NP4.1

The mapping of faults and fractures is a problem of high relevance in Earth Sciences. However, their identification in digital elevation models is a time-consuming task given the resulting networks' fractal nature. The effort is especially challenging in submarine environments, given their inaccessibility and difficulty in collecting direct observations. Here, we propose a semi-automated method for detecting faults in high-resolution gridded bathymetry data (~1 m horizontal and ~0.2 m vertical) of the Pescadero Basin in the southern Gulf of California, which were collected by MBARI's D. Allan B autonomous underwater vehicle. This problem is well suited to be explored by machine learning and deep-learning methods. The method learns from a model trained to recognize fault-line scarps based on key morphological attributes in the neighboring Alarcón Rise. We use the product of the mass diffusion coefficient with time, scarp height, and root-mean-square error as training attributes. The method consists of projecting the attributes from a three-dimensional space to a one-dimensional space in which normal probability density functions are generated to classify faults. The LDA implementation results in various cross-sectional profiles along the Pescadero Basin show that the proposed method can detect fault-line scarps of different sizes and degradation stages. Moreover, the method is robust to moderate amounts of noise (i.e., random topography and data collection artifacts) and correctly handles different fault dip angles. Experiments show that both isolated and linkage fault configurations are detected and tracked reliably.

How to cite: Vega Ramirez, L. A., Splez Madero, R. M., Contreras Perez, J., Caress, D., Clague, D. A., and Paduan, J. B.: A new Method for Fault-Scarp Detection Using Linear Discriminant Analysis (LDA) in High-Resolution Bathymetry Data From the Alarcón Rise and Pescadero Basin, Gulf of California., EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-417, https://doi.org/10.5194/egusphere-egu21-417, 2021.

Random forests is a supervised machine learning algorithm which has witnessed recently an exponential increase in its implementation in water resources. However, the existing implementations have been restricted in applications of Breiman’s (2001) original algorithm to regression and classification models, while numerous developments could be also useful for solving diverse practical problems. Here we popularize random forests for the practicing hydrologist and present alternative random forests based algorithms and related concepts and techniques, which are underappreciated in hydrology. We review random forests applications in water resources and provide guidelines for the full exploitation of the potential of the algorithm and its variants. Relevant implementations of random forests related software in the R programming language are also presented.

How to cite: Tyralis, H., Papacharalampous, G., and Langousis, A.: Random forests in water resources, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-2105, https://doi.org/10.5194/egusphere-egu21-2105, 2021.

Nowadays a wide range of methods and tools to study and forecast time series is available. An important problem in forecasting concerns embedding of time series, i.e. construction of a high dimensional space where forecasting problem is considered as a regression task. There are several basic linear and nonlinear approaches of constructing such space by defining an optimal delay vector using different theoretical concepts. Another way is to consider this space as an input feature space – IFS, and to apply machine learning feature selection (FS) algorithms to optimize IFS according to the problem under study (analysis, modelling or forecasting). Such approach is an empirical one: it is based on data and depends on the FS algorithms applied. In machine learning features are generally classified as relevant, redundant and irrelevant. It gives a reach possibility to perform advanced multivariate time series exploration and development of interpretable predictive models.

Therefore, in the present research different FS algorithms are used to analyze fundamental properties of time series from empirical point of view. Linear and nonlinear simulated time series are studied in detail to understand the advantages and drawbacks of the proposed approach. Real data case studies deal with air pollution and wind speed times series. Preliminary results are quite promising and more research is in progress.

How to cite: Kanevski, M.: Empirical analysis of time series using feature selection algorithms, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-6697, https://doi.org/10.5194/egusphere-egu21-6697, 2021.

Multi-dimensional arrays (also known as raster data, gridded data, or datacubes) are key, if not essential, in many science and engineering domains. In the case of Earth sciences, a significant amount of the data that is produced falls into the category of array data. That being said, the amount of data that is produced daily from this field is huge. This makes it hard for researchers to analyze and retrieve any valuable insight from it. 1-D sensor data, 2-D satellite imagery, 3-D x/y/t image time series and x/y/z subsurface voxel data, 4-D x/y/z/t atmospheric and ocean data often produce dozens of Terabytes of data every day, and the rate is only expected to increase in the future. In response, Array Databases systems were specifically designed and constructed to provide modeling, storage, and processing support for multi-dimensional arrays. They offer a declarative query language for flexible data retrieval and some, e.g., rasdaman, provide federation processing and standard-based query capabilities compliant with OGC standards such as WCS, WCPS, and WMS. However, despite these advances, the gap between efficient information retrieval and the actual application of this data remains very broad, especially in the domain of artificial intelligence AI and machine learning ML.

In this contribution, we present the state-of-art in performing ML through Array Databases. First, a motivating example is introduced from the Deep Rain Project which aims at enhancing rainfall prediction accuracy in mountainous areas by implementing ML code on top of an Array Database. Deep Rain also explores novel methods for training prediction models by implementing server-side ML processing inside the database. A brief introduction of the Array Database rasdaman that is used in this project is also provided featuring its standard-based query capabilities and scalable federation processing features that are required for rainfall data processing. Next, the workflow approach for ML and Array Databases that is employed in the Deep Rain project is described in detail listing the benefits of using an Array Database with declarative query language capabilities in the machine learning pipeline. A concrete use case will be used to illustrate step by step how these tools integrate. Next, an alternative approach will be presented where ML is done inside the Array Database using user-defined functions UDFs. Finally,  a detailed comparison between the UDF and workflow approach is presented explaining their challenges and benefits.

How to cite: Campos Escobar, O. J. and Baumann, P.: Towards AI in Array Databases, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-7409, https://doi.org/10.5194/egusphere-egu21-7409, 2021.

The quest to understand cause and effect relationships is at the basis of the scientific enterprise. In cases where the classical approach of controlled experimentation is not feasible, methods from the modern framework of causal discovery provide an alternative way to learn about cause and effect from observational, i.e., non-experimental data. Recent years have seen an increasing interest in these methods from various scientific fields, for example in the climate and Earth system sciences (where large scale experimentation is often infeasible) as well as machine learning and artificial intelligence (where models based on an understanding of cause and effect promise to be more robust under changing conditions.)

In this contribution we present the novel LPCMCI algorithm for learning the cause and effect relationships in multivariate time series. The algorithm is specifically adapted to several challenges that are prevalent in time series considered in the climate and Earth system sciences, for example strong autocorrelations, combinations of time lagged and contemporaneous causal relationships, as well as nonlinearities. It moreover allows for the existence of latent confounders, i.e., it allows for unobserved common causes. While this complication is faced in most realistic scenarios, especially when investigating a system as complex as Earth's climate system, it is nevertheless assumed away in many existing algorithms. We demonstrate applications of LPCMCI to examples from a climate context and compare its performance to competing methods.

Related reference:
Gerhardus, Andreas and Runge, Jakob (2020). High-recall causal discovery for autocorrelated time series with latent confounders. In Advances in Neural Information Processing Systems 33 pre-proceedings (NeurIPS 2020). 

How to cite: Gerhardus, A. and Runge, J.: LPCMCI: Causal Discovery in Time Series with Latent Confounders, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-8259, https://doi.org/10.5194/egusphere-egu21-8259, 2021.

The Earth’s climate is a highly complex and dynamical system. To better understand and robustly predict it, knowledge about its underlying dynamics and causal dependency structure is required. Since controlled experiments are infeasible in the climate system, observational data-driven approaches are needed. Observational causal inference is a very active research topic and a plethora of methods have been proposed. Each of these approaches comes with inherent strengths, weaknesses, and assumptions about the data generating process as well as further constraints.
In this work, we focus on the fundamental case of bivariate causal discovery, i.e., given two data samples X and Y the task is to detect whether X causes Y or Y causes X. We present a large-scale benchmark that represents combinations of various characteristics of data-generating processes and sample sizes. By comparing most of the current state-of-the-art methods, we aim to shed light onto the real-world performance of evaluated methods. Since we employ synthetic data, we are able to precisely control the data characteristics and can unveil the behavior of methods when their underlying assumptions are met or violated. Further, we give a comparison on a set of real-world data with known causal relations to complete our evaluation.

How to cite: Käding, C. and Runge, J.: A Benchmark for Bivariate Causal Discovery Methods, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-8584, https://doi.org/10.5194/egusphere-egu21-8584, 2021.

Surface winds — both wind speed and vector wind components — are fields of fundamental climatic importance. The character of surface winds greatly influences (and is influenced by) surface exchanges of momentum, energy, and matter. These wind fields are of interest in their own right, particularly concerning the characterization of wind power density and wind extremes. Surface winds are influenced by small-scale features such as local topography and thermal contrasts. That is why accurate high-resolution prediction of near‐surface wind fields is a topic of central interest in various fields of science and industry. Statistical downscaling is the way for inferring information on physical quantities at a local scale from available low‐resolution data. It is one of the ways to avoid costly high‐resolution simulations. Statistical downscaling connects variability of various scales using statistical prediction models. This approach is fundamentally data-driven and can only be applied in locations where observations have been taken for a sufficiently long time to establish the statistical relationship. Our study considered statistical downscaling of surface winds (both wind speed and vector wind components) in the North Atlantic. Deep learning methods are among the most outstanding examples of state‐of‐the‐art machine learning techniques that allow approximating sophisticated nonlinear functions. In our study, we applied various approaches involving artificial neural networks for statistical downscaling of near‐surface wind vector fields. We used ERA-Interim reanalysis as low-resolution data and RAS-NAAD dynamical downscaling product (14km grid resolution) as a high-resolution target. We compared statistical downscaling results to those obtained with bilinear/bicubic interpolation with respect to downscaling quality. We investigated how network complexity affects downscaling performance. We will demonstrate the preliminary results of the comparison and propose the outlook for further development of our methods.

This work was undertaken with financial support by the Russian Science Foundation grant № 17-77-20112-P.

How to cite: Rezvov, V., Krinitskiy, M., Gavrikov, A., and Gulev, S.: Comparison of AI-based approaches for statistical downscaling of surface wind fields in the North Atlantic, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-8844, https://doi.org/10.5194/egusphere-egu21-8844, 2021.

Recently, Unmanned Aerial Vehicles (UAVs) equipped with high-resolution imaging sensors have become a viable alternative for ecologists to conduct wildlife censuses, compared to foot surveys. They cause less disturbance by sensing remotely, they provide coverage of otherwise inaccessible areas, and their images can be reviewed and double-checked in controlled screening sessions. However, the amount of data they generate often makes this photo-interpretation stage prohibitively time-consuming.

In this work, we automate the detection process with deep learning [4]. We focus on counting coastal seabirds on sand islands off the West African coast, where species like the African Royal Tern are at the top of the food chain [5]. Monitoring their abundance provides invaluable insights into biodiversity in this area [7]. In a first step, we obtained orthomosaics from nadir-looking UAVs over six sand islands with 1cm resolution. We then fully labelled one of them with points for four seabird species, which required three weeks for five annotators to do and resulted in over 21,000 individuals. Next, we further labelled the other five orthomosaics, but in an incomplete manner; we aimed for a low number of only 200 points per species. These points, together with a few background polygons, served as training data for our ResNet-based [2] detection model. This low number of points required multiple strategies to obtain stable predictions, including curriculum learning [1] and post-processing by a Markov random field [6]. In the end, our model was able to accurately predict the 21,000 birds of the test image with 90% precision at 90% recall (Fig. 1) [3]. Furthermore, this model required a mere 4.5 hours from creating training data to the final prediction, which is a fraction of the three weeks needed for the manual labelling process. Inference time is only a few minutes, which makes the model scale favourably to many more islands. In sum, the combination of UAVs and machine learning-based detectors simultaneously provides census possibilities with unprecedentedly high accuracy and comparably minuscule execution time.

Fig. 1: Our model is able to predict over 21,000 birds in high-resolution UAV images in a fraction of time compared to weeks of manual labelling.

References

1. Bengio, Yoshua, et al. "Curriculum learning." Proceedings of the 26th annual international conference on machine learning. 2009.

2. He, Kaiming, et al. "Deep residual learning for image recognition." Proceedings of the IEEE conference on computer vision and pattern recognition. 2016.

3. Kellenberger, Benjamin, et al. “21,000 Birds in 4.5 Hours: Efficient Large-scale Seabird Detection with Machine Learning.” Remote Sensing in Ecology and Conservation. Under review.

4. LeCun, Yann, Yoshua Bengio, and Geoffrey Hinton. "Deep learning." nature 521.7553 (2015): 436-444.

5. Parsons, Matt, et al. "Seabirds as indicators of the marine environment." ICES Journal of Marine Science 65.8 (2008): 1520-1526.

6. Tuia, Devis, Michele Volpi, and Gabriele Moser. "Decision fusion with multiple spatial supports by conditional random fields." IEEE Transactions on Geoscience and Remote Sensing 56.6 (2018): 3277-3289.

7. Veen, Jan, Hanneke Dallmeijer, and Thor Veen. "Selecting piscivorous bird species for monitoring environmental change in the Banc d'Arguin, Mauritania." Ardea 106.1 (2018): 5-18.

How to cite: Kellenberger, B., Veen, T., Folmer, E., and Tuia, D.: Deep Learning Enhances the Detection of Breeding Birds in UAV Images, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-12157, https://doi.org/10.5194/egusphere-egu21-12157, 2021.

How to cite: Sudmanns, M., Augustin, H., van der Meer, L., Baraldi, A., and Tiede, D.: The Sen2Cube.at national semantic Earth observation data cube for Austria, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-12486, https://doi.org/10.5194/egusphere-egu21-12486, 2021.

Total cloud score is a characteristic of weather conditions. At the moment, there are algorithms that automatically calculate cloudiness based on a photograph of the sky These algorithms do not know how to find the solar disk, so their work is not absolutely accurate.

To create an algorithm that solves this data, the data used, obtained as a result of sea research voyages, is used, which is marked up for training the neural network.

As a result of the work, an algorithm was obtained based on neural networks, based on a photograph of the sky, in order to determine the size and position of the solar disk, other algorithms can be used to work with images of the visible hemisphere of the sky.

How to cite: Borisov, M. and Krinitskiy, M.: Artificial neural networks in the problem of determining the position and size of the solar disk in wide-angle photographs of the sky, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-13051, https://doi.org/10.5194/egusphere-egu21-13051, 2021.

Within the framework of meteorology and oceanology, the importance of the cloud mass and the type of clouds cannot be underestimated. When describing and studying weather, precipitation and the movement of air masses over the ocean, the amount and type of clouds determines the flows of precipitation, their intensity, helps to predict the weather and the content of various impurities in the air, which makes the study of the properties of cloud cover one of the key aspects of meteorological and oceanological research.

The types of clouds are determined by the specialist, visually comparing the picture of the sky over the ocean with the guideline documents, the use of which reduces the possibility of the human factor affecting the determination of these parameters.

For an accurate study, study of the dynamics and dependence of climatic models on the conditions of cloud types, long-term measurements of the same type and the continuity of their methods are required. However, all these data are very unevenly distributed over the Earth's surface, and the number of ship observations is greatly reduced.

Thus, taking into account the importance of reliable determination of data related to cloudiness and the problems of their accuracy, the relevance and need to automate the determination of cloud types are obvious.

As a result of the work, an algorithm was obtained that allows classifying cloud types based on photographs taken during long-term sea expeditions.

How to cite: Veremev, N.: The use of artificial neural networks in the problem of classifying cloud types in wide-angle images of the visible hemisphere of the sky., EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-14470, https://doi.org/10.5194/egusphere-egu21-14470, 2021.

The use of artificial intelligence and specifically deep learning (DL) approaches in the domain of remote sensing is increasing. Such methods provide excellent results and show great potential for future applications. Earth observation sensors are able to deliver data with higher spatial, spectral and temporal resolutions. In this project, we use Sentinel-2 multispectral data and couple this input with a crowd annotated very high resolution (VHR) map which is generated in the video-game Cerberus, developed by the company BlackShore. In Cerberus, players are able to map features, like buildings, forest and specific types of crop fields, that are subsequently used as input for the Machine Learning (ML) pipeline. The ML pipeline is applied to classify crop fields in a larger region.

The main objective of this research is to study the accuracy of a model in detecting and describing the type of crop and whether the addition of a temporal dimension increases the accuracy. We will be experimenting with different methods that take their root in DL. The study region shown to Cerberus-players is Oromia in Ethiopia, south of the capital Addis Ababa. Using Sentinel-2 data, we aim to extend the generated maps to cover Ethiopia.

First, we will implement two DL methods; Random Forest (RF), and a 3D Convolutional Neural Network (CNN) that do not make use of the temporal dimension in order to have a baseline of the expected accuracy from a single multi-spectral image. Next, we will investigate four models that make use of time series: 1) a hybrid convolutional neural network-random forest (CNN-RF); 2) a 3D CNN that takes as input the output of a stack of 3D CNNs; 3) a model based on Recurrent Neural Networks (RNNs) performing pixel-based classification; and 4) an innovative method that combines the strength of RNNs, CNNs and Generative Adversarial Networks. 

We are now implementing the methods and shall report on results at EGU April 2021. For future research, it could be a very interesting case to study the possibility of generalizing the combined approach of crowd annotated training data with extended classification over larger regions and generalizing to other areas.

How to cite: Guillot, A., You, S., van't Woud, H., Perenboom, M., Kruijver, A., and Foing, B.: Crowd, Crops and Machines: How crowdsourced annotations can help towards crops classification, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-14553, https://doi.org/10.5194/egusphere-egu21-14553, 2021.

Remote sensing (RS) imagery, generated by e.g. cameras on satellites, airplanes and drones, has been used for a variety of applications such as environmental monitoring, detection of craters, monitoring temporal changes on planetary surfaces.

In recent years, researchers started applying Computer Vision [TP1] methods on RS data. This led to a steady development of remote sensing classification, providing good results on classification and segmentation tasks on RS data.  However, there are still problems with current approaches. Firstly, the main focus is on high-resolution RS imagery. Apart from the fact that these data are not accessible to everyone, the models fail to generalize on lower resolution data. Secondly, the models fail to generalize on more fine-grained classes. For example, models tend to generalize very well on detecting buildings in general, however they fail to distinguish if a building belongs to a fine-grained subclass like residential or commercial buildings. Fine-grained classes often appear very similar to each other, therefore, models have problems to distinguish between them. This problem occurs both in high-resolution and low-resolution RS imagery, however the drop in accuracy is much more significant when using lower resolution data.

For these reasons, we propose a Multi-Task Convolutional Neural Network (CNN) with three objective functions for segmentation of RS imagery. This model should be able to generalize on different resolutions and receive better accuracy than state-of the-art approaches, especially on fine-grained classes.

The model consists of two main components. The first component is a CNN that transforms the input image to a segmentation map. This module is optimized with a pixel-wise Cross-Entropy loss function between the segmentation map of the model and the ground truth annotations. If the input image is of lower resolution, this segmentation map will miss out on the complete structure of input images. The second component is another CNN to build a high-resolution image from the low-resolution input image in order to reconstruct fine-grained structure information. This module essentially guides the model to learn more fine-grained feature representations. The transformed image from this module will have much more details like sharper edges and better color. The second CNN module is optimized with a Mean-Squared-Error loss function between the original high-resolution image and the transformed image. Finally, the two images created by the model are then evaluated by a third objective function that aims to learn the distance of similarity between the segmented input image and the super-high resolution segmentation. The final objective function consists of a sum of the three objectives mentioned above. After the model is finished with training, the second module should be detached, meaning high-resolution imagery is only needed during the training phase.

At the moment we are implementing the model. Afterwards, we will benchmark the model against current state of the art approaches. The status will be presented at EGU 2021.­

How to cite: Zamarialai, S., Perenboom, T., Kruijver, A., Shi, Z., and Foing, B.: Monitoring Temporal Developments from Remote Sensing Data using AI Fine-Grained Segmentation, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-15917, https://doi.org/10.5194/egusphere-egu21-15917, 2021.

In order to make predictions on how the climate would respond to changes in global and regional emissions, we typically run simulations on Global Climate Models (GCMs) with perturbed emissions or concentration fields. These simulations are highly expensive and often require the availability of high-performance computers. Machine Learning (ML) can provide an alternative approach to estimating climate response to various emissions quickly and cheaply. 

We will present a Gaussian process emulator capable of predicting the global map of temperature response to different types of emissions (both greenhouse gases and aerosol pollutants), trained on a carefully designed set of simulations from a GCM. This particular work involves making short-term predictions on 5 year timescales but can be linked to an emulator from previous work that predicts on decadal timescales. We can also examine uncertainties associated with predictions to find out where where the method could benefit from increased training data. This is a particularly useful asset when constructing emulators for complex models, such as GCMs, where obtaining training runs is costly. 

How to cite: Mansfield, L., Nowack, P., and Voulgarakis, A.: Predicting climate model response to changing emissions, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-15944, https://doi.org/10.5194/egusphere-egu21-15944, 2021.

When working with ungridded remote sensing data, such as swath surface reflectance like Moderate Resolution Imaging Spectroradiometer (MODIS) MOD09 or Visible Infrared Imaging Radiometer Suite (VIIRS) VNP09, extracting targeted information of interest from a collection of granules can be a challenging exercise. Given a region of interest (ROI), it is tedious both to determine the subset of granules that intersect the ROI, as well as identifying, within the granules, the individual instantaneous field of views (IFOVs) contained by the ROI.

The SpatioTemporal Adaptive-Resolution Encoding (STARE) is an indexing scheme that recursively divides the Earth's surface into quadtree hierarchies, allowing triangular elements ("trixels") of varying sizes (resolutions) to be identified with unique index values. STARE is also a software library that operates on STARE indices. It can efficiently determine the spatial relationship between two trixels, by evaluating their index values, if the trixels share a common path in the STARE tree structure. By representing geographical regions as the sets of trixels with adaptive resolutions that tesselating them, STARE provides an elegant method to determine geospatial coincidence of arbitrarily shaped geographic regions, with accuracy up to ~7-8 cm in length. 

In this presentation, we introduce STARELite, a SQLite STARE extension and its use for cataloguing volumes of remote sensing granules that researchers often possess in their local storage. In this application, STARELite is used to determine subsets of granules intersecting arbitrary ROIs. Further, STARELite can be used for the inverse search problem: Determining all spatially coincident granules of an individual granule. STARELite leverages other components of the STARE ecosystem; namely STARE sidecars, which hold the trixel index values of each iFOV and a set of trixels representing the cover of each granule; STAREMaster, which is used to generate STARE sidecar files; and STARPandas, a Python Pandas extension used to bootstrap STARELite databases.

Given the limitations of SQLite, STARELite is to be understood as a proof of concept for the integration of STARE into relational databases in general. 

How to cite: Griessbaum, N., Rilee, M., Frew, J., and Kuo, K.-S.: Integrating STARE with relational databases, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-16410, https://doi.org/10.5194/egusphere-egu21-16410, 2021.