Amodal depth estimation aims to predict the depth of occluded (invisible) parts of objects in a scene. This task addresses the question of whether models can effectively perceive the geometry of occluded regions based on visible cues. Prior methods primarily rely on synthetic datasets and focus on metric depth estimation, limiting their generalization to real-world settings due to domain shifts and scalability challenges. In this paper, we propose a novel formulation of amodal depth estimation in the wild, focusing on relative depth prediction to improve model generalization across diverse natural images. We introduce a new large-scale dataset, Amodal Depth In the Wild (ADIW), created using a scalable pipeline that leverages segmentation datasets and compositing techniques. Depth maps are generated using large pre-trained depth models, and a scale-and-shift alignment strategy is employed to refine and blend depth predictions, ensuring consistency in ground-truth annotations. To tackle the amodal depth task, we present two complementary frameworks: Amodal-DAV2, a deterministic model based on Depth Anything V2, and Amodal-DepthFM, a generative model that integrates conditional flow matching principles. Our proposed frameworks effectively leverage the capabilities of large pre-trained models with minimal modifications to achieve high-quality amodal depth predictions. Experiments validate our design choices, demonstrating the flexibility of our models in generating diverse, plausible depth structures for occluded regions. Our method achieves a 69.5% improvement in accuracy over the previous SoTA on the ADIW dataset.

In recent years, free access to high and medium resolution data has become available, providing researchers with the opportunity to work with low resolution satellite images on a global scale. Sentinel-1 and Sentinel-2 are popular sources of information due to their high spectral and spatial resolution. To obtain a final product with a resolution of 10 meters, we have to use bands with a resolution of 10 meters. Other satellite data with lower resolution, such as Landsat-8 and Landsat-9, can improve the results of land monitoring, but their harmonization requires a process known as super-resolution. In this study, we propose a method for improving the resolution of low-resolution images using advanced deep learning techniques called Generative Adversarial Networks (GANs). The state-of-the-art neural networks, namely transformers, with the combination of channel attention and self-attention blocks were employed at the base of the GANs. Our experiments showed that this approach can effectively increase the resolution of Landsat satellite images and could be used for creating high resolution products.

In the past few years, medium and high-resolution data became freely available for downloading. It provides great opportunity for researchers not to select between solving the task with high-resolution data on small territory or on global scale, but with low-resolution satellite images. Due to high spectral and spatial resolution of the data, Sentinel-1 and Sentinel-2 are very popular sources of information. Nevertheless, in practice if we would like to receive final product in 10 m resolution we should use bands with 10 m resolution. Sentinel-2 has four such bands, but also has other bands, especially red-edge 20 m resolution bands that are useful for vegetation analysis and often are omitted due to lower resolution. Thus, in this study we propose methodology for enhancing resolution (super-resolution) of the existing low-resolution images to higher resolution images. The main idea is to use advanced methods of deep learning - Generative Adversarial Networks (GAN) and train it to increase the resolution for the satellite images. Experimental results for the Sentinel-2 data showed that this approach is efficient and could be used for creating high resolution products.

One of the parts of the Land Monitoring Service is Copernicus Urban Atlas, which provides reliable and comparable land use maps with high accuracy for large number European functional urban areas and their neighbors for every 6 years (2006, 2012, 2018). Unfortunately, there is no available such products for Ukrainian cities and there is no possibility to reproduce the technology by which they are obtained. This is due to the unavailability of sufficient high resolution satellite data information at the cities level, which is an integral part of the European methodology for obtaining the Urban Atlas. That is why we have proposed new approach on the base of open data which can be applicable to any other city. Kyiv (Ukraine) became the first city outside the Europe, for which the methodology by creating Urban Atlas was developed, which is compliant in structure and functionality to the European Copernicus Urban Atlas. The methodology was scaled for Lviv City, as well as applied and tested for other cities, in particular for Rivne, Irpin (Ukraine) and Lublin (Poland). In addition to the main management tasks that the Urban Atlas helps to solve, the obtained products can be used to unify and air quality monitoring in cities, and as a base for assessing the sustainable development goals indicator 11.6.2 “Annual mean levels of fine particulate matter in cities”.

Illegal logging in Ukraine is a big problem that negatively affects both environmental and socio-economic indicators of the country. The main reason for this problem is the lack of independent control over the forest industry. Lack of control, in turn, makes it possible to provide inaccurate information about the permitted logging and to hide the fact of logging. The solution to this problem is the use of modern approaches of Remote Sensing and deep learning to implement mechanisms for forestry monitoring and logging detection based on the satellite data. Most researches on satellite-based logging detection technology are based on the optical satellite missions. However, for countries with temperate and cold climates, the use of such approaches is problematic in winter and autumn due to the lack of vegetative biomass and the high percentage of clouds and snow in satellite images. In this study, we assessed a methodology for detecting logging based on optical and radar images of Copernicus satellite missions, namely Sentinel-l and 2. The obtained results show that when using this approach, it is possible to monitor and detect logging with high accuracy both in summer and in winter with the frequency of data updates once a week. The basis of this methodology is a convolutional neural network with U -Net architecture, which input is a stack of optical and radar images in summer and spring, and works on radar images only in winter and autumn.

In the past few years, deep learning methods developed and progressed in the many applied fields of science due to the appearance of large amounts of freely available data and improvement of computing resources. The task of crop mapping based on satellite data is not an exclusion. However, the appearance of clouds and shadows on the optical images cause difficulties on applying typical deep learning methods that recommend themselves in other issues. The easiest solution was to consider only images with a small percentage of clouds, but such way decreased the available data informativeness. In this study, we propose the new deep learning method based on recurrent neural network for efficient and precise crop mapping based on Sentinel-1 and Sentinel-2 imagery. The main idea of the study is to utilize all available information from the satellites and to provide an opportunity for neural network to extract the necessary features without any expert knowledge. Taking into account that regular feedforward neural networks could not deal effectively with such issue, authors suggest recurrent neural network with long short-term memory cells. Experimental results for the part of Kyiv region showed that this approach is rather efficient and outperformed traditional machine learning approaches and deep U-net architecture in terms of overall accuracy.

Ukrainian agricultural industry is one of the main sectors of economic growth. Nevertheless, Ukraine is way behind in the development. This is mostly due to the low level of modern technologies utilization by businesses and state entities. To ensure transparency, equity and reliability of Ukrainian land market, objective information on land use and crop state is required. The World Bank program “Supporting Transparent Land Governance in Ukraine” addresses these issues. Within the project, we performed satellite monitoring of land use in Ukraine, analyzed the feasibility of Google's cloud-based technology for processing large amount of data and developed a new platform to analyze the crop state using open and free Sentinel-1/2 satellite data. It is a 5-year project, which is extended for the whole country this year. We plan to make the technology of satellite monitoring operational and deployed in governmental institutions in 2023.

For accurate crop classification, it is necessary to use time-series of high-resolution satellite data to better discriminate among certain crop types. This task brings the following challenges: a large amount of satellite data for download, Big data processing and computational resources for utilization of state-of-the-art classification approaches. For solving these problems, we have developed an automated crop classification workflow, which is based on machine-learning techniques. By deployment of the workflow on the cloud platform, we can overcome challenges of Big data downloading and processing. In this paper, we present the system architecture and describe the experiments on structural and parametric identification of machine learning models utilized in the system.

For evaluating how far we are from achieving the Sustainable Development Goals and how big the progress is a global indicator framework was developed by the Inter-Agency and Expert Group on Sustainable Development Goals Indicators (IAEG-SDGs). In this paper, we propose an improved methodology for calculating indicator 2.4.1 "Proportion of agricultural area under productive and sustainable agriculture" within the ERA-PLANET Horizon 2020 project "The European Network for Observing our Changing Planet". The main improvements are in using accurate 10 m resolution crop classification map and Leaf Area Index derived from satellite data and WOFOST biophysical model. Based on proposed methodology indicator 2.4.1 has been calculated for the territory of Ukraine.

Crop classification maps based on high resolution remote sensing data are essential for supporting sustainable land management. The most challenging problems for their producing are collecting of ground based training and validation datasets, non-regular satellite data acquisition and cloudiness. To increase the efficiency of ground data utilization it is important to develop classifiers able to be trained on the data collected in the previous year. In this study, we propose new deep learning method for providing crop classification maps using in-situ data that has been collected in the previous year. Main idea of the study is to utilize deep learning approach based on sparse autoencoder. At the first stage it is trained on satellite data only and then neural network fine-tuning is conducted based on in-situ data form the previous year. Taking into account that collecting ground truth data is very time consuming and challenging task, the proposed approach allows us to avoid necessity for annual collecting in-situ data for the same territory. Experimental results for the territory of Ukraine show that this technique is rather efficient and provides reliable crop classification maps with overall accuracy higher than 85.9%.

In this paper, we propose a novel method for an object-based post-classification filtering, specifically tailored to improve agricultural land use maps. That has significant impact on the solving other applied tasks like detection of land cover changes and crop rotation violation, area estimation and crop yield forecasting. The main idea of this method is to divide classification map into separate objects (group of pixels with the same class value) and investigate the properties of them, taking into account the specificity of each class, independently. The most challenging task in post-classification filtering is preserving edges and boundaries between different fields. Often these boundaries are narrow and some traditional filters tend to treat this like noise and remove them. To deal with this, our method identifies boundaries of objects like crop fields, based on a modified version of the Sobel algorithm. The accuracy and effectiveness of our method has been tested and compared with other methods, based on accuracy assessments and visual comparison.

There are a lot of various satellite air quality products with coarse resolution at the moment. They are successfully used in numerous environmental applications, but still not sensitive enough to capture all the variability of air conditions that is necessary for Air Quality (AQ) monitoring in the city. A few months ago Sentinel-5 satellite has been launched able to provide higher resolution air quality products. Air quality is one of the priority areas within ERA-PLANET project of EU Horizon-2020 program. The overarching goal of ERA-PLANET is to strengthen the European Research Area in the domain of Earth Observation in coherence with the European participation to Group on Earth Observation (GEO) and the Copernicus. Within ERA-PLANET project of Horizon-2020 program it is planned to develop air quality monitoring service for urban areas based on remote sensing data and network of air quality sensors that can be used for real time AQ monitoring with high spatial and temporal resolution. Kyiv is selected as one of pilot cities for Smart City concept implementation. This study provides an analysis of existing satellite products and ground based observations in Kyiv and describes the concept of air quality monitoring in Kyiv Smart City project.

In this paper we propose method for land degradation indicators identification within the ERA-PLANET project. We also consider data sources that can be used for these purposes and consider the possibility of using high spatial resolution satellite-based classification maps. Project GEOEssential within the framework of the ERA-PLANET Project, using heterogeneous satellite data has the potential to make positive changes in land use. In this paper the ways of forming essential variables are described with specific attention to domain where and how it is possible to improve the state of productivity of land and to identify the non-proper use of land, which leads to desertification, forest degradation and the general degradation of land resources.

Data provided by synthetic aperture radar (SAR) of Sentinel satellite can be useful for many applications. However, as for any SAR image, speckle noise is present in acquired images. Speckle properties are important for different operations of SAR image processing as filtering, edge detection, segmentation, classification. Thus, we first carry out preliminary analysis of speckle statistics and show that speckle PDF is quite close to Gaussian whilst noise is of practically multiplicative nature. Second, spatial correlation properties of speckle are analyzed. The study is performed in local DCT domain. This is done since then the obtained 2D spectrum is employed in image despeckling based on DCT. Peculiarities of several possible approaches to despeckling are discussed. Several examples for one component and dual polarization data are presented.

Agriculture is one of the key areas where Remote Sensing (RS) techniques can be efficiently implemented for solving wide range of tasks (crop mapping, crop monitoring, crop yield forecasting etc.) on regular basis. Sentinel mission represents really new opportunities in agricultural domain - free of charge for non-commercial use satellite images with 10-20 m spatial resolution, 5-day revisit frequency with global coverage and compatibility to the Landsat missions. In this paper we present the results of Sentinel-2 national demonstration project in Ukraine executed during vegetation period of 2016 and coordinated by Universite catholique de Louvain (UCL). Within this demonstration Ukraine was selected as one of three sites for national demonstration due to high variability of agroclimatic conditions, relatively big fields and wide range of major crops over the territory of the country.

Ukraine is a large agricultural country situated in Eastern Europe (603,500 km2). Nowadays in Ukraine, there is no any land market due to the moratorium on land sales. Nevertheless, in all areas preparation for land market is undergoing. Cropland productivity assessment based on satellite data is a challenging task for Ukraine because of a large territory and big diversity of agricultural crops. Cropland productivity is one of the major factors for forming the land price. In this paper, we aim to provide land productivity maps based on analysis of MODIS and Landsat-8 data due to availability long term time-series of Normalized Difference Vegetation index (NDVI) from sensors aboard those remote sensing satellites. Taking into account the huge amount of satellite products to be analyzed, in the study we propose to exploit the Google Earth Engine (GEE) cloud platform. It was found that land productivity maps provided from MODIS data for different time periods are strongly correlated. The experiment shows that land productivity maps should have high resolution. That is why, Landsat-8 data is more appropriate for land market purpose, despite of some bias in values comparing to results based on MODIS data. Comparing crop mask from ESA Sen2Agri project and obtained results it was found the dependence of land productivity value and crop/non-crop cover. It was found that irrigated fields from the south part of the study area are the most productive lands in Ukraine.

For many applied problems in agricultural monitoring and food security it is important to provide reliable crop classification maps in national or global scale. Large amount of satellite data for large scale crop mapping generate a “Big Data” problem. The main idea of this paper was comparison of pixel-based approaches to crop mapping in Ukraine and exploring efficiency of the Google Earth Engine (GEE) cloud platform for solving “Big Data” problem and providing high resolution crop classification map for large territory. The study is carried out for the Joint Experiment of Crop Assessment and Monitoring (JECAM) test site in Ukraine covering the Kyiv region (North of Ukraine) in 2013. We found that Google Earth Engine (GEE) provided very good performance in enabling access to remote sensing products through the cloud platform, but our own approach based on ensemble of neural networks outperformed SVM, decision tree and random forest classifiers that are available in GEE.

Deep learning (DL) is a powerful state-of-the-art technique for image processing including remote sensing (RS) images. This letter describes a multilevel DL architecture that targets land cover and crop type classification from multitemporal multisource satellite imagery. The pillars of the architecture are unsupervised neural network (NN) that is used for optical imagery segmentation and missing data restoration due to clouds and shadows, and an ensemble of supervised NNs. As basic supervised NN architecture, we use a traditional fully connected multilayer perceptron (MLP) and the most commonly used approach in RS community random forest, and compare them with convolutional NNs (CNNs). Experiments are carried out for the joint experiment of crop assessment and monitoring test site in Ukraine for classification of crops in a heterogeneous environment using nineteen multitemporal scenes acquired by Landsat-8 and Sentinel-1A RS satellites. The architecture with an ensemble of CNNs outperforms the one with MLPs allowing us to better discriminate certain summer crop types, in particular maize and soybeans, and yielding the target accuracies more than 85% for all major crops (wheat, maize, sunflower, soybeans, and sugar beet).

Along the season crop classification based on satellite data is challenging task for Ukraine because of a big diversity of different agricultural crops with different phenology (crop calendars). Taking into account the availability for free of high resolution (10 to 30 meter) optical and SAR data from different satellite, the most resource consuming task is ground data collecting. That is why the proper time of ground surveys and crop classification maps developing is very important. In the study we propose to build three crop classification maps for JECAM Ukraine test site in Kyiv region during the vegetation season. The first one is built in the middle of May to classify winter cereals and rapeseeds. The next crop classification map is developing in July to discriminate major summer crops (spring cereals, maize, soybeans, sunflowers). The final crop map is built in autumn to refine summer crops and sugar beet discrimination. Time series of multi-temporal satellite images with restored missing (clouded and shadowed) data are classified using neural network approach, in particular ensemble of multi-layer perceptrons (MLPs). It is shown, that addition of satellite data from the end of previous year to the spring imagery allows to significantly improve the accuracy of winter crops classification. In July it is possible to deliver the map with major summer crops with overall accuracy higher than 87%, and the overall accuracy of final map at the end of the season is 94%.

In the paper we propose the methodology for solving the large scale classification and area estimation problems in the remote sensing domain on the basis of deep learning paradigm. It is based on a hierarchical model that includes self-organizing maps (SOM) for data preprocessing and segmentation (clustering), ensemble of multi-layer perceptrons (MLP) for data classification and heterogeneous data fusion and geospatial analysis for post-processing. The proposed methodology is applied for generation of high resolution land cover and land use maps for the territory of Ukraine from 1990 to 2010 and 2015.

Many applied Earth observation problems are based on land cover and land use maps, derived from satellite data. That is why it is important to assess their accuracy. We have developed retrospective regional 30 meter resolution land cover maps for Ukraine based on Landsat data for 1990, 2000 and 2010. As there is no reference data for validating retrospective periods, validation of the maps could be done only with photo-interpretation. In this paper we investigate two different sampling schemes for reference samples selection: pseudo-random (purposeful) samples selection (first approach) and systematic on regular grid (second approach). With systematic samples selection we receive the lower accuracy of classification (overall, user and producer), then with pseudo-random Nevertheless we consider the validation results with the systematic sampling scheme (the second approach) to be more reliable comparing to the first one, because the second sampling scheme is less subjective. Moreover, samples proportion within the second approach better corresponds to the statistics.

In this paper, we propose a new approach to pixel and parcel-based classification of multi-temporal optical satellite imagery. We first restore missing data due to clouds and shadows based on vector and raster data fusion in different phases of classification methodology. Pixel-based classification maps are derived from an ensemble of neural networks, in particular multilayer perceptrons (MLPs). The proposed approach is applied for regional scale crop classification using multi-temporal Landsat-8 images for the JECAM site in the Kyivska oblast of Ukraine in 2013. The obtained results on crop area estimates are also compared to official statistics.

In this paper we propose a new methodology to automatically generate retrospective high resolution land cover maps on a regular basis for the whole territory of Ukraine. An ensemble of neural networks, in particular multilayer perceptrons (MLPs), is used for multi-temporal Landsat-4/5/7 satellites imagery classification with previously restored missing data due to clouds, shadows and non-regular coverage. This methodology was used to obtain land cover maps for the territory of Ukraine for three decades, namely 1990s, 2000s and 2010s, with overall accuracy more than 97%.

In this paper, we propose an approach for estimation biophysical parameters, namely LAI effective, FAPAR, and FCOVER, based on in-situ and satellite measurements. In-situ data were collected during 2013-2014 within several field campaigns at the JECAM test site in Ukraine. We have built 30-meter resolution crop specific maps of biophysical parameters based on regression dependencies between ground measurements and NDVI derived from high resolution imagery (Landsat, SPOT) and Proba-V (100 m). In this paper, we discuss the best model selection for LAI effective, FAPAR and FCOVER mapping as well as selection of optimal source of satellite images. Obtained results are compared to available coarse resolution global biophysical products such as MODIS and SPOT-Vegetation.

In this paper, a new method for automatic identification of ground control points (GCPs) on optical remote sensing images is presented. An elastic Radial Basis Function (RBF) neural network based model for nonlinear coordinate transformation and image rectification is proposed. The new method can be used to produce dense fields of about thousands of GCPs per image to train highly deformable transformation models. As a result, an accuracy improvement of order of 4 in comparison with the Automated Precise Orthorectification Package (AROP) can be obtained. The proposed method is applied for the Ukrainian remote sensing satellite Sich-2. The obtained average RMSE error by the new method for Sich-2 images is estimated at 17.8 m.