Conference Agenda

Because of the sheer size of grasslands, grassland degradation makes a far more important contribution than cropland degradation to the global environmental change phenomenon of desertification, which is land degradation in dry areas. About 42% of China's national land area consists of grasslands, and if all the grasslands in the world were brought together they would cover 50% more than the total area of Europe. Yet since there are no reliable estimates of the scale and degree of grassland degradation in the world, the global extent of desertification is equally uncertain, and this also makes it difficult to monitor compliance with the Land Degradation Neutrality (LDN) target of the Sustainable Development Goals (SDGs).

Devising a reliable method for mapping grassland degradation is therefore imperative, and is the aim of this project. In the first part of this project we have critically evaluated all previous research into mapping desertification with satellite data in order to devise such a method. We report eight main findings. First, grassland degradation, like desertification, is a very complex phenomenon, comprising multiple forms of soil degradation and vegetation degradation. Second, grassland degradation is poorly monitored at national scale in most countries. Third, even though monitoring dryland degradation is essential to estimate compliance with the LDN target, little progress in using satellite data and other types of "big data" to compensate for the lack of national data has been reported in the literature since the SDGs were finalized. Fourth, matching the spatial resolution of a sensor to the areal variability of each form of vegetation degradation and soil degradation is necessary to minimize spatial systematic errors, while matching the temporal resolution of a sensor to the turnover time of each component is necessary to minimize temporal systematic errors. Fifth, consequently, a multiple sensor approach is needed to monitor grassland degradation, since while medium (20-80 m) resolution data can be used for initial land use mapping, key grassland biophysical parameters(e.g. biomass, productivity, soil carbon content) estimation and the mapping of some forms of land degradation (such as sandy area expansion), very high (≤ 1 m) resolution data is often needed for mapping most forms of vegetation degradation (e.g. tree density, shrub encroachment) and soil degradation (e.g. gully expansion) in dry areas. Sixth, a multilayer multisensor approach, with land use classification and key grassland parameters estimation preceding degradation measurement, will optimize image processing time and the overall accuracy of grassland degradation assessment. Seventh, reliable methods are still lacking for using satellite data to measure wind erosion and soil compaction. Eighth, there is great scope to experiment with radar and LIDAR sensors to improve the accuracy and coverage of grassland degradation monitoring.

The continuous monitoring of natural multi-hazards, namely large landslides, is of fundamental importance to minimize and prevent actual and future risks. In this Dragon-5 project, Earth Observation (OE) data are used and integrated into various methods for the identification and monitoring of landslide phenomena at different spatio-temporal scales. The upper-Jinsha River region, composed of steep terrain and broken structures, is known to cause frequent landslide disasters. Distributed scatterer SAR interferometry (DS-InSAR) was used to monitor and analyze the Woda landslide area. In the study, 106 Sentinel-1A ascending and 102 Sentinel-1A descending images, covering a time period of almost 5 years (from 2014/11/05 to 2019/09/04 and from 2014/10/31 to 2019/09/11, respectively). The two-dimensional deformation of the landslide revealed that the maximum surface deformation rate in the normal direction is about -80 mm/yr, and in the east-west direction is about 118 mm/yr. According to the rescaled range (R/S) analysis, the Hurst index the deformation trend will continue for some time.

The Gilgit region, Pakistan, was also analyzed using a time series of Sentinel-1 images, covering a time period of about a year (2019 – 2020). Several deformation areas were mapped in the region. A 3D prototype platform, enabling the visualization of the deformation results of the typical active landslides, has been developed, allowing the identification of 3 landslides with potential risk.

The use of Artificial Intelligence (AI) techniques in EO data, to develop/optimize data processing and analysis methods, is also one of the major goals of this project. In this regard, the multitemporal difference interferometric phase diagram of subsidence mining area is obtained by using synthetic aperture radar differential interferometry (DInSAR) techno machine learning (ML) FCN-8s, PSPNet Deeplabv3 and U-Net models were applied to DinSAR data to extract mining subsidence. The results show that the U-Net model presents high detection accuracy and takes short time to run. In order to improve the semantic segmentation and extraction accuracy of mining subsidence, the efficient channel attention (ECA) module is introduced into the traditional U-Net model for training. The detection of earthquake fringes with Vision Transformer (ViT) technic was also tested in the scope of this project. In our methodology we use a dataset composed of 5110 interferograms from LicSAR dataset, that were used to apply DL techniques to recognize patterns. The F1-Score and AUC clearly show that the ViT outperforms VGG19, respectively, with 0.88 against 0.69, and 0.97 against 0.86. However, the major disadvantage of this method consists in the detection of false negatives in overlapped patches. Sometimes the model identifies as “no deformation” patches with deformation fringes but that change in the overlapped patches for the same interferogram. For a better confirmation the method is evaluated applied to the whole interferogram. It can be concluded that the results are higher in both models, proving that some false negatives are adjusted in overlapped patches. ViT remains the model with the best performance. The precision is now equal for both (1.0) and with a significant rise of the Recall and F1-Scores. As main conclusions, we can say that the new and different model proposed in this work to detect fringes in SAR interferograms achieves better results, comparing to Convolutional Neural Networks. This way, we are even closer of obtaining results that can create practical products that may help providing a faster response in earthquakes scenarios.

Monitoring structures of great heritage and historical value, more quickly and effectively, is also one of the major goals of our project. However, only the use of Artificial Intelligence techniques will allow to deal with the huge amount of data that will be generated. The Vilariça Valley, located in the north of Portugal is crossed by an active fault and 3 classified buidings (churches of Torre de Moncorvo, Freixo de Espada à Cinta and Foz Côa, in Portugal) are being used as test sites. More than 1000 Sentinel-1, ERS and Envisat images are already processed providing thousands of points in the villages where these buildings are located. However, only a few points were identified directly in the buildings and in their proximity. To increase the density of points in the building and in its proximity, a time series of 25 PAZ images was commissioned, with the support of ESA. In the near future, these results will be integrated with data from field sensors and LiDAR models, using AI techniques.

Based on the big archived Sentinel-1 satellite radar data during the last 7 years, we developed parallel processing techniques for the detection of subtle ground motion on a phase-stable point basis, so that InSAR deformation could be used for accurately determining tectonic motion or strains induced by anthropogenic activities. In particular, we strengthen our analysis results by joint analysis with 3D GPS data within the same time span if available. In this way, we are able to separate ground motion with long-wavelength errors of InSAR results, which is the major source of errors when tectonic motion is considered. Joint analysis of GPS and InSAR data also provides us a way to separate InSAR surface motion in the Line-Of-Sight direction into vertical and horizontal deformation in particular orientations. But it is worthless to have 3D measurements with InSAR only as it cannot effectively discern the signals in a nearly north-south direction.

We apply our method in various tectonic regions for seismic deformation detections, including the population densified North China Plain, the fast shale-gas development Sichuan basin, and the tectonic active Eastern Tibetan Plateau, etc. In this study, we reduce various InSAR errors, in particular the unwrapping errors as before, while the most important long-wavelength errors are overcome by GPS-assistant joint analysis in this version of processing, so that tectonic motion can be well detected and compared with GPS measurements.

The project aims to develop and apply innovative data analytic methods underpinned with machine (deep) learning technology to analyze and detect seismic anomalies from electromagnetic data observed by the SWARM and CSES satellites along with CSELF network. In this report we will present the progress of anomaly detection algorithms underpinned with Deep Neural Networks (DNN), which focuses on predicting Swarm data as long as possible from the Swarm historic data. We report our investigation into several architectures of Recurrent Neural Networks (RNN), in particular we investigate the development of a Long-Short Term Memory (LSTM) based methods. We present our design and implementation of the LSTM algorithm and predicted results of applying the algorithm to learn from the Swarm historic data. We will also describe error metrics for measuring the accuracy of Swarm data reconstruction and the methods of detecting anomalies in the Swarm data, which are in relation to three earthquakes.