, we also investigate the use of unsupervised techniques by means of convolutional auto-encoders (CAEs) to cluster forest and non-forest data at any given time of the year.
In the final paper we will present the pros and cons of each approach and their viability when being implemented at a larger scale. The final goal is the development of a pre-operational framework, running on high performance computing facilities, for the regular monitoring of the Amazon basin, the detection of changes and the set up of a reliable early warning system for deforestation activities and forest degradation phenomena independently of the cloud cover.