This paper presents a comparative analysis of three Long Short-Term Memory (LSTM) architectures for precipitation prediction and regression in urban areas. We implemented and evaluated a simple stacked bidirectional LSTM, a two-part model with separate classification and regression networks, and an integrated hurdle LSTM model. The architectures were tested on two distinct datasets: a large, highly variable dataset from Romania's National Air Quality Monitoring Network (NAQMN) and the more uniform CAMS ERA5 dataset from the European Centre for Medium-Range Forecasts (ECMWF). Results indicate that model performance is highly dependent on data characteristics. For the CAMS ERA5 dataset, all models achieved good performance, with the simple LSTM explaining data variability most effectively (R2-score of 0.63). On the more challenging and zero-biased NAQMN dataset, the hurdle LSTM model was clearly superior, providing more accurate predictions with a significantly lower Mean Absolute Error (MAE) of 0.23 compared to the other models. This study demonstrates that for complex and varied real-world data, sophisticated architectures like the hurdle LSTM are required for robust prediction, while simpler models can be effective on more homogenous datasets.

Despite the permanent efforts to reduce emissions and achieve carbon neutrality a warmer climate is no longer to be avoided. The European mission „Adaptation to Climate Change" aims to build resilience by 2030 in at least 150 European communities and regions. At the same time, the „Destination Earth" (DestinE) initiative promotes the use of digital twins of the Earth enabling a thorough assessment of climate change by leveraging an accurate digital model of the Earth that can be used to monitor, model, and predict natural and human activity, and to develop and test scenarios for a more sustainable growth. Climate models describe changes at scales of 50km to 150km. However, adaptation measures shall be applied at human activities scales, from 10m to 1km. We propose to achieve this by scale-out novel paradigms of Artificial Intelligence for Earth Observation (AI4EO) including the use of coupled models across domains and spatiotemporal scales. The envisaged R&D work will be carried out in the project Competence Center for Climate Change Digital Twin Earth for forecasts and societal redressement: DTEClimate, in the frame of Romania National Recovery and Resilience Plan.

Satellite image time series are a worthwhile source of information for a broad range of applications, especially in the context of future global challenges. The challenge to discover correlations, patterns or anomalies would be eased if the data analysts might benefit of visualization tools enabling them to grasp, briefly, the characteristics of the time series. Responding to these needs, this paper proposes a graphical user interface focused on SITS visualization. While overcoming the current software limitations we developed a Python GUI interface enabling remote sensing scientists who use Python for their research, to visually analyze the satellite image time series in the same environment.

In the context of accelerated urbanization, metropolitan green infrastructure is considered a strategic approach to ensure healthy and sustainable living environments. Earth Observation (EO) offers the right means for large scale and long term assessment and monitoring of such green areas and the entire urban environment. The methodology presented in this paper leverages one of the most common satellite missions for vegetation assessment, the Sentinel-2 mission, applies super-resolutions techniques to increase the image spatial resolution and quantifies the spectral radiation reflected by the ground in order to map the Earth’s biophysical properties. By considering multiple acquisitions over the same area, time series of spectral indices are generated and processed using LDA, a generative model well known for hierarchical latent information extraction in both text and image analysis. The resulting temporal signature of each topic is further correlated with the evidence of environmental indicators to underline the vegetation vulnerability and specificity of the species. A use case centered for the Bucharest city in Romania, was included.

Sentinel-2 image super-resolution (SR) has proven advantageous in multiple data analysis pipelines, leading to a more comprehensive assessment of different environment-related metrics. This research aims to provide a method for super-resolving the 60-m bands provided by Sentinel-2 up to 10-m spatial resolution, using Gaussian process regression (GPR). While common GPR methods directly operate on raw data using carefully designed kernels, we propose a convolutional neural network (CNN)-based feature extraction kernel to directly process the input 10-m patches, applied in constructing the elements of the integrated covariance matrices. For each scene, a small number of training patches are sampled to optimize the CNN parameters and to construct the predictive mean function, the latter being further used for predicting super-resolved pixels for new input areas. We prove that our method is a reliable SR mechanism by assessing its performance both quantitatively, using metrics against other methods from literature, and qualitatively, through visual analysis of the results.

Deep learning methods have become ubiquitous tools in many Earth observation applications, delivering state-of-the-art results while proving to generalize for a variety of scenarios. One such domain concerns the Sentinel-2 (S2) satellite mission, which provides multispectral images in the form of 13 spectral bands, captured at three different spatial resolutions: 10, 20, and 60 m. This research aims to provide a super-resolution mechanism based on fully convolutional neural networks (CNNs) for upsampling the low-resolution (LR) spectral bands of S2 up to 10-m spatial resolution. Our approach is centered on attaining good performance with respect to two main properties: consistency and synthesis. While the synthesis evaluation, also known as Wald’s protocol, has spoken for the performance of almost all previously introduced methods, the consistency property has been overlooked as a viable evaluation procedure. Recently introduced techniques make use of sensor’s modulation transfer function (MTF) to learn an approximate inverse mapping from LR to high-resolution images, which is on a direct path for achieving a good consistency value. To this end, we propose a multiobjective loss for training our architectures, including an MTF-based mechanism, a direct input–output mapping using synthetically degraded data, along with direct similarity measures between high-frequency details from already available 10-m bands, and super-resolved images. Experiments indicate that our method is able to achieve a good tradeoff between consistency and synthesis properties, along with competitive visual quality results.

The Normalized Difference Vegetation Index (NDVI) is an important factor to be considered in vegetation tracking and analysis, which can be easily derived from multispectral (MS) images. However, the limitation imposed by the atmospheric conditions makes the calculation of this index difficult. Because of the clouds, only a limited number of multispectral bands can capture the land appropriately. Furthermore, the multispectral sensors are dependent on the sunlight, which makes the acquisition of data more limited. These limitations do not hinder other types of Earth Observation (EO) data, like the scenes captured by the Synthetic Aperture Radar (SAR). However, SAR images cannot be used in NDVI calculation. In this article, we propose a deep learning (DL) based method for NDVI estimation from SAR data. Using a database with corresponding MS and SAR patches, we calculate the NDVI for each sample, then use a convolutional neural network (CNN) for predicting the NDVI of SAR images. This simple method leads to a precision of 70% in NDVI estimation from SAR images.

Although the number of labeled datasets in Earth Observation (EO) is increasing, there is still a major gap between the Deep Learning (DL) classifiers designed in this field versus the models in Computer Vision. This gap is produced mainly by the number of datasets available, but also by the diversity of data. In EO, there are different sensors acquiring images, from multispectral (MS) or hyperspectral data, to SAR imagery. In this paper, we want to demonstrate how to reduce the divergence created by the diversity of data. We trained several DL architectures on Bag-of-Words from large-scale MS and SAR datasets, and then we used transfer learning on smaller ones and evaluated the results. With this method, we demonstrate that a DL architecture can be trained with any type of large-scale data, transformed into Bag-of-Words, and the trained model can be used further on other types of data, without regard on the number of channels.

In this paper, we present three methods that reduce the computational time of training Deep Neural Networks with multispectral images, optimize the resource occupation of the dataset, and obtain high performance for reduced datasets. In the first two methods, we reduce the dimension of the input data with either histograms of pixel intensity or Bag-of-Words. Then we train a Convolutional Neural Network with either histograms or Bag-of-Words and we achieve an accelerated training. Moreover, storing the image patches from the dataset in the form of histograms or Bag-of-Words reduced the memory storage significantly. In the last method, we subsample the training dataset randomly to 50%, 20% and 10% of the original dataset, thus training a Convolutional Neural Network on a smaller number of samples (in the form of histograms or Bag-of-Words), and the classification performance is almost unaffected. This is an important achievement, as there are few labelled datasets for Earth Observation and the number of images in these datasets is small. Our results show that the training time is reduced by a maximum of 387 times and the datasets with histograms or Bag-of-Words occupy 633 times less space.

The two Sentinel-2 satellites provide, since March 2017, high-resolution worldwide images every five days, freely distributed, generating terabytes of high-dimensional data. An intuitive manner to summarize the main characteristics of the data and gather knowledge is visual exploratory analysis, which is often based on dimensionality reduction methods to represent high-dimensional data. From previous research and the state-of-the-art literature, turned out that t-distributed Stochastic Neighbour Embedding is one of the most appropriate technique to reduce the dimensionality of a dataset, but it requires very high computational power. To overcome this inconvenience, we proposed two hybrid DR algorithms, which combine the DR with the nearest neighbour technique or random forest regression. The main conclusion is that our approaches reduce computational power without compromising the representation quality.

Ecosystems are providing essential services to humans. Food production, drinkable water, clean air, biodiversity support, climate regulation and nutrient cycling, mitigation of natural disasters, regulation of pests and diseases are just few examples of ecosystem services of vital importance. This paper proposes an integrated approach for terrestrial and ecosystems' studies targeting to integrate in situ measurements acquired by automate ground sensors with Earth observation data analysis to provide data quality services for multiple applications. The resulted services will be delivered through a web platform, acting both as a highly instrumented site and as a prototype for a national distributed monitoring network to support transparent and knowledge based conservation and management policies.

This paper introduces a tool designed to provide an innovative and insightful way of exploring Earth observation data content beyond visualization, by addressing a visual analytics process. The considered framework combines machine learning and visualization techniques, empowered through human interaction, to gain knowledge from the data. The proposed tool- eVADE leverages the methodologies developed in the fields of information retrieval, data mining and knowledge representation by the means of a visual analytics component. eVADE increases users capability to understand and extract meaningful semantic clusters together with quantitative measurements, presented in a suggestive visual way.

Exploratory visual analysis is often required to assist human operator to understand and interpret Earth Observation (EO) images. Optimal image representation offers cognitive support in discovering relevant facts about the scene with respect to a particular application. This is of crucial importance for training data sets selection in all Machine Learning tasks, particularly in the design of active learning tools for multispectral (MS) EO data. This paper proposes a deep neural network (DNN) based method to compress, learn and reveal the most significant information included in the spectral bands of EO data in support of relevant visualization for image content analysis. The advanced method uses a DNN to discover the most suggestive pseudo-color representation able to highlight the entire MS image content better than the particular 3 bands selection (R, G, B). We propose the use of information theory and the concept of mutual information to rank the spectral bands based on the amount of information contained, by applying the minimum-redundancy-maximum-relevance (mRMR) criterion on a the image so that we obtain the ranked bands. A DNN stacked autoencoder based paradigm is developed in order to extract and compress in three bands the overall information from the MS EO data. The developed method is demonstrated and validated for Sentinel 2 dataset.

The data mining systems solve the problem of handling Earth Observation archives counting on a feature vectors based description of the data. Increasing the dimensionality of the feature vectors would offer an effective perspective of the dataset's content. The modern systems provide visual exploration of data projecting their high-dimensional feature space in a 3-D space. The dimensionality reduction methods represent the main way to achieve such representation. Several dimensionality reduction methods have been proposed to identify the mapping, bot not all of them retain the same dataset properties. In order to compare their performance, the development of formal measures like “Trustworthiness” or the measures based on Co-ranking matrix was mandatory. These measures objectively evaluate the similarity between the structure detected in the original and the reduced space. In this paper we evaluate six dimensionality reduction methods using “Trustworthiness” and “Continuity” measures. In this regard three datasets have been used: an artificial one and two remote sensing datasets. Each of them have been described by a high-dimensional feature space.

Modern knowledge discovery systems, empowered by visual data exploration techniques, enable the user to discover and understand the data content. Considering patch-level processing, the visual exploration of Earth Observation archives aims to identify groups of items sharing similar semantic content. Each patch is further represented by certain descriptors, i.e., spectral signatures or Weber local descriptors, to capture structural signature. Later on, the content of the archive is illustrated by a 3-D projection of the high-dimensional space of the descriptors. Aspiring to prove the visual data mining potential, this letter intends to determine the capability of dimensionality reduction techniques to achieve a meaningful 3-D projection of the high-dimensional space. Several real-world data sets were used, i.e., University of California, Merced Land Use data set and a Landsat 7 Enhanced Thematic Mapper Plus image tiled into patches.

Visual data mining methods are of great importance in exploratory data analysis having a high potential for mining large databases. As the data feature space is generally n-dimensional, visual data mining relies on dimensionality reduction techniques. This is the case for image feature spaces which can be visualized by giving each data point a location in a three dimensional space. This paper aims to present a comparative study of several dimensionality reduction methods considering as input image feature spaces, in order to detemine an optimal visualization method to illustrate the separation of the classes. At the beginning, to check the performance of the envisaged method, an artificial dataset consisting of random vectors describing six, 20-dimensional Gaussian distributions with spaced means and low variances was generated. Further, two real images datasets are used to evaluate the contributions of dimensionality reduction algorithms related to data visualization. The analysis focuses on the PCA, LDA and t-SNE dimensionality reduction techniques. Our tests are performed on images for which the computed features include the color histogram and Weber descriptors.

In this paper we present the result of data analytics techniques applied to a database comprising of 32 SLC SM TerraSAR-X images, acquired over the area of Bucharest, Romania. The methodology follows a two step approach. The first stage consists of a coarse identification of potentially changed areas using a supervised learning image annotation tool with relevance feedback. Gabor texture features are used to describe image patches. The patch size is derived as a function of the resolution and pixel spacing of the data. In the second stage we apply an information theory strategy to refine the regions previously shown to exhibit class dynamics within the image stack, with pixel accuracy. Finally, a series of analytical indicators (absolute extent of areas affected by change, class evolution trends, inter-class correlations) are derived, in order to generate a predictive model for the selected test site.

The ill-defined nature of the segmentation problem makes the selection of the optimal image partition difficult. One can characterize image segmentation as an attempt to find the best possible representation of a data set using a certain number of ldquoobjects.rdquo This can be regarded as data information compression, resulting in the distortion of the original values. Data sets are well represented when the correct number of regions is chosen. The concept behind this approach is similar to the main problem of rate distortion theory: A finite set of code words is chosen to approximate the numbers or source symbols as well as possible. In our approach, the number of regions is equivalent to the number of code words. The mean of a region provides canonical representation of respective group members, and the distortion function is the mean-square error assuring a good evaluation method for image segmentation.