Conference Agenda

Hyperspectral satellite missions, equipped with contiguous visible-to-shortwave infrared spectral information, present unprecedented opportunities for retrieving a diverse range of vegetation traits with enhanced accuracy through novel retrieval techniques. In this context, we harnessed hyperspectral data cubes collected by the PRecursore IperSpettrale della Missione Applicativa (PRISMA) satellite, operated by the Italian Space Agency. Our objective is to develop and evaluate a hybrid retrieval workflow for crop trait mapping.

Vegetation property retrieval from optical data is inherently challenging due to its ill-posed nature. Multiple trait combinations can yield similar effects on the optical properties of the canopy. To address this, we leveraged non-parametric regression methods, well-suited for hyperspectral data. These models exploit the full spectral information to construct flexible, non-linear models. However, they face two significant limitations: the requirement of a substantial set of field data for effective model training and limited generalization capacity. The latter is crucial for future operational hyperspectral missions, where retrieval schemes must apply across diverse vegetation types and conditions. Non-parametric regression methods are particularly powerful and suited for hyperspectral data since they leverage all the spectral information to build flexible models in a non-linear way. The hybrid approaches combine the efficiency and adaptivity of the non-parametric approaches and the generic properties of the physically-based methods. However, crucial for these methods is the statistical distribution of the parameters of the Look-Up Tables (LUTs) generated by the radiative transfer model. This implies prior knowledge of the crops of which we want to estimate the parameters.

In this study, we explore the application of Domain Adaptation techniques to enhance the metrics of algorithms trained on LUTs generated by the PROSAIL radiative transfer model. Specifically, we investigate the adaptation of these algorithms to PRISMA hyperspectral images for the estimation of Leaf Area Index (LAI) and Chlorophyll (Chl) content. To validate the estimated values, we used different study sites in Europe and China, where in situ measurements are available along with PRISMA Hyperspectral images. The study sites are Maccarese near Rome, Italy and Youyi Farm, northeast China's in Heilongjiang Province.

We begin by generating a comprehensive LUT using the PROSAIL model, which combines the leaf optical properties model (PROSPECT-D) with the canopy bidirectional reflectance model (4SAIL). From this large LUT, we extract smaller sub-LUTs based on the specific geometry of PRISMA satellite acquisitions. The three critical parameters considered are the sun zenith angle, view zenith angle and relative azimuth angle. These sub-LUTs are tailored to match the acquisition conditions. Next, we employ various machine learning algorithms, including Gaussian Process Regression (GPR) and Partial Least Square Regression (PLSR). These models are trained on simulated spectra of the sub-LUTs and used to estimate crop biophysical variables on satellite images by applying the Domain Adaptation techniques.

Our research is currently in its early stages and the results are still being analyzed. They will be presented at the poster publication.

As one of the world's largest producers and consumers of grain, food security is essential for the country's political stability, economic development and social harmony. Keeping abreast of the biomass during wheat growth can help farmers and agricultural managers keep abreast of crop growth, identify problems in crop growth, and take corresponding management measures. Timely and accurate monitoring of winter wheat biomass is of great significance to ensure the stability of grain production and increase yield. At present, the main problems faced by remote sensing monitoring of crop biomass include the need for a large number of sample data, complex parameters, and poor temporal and spatial mobility. The CBA-Wheat Biomass Model (CBA-Wheat) is based on ground experimental data and a large number of measured data, which explains the relationship between the coefficient values of the ordinary least squares regression model and ZS of biomass and vegetation index at different growth stages of winter wheat. In this study, the parameters of the CBA-Wheat model were optimized by using the optimization algorithm to solve the problem that the existing model involves the amount of modeling data and it is difficult to apply and promote it at the satellite scale, and a winter wheat biomass estimation model suitable for MODIS data was obtained, and a 500-meter resolution winter wheat biomass dataset covering the main winter wheat producing areas in China was created by coupling ERA5 and MODIS multi-source remote sensing data. The results showed that the biomass estimation effect of the coupled CBA-Wheat model was better than that of other traditional models, with R2=0.64 and RMSE of 2.96 t/ha in the modeling dataset and R2=0.68 and RMSE=3.11 t/ha in the validation set. The RMSE of cross-validation accuracy for one year was 2.49-4.68 t/ha, and the RMSE of reserve area verification was 2.24-4.23 t/ha. The performance of the model was comprehensively evaluated by comparing it with the machine learning model and cross-validating it. The model has high inversion accuracy and is suitable for inversion during the whole growth period, and has good application potential in the prediction of large-area biomass in the region. In this study, a winter wheat biomass dataset was generated from 2009 to 2023 in China's main winter wheat producing areas, covering the period from early March to harvest every eight days, and the biomass characteristics of winter wheat were analyzed and discussed from two dimensions: space and time. Overall, the dataset is of great significance for agricultural production management and yield forecasting.

Pine wilt disease (PWD) is a lethal wilting disease caused by the pine wood nematode (Bursaphelenchus xylophilus; PWN). After becoming infected with the disease, pine trees show certain symptoms, in which the needles of the tree gradually change color, the pine resin stops flowing, and the tree wilts, withers and eventually dies, with the whole process occurring over a few months. Studies have shown that PWD has already become one of the most dangerous forest biological disasters in Asia and Europe, not only causing serious damage to pine forest resources and endangering the ecological security of scenic spots and other areas but also seriously affecting import and export trade.

At present, it is usually believed that there is currently no effective method to completely eradicate the disease, and the timely removal of the diseased trees is the most effective way to control the disease once the infected trees are identified. Using remote sensing technology to accurately locate the trees that are infected with pine wilt disease can enable the timely monitoring of the health status of the forest and the cleanup of the diseased trees while helping elucidate the prevalence pattern of the disease at multiple scales and its interactions with humans and the natural environment.

We established a spatiotemporal mining (STM) framework to characterize the spatial and temporal pattern of wilting process caused by PWD: (1) a spectral index was calculated for the two remote sensing images with similar dates in adjacent years; (2) a bi-temporal change analysis was used to obtain the differences between the calculated indices; and (3) the resulting image was enhanced through spatial filtering based on a proposed kernel which is fitted to the spatial pattern of the wilted pine trees. The STM focused on distinguishing between the wilted pine trees and other objects with similar spectral characteristics but diverse spatial and temporal patterns. Finally, a set of criteria were used to extract the wilted pine trees caused by PWD. The results were validated using the locations of the wilted pine trees recorded in the in-situ investigations. Meanwhile, A supervised classification method based on single-date imagery is also used for comparison.

The results showed that the producer’s accuracy for the two methods (the proposed STM and single-date supervised classification) were relatively close (84.7% and 83.3%), while the user’s accuracy for the STM-based method (86.9%) was significantly better than that of the single-date classification method (73.5%). In STM, dual-time-point analysis can effectively eliminate discolored deciduous trees, sparsely vegetated regions, and other ground objects that are easily confused with wilted pine trees, while spatial filtering combined with attribute filtering can also eliminate interference factors such as changes in ground feature types, radiation, and atmospheric differences introduced by dual-time-point analysis. When compared with single-date classification, the STM-based method has successfully reduced the omission frequency of wilted pine tree detection. By not limiting ourselves to a specific method, we are more focused on proposing a spatiotemporal feature extraction strategy that effectively address targets that cannot be distinguished solely by spectral information, with a goal toward advancing research on remote sensing monitoring of forest pests and diseases.

Wheat Fusarium Head Blight (FHB) not only damages the yield and quality of crops, but the toxins it produces also pose a serious threat to humans. Achieving precise monitoring of regional FHB is of great significance for ensuring the safety of grain production. This study is based on the survey data of FHB in 93 farmlands in Hefei, Chuzhou, and Bengbu cities of Anhui Province, China, in May 2021. A regional-scale FHB remote sensing monitoring model was constructed using the co-training semi-supervised learning (CTSS) method. The specific process of model construction and validation includes: 1) Extracting the average temperature, average humidity, and average rainfall from the flowering stage to the filling stage of wheat, as well as OSAVI, RDVI, and REHBI during the filling stage. These factors are selected as the environmental and host factors influencing the occurrence and development of FHB and participate in model construction. 2) Constructing a CTSS model using logistic regression (LR) and support vector machine (SVM) as base classifiers. Utilizing the CTSS model to extract valuable information from non-ground survey sample points in the study area, assisting ground survey points in jointly achieving the training of the wheat stripe rust remote sensing monitoring model. 3) Comparing the FHB remote sensing monitoring model constructed in this study with models solely based on LR or SVM for FHB remote sensing monitoring. Simultaneously, conducting a comparative analysis with models solely based on LR or SVM as base classifiers in CTSS. The FHB remote sensing monitoring model constructed in this study was found to perform the best, with a monitoring accuracy of 85%. This study indicates that non-ground survey sample points can provide effective information for constructing a regional-scale FHB remote sensing monitoring model. The approach of integrating non-ground survey sample points with ground survey points using a CTSS model in this study provides a reference for the regional-scale crop disease remote sensing monitoring under limited samples.

Land monitoring plays a vital role in modern farming practices and agricultural management. An accurate tracking of status and conditions for agricultural parcels could help farmers and landowners make informed decisions about irrigation, fertilisation, pest control and crop rotation choices. Moreover, it enables early detection of potential problems such as soil erosion, plant disease or invasive species infestations, allowing for timely intervention and mitigation measures. Consequently, this leads to promoting sustainable farming practices, improving crop yields and supporting the long-term health of both farmland and ecosystems.

One-off acquisitions of satellite images may not be sufficient to accurately identify land crop categories due to temporal variations in agricultural activities and other natural phenomena. Satellite image time series (SITS) data provide a comprehensive view over time, capturing seasonal changes, crop growth cycles and land dynamics, enabling the development of robust classification models and for a more informed decision-making process in agriculture and land management.

In this paper, we tackled the task of semantic segmentation of agricultural fields, using both optical and radar modalities. Our experiments made use of the PASTIS dataset, containing over 2.4k 128 x 128 time series, each acquisition encapsulating 10 relevant Sentinel-2 (S2) bands (out of the 13 bands provided by Sentinel-2, bands B1, B9 and B10 were excluded). We have also experimented with it's multimodal counterpart, namely PASTIS-R, containing corresponding Sentinel-1 (S1) time-series, in both ascending and descending orbit. The versatility of these datasets, acquired over the French metropolitan area, motivates their usage in training algorithms which could be further applied on a variety of environments. While S2-based SITS segmentation has been the most attempted approach in recent years, very few works succesfully address the multimodal fusion of optical and radar data (S1 + S2). In our study, we experimented with both trajectories, aiming to find a good trade-off between accuracy and inference speed. We tested multiple fusion techniques for S1 + S2 prediction, drawing conclusions regarding the best approach, while also proposing a technique to retrieve the most influential/non-influential timestamps, through a cross-attention module encapsulated in a mid-fusion strategy. Our experiments reveal that we can achieve comparable accuracy to state-of-the-art (SOA) methods, albeit with improved computational speed.

The development of fast and accurate automatic systems for crop segmentation is mostly driven by the need of rapid climate change adaptation, lowering the risk of future unpredictable losses, while also increasing resilience through the pre-emptive adoption of climate-smart agricultural techniques.

Multitemporal Synthetic Aperture Radar (SAR) Interferometry and SAR Tomography are powerful coherent processing tools of the satellite radar image series, with large applicability in Earth’s surface monitoring. Persistent Scatterers (PS) based approaches aim to identify the coherent targets, whose electromagnetic backscattering is correlated on acquisitions over all the acquisition period , thus allowing measuring, through the exploitation of phase models related to the geophysical parameters of interest, the topography with meter accuracy and the linear deformation rates with subwavelength accuracy – up to several mm/year. In SAR Tomography techniques, detection theory is exploited to identify the presence in the resolution cell (i.e. image pixel) of the single scatterers, defining binary hypothesis tests to discriminate between the absence and the presence of a stable target, with a fixed level of false alarm. Single Look Generalized Likelihood Test (SL-GLRT) is among the most popular processing techniques, being also a Constant False Alarm Rate (CFAR) detector, with the probability of false alarm independent of the noise power. The density of the measured points can be increased by analyzing also the Distributed Scatterers (DS) – targets with similar statistical behavior, present in low coherence areas. Their contribution spreads along multiple pixels, requiring multi-look processing techniques. Besides multi-look GLRT (MGLRT), recent research advancements in the direction of weak scatterers detection led to the development of SqueeSAR and CAESAR methods. Both techniques employ the processing of the sample covariance matrix, which is estimated from the multitemporal data.

The objective of the work consisted in the coherent processing of a multitemporal SAR images dataset, for estimation of residual topography and linear deformation rates of Constanta city, Romania. Located on the shore of the Black Sea, Constanta is one of the largest cities from the area, having also a rich historical heritage, being founded in 7^th Century BC as Tomis.

A Tomographic SAR processing involving images time series acquired by the Sentinel 1 satellites was constructed and exploited. A two-scale processing, , consisting in an initial pre-processing at lower resolution, of the of the multitemporal dataset and the subsequent high resolution analysis, has been applied.

The low-resolution processing exploits data whose resolution has been degraded through the use of moving average filters in order to mitigate the effect of noise. This step is necessary to allow an easier implementation of phase unwrapping that allows separating atmospheric phase contribution as well as distributed non-linear deformation acting on a large spatial scale. Estimation and compensation of these latter terms is mandatory for the implementation of Tomographic-based processing at full (or close-to-full) available resolution of the data. CAESAR algorithm has been chosen for the high resolution processing, to estimate the residual topography and subsidence. Since it employs the multi-look analysis, it can also extract information from decorrelating areas, characterized by the presence of DS. Our recent experiments indicated that CAESAR based detector has better performances than the one based on SqueeSAR method, in terms of probability of detection. Furthermore, its false alarm rate behavior is independent on the number of averaged looks, making its implementation more facile than the one of MGLRT, where the detection threshold needs to be adapted according to the multi-look degree.

The multi-temporal dataset consists of 183 SAR images of Constanta area, acquired by Sentinel-1 constellation, covering a temporal span of more than 4 years, between January 2020 and February 2024. The temporal distance between two consecutive acquisitions is 6 days in the interval when both Sentinel 1A and 1B satellites were operational – up to December 2021, and increased to 12 days after that date when only Sentinel 1A was operating. Images are acquired in Interferometric Wide swath (IW) mode, VV polarization, with the spatial resolution equal to approximately 5 m in slant range and 20 m in azimuth directions. Acquitions are carried out during descending passes of the satellite, with the look-angle equal to ~37 degrees. The sensor operates in C band, with the radar frequency of 5.4 GHz. The span of the dataset’s perpendicular baselines is 436 m, leading to a resolution of 55.21m in the elevation direction. The Rayleigh resolution in the temporal domain is equal to 0.68 cm/year.

Among all the available Sentinel-1 slice, the processing has been focused on two bursts belonging to the IW2 subswath. The test area contains the upper half of the city, including the old part and a shore region in which consolidation works were conducted several years ago due to the presence of subsidence. The dimensions of the test site are 500 pixels in azimuth and 2100 pixels in range directions.

CAESAR algorithm, used for estimation of scene’s parameters, exploits the Principal Components Analysis (PCA) of the data – more specifically, the eigendecomposition of the sample covariance matrix. The dominant component, associated to the highest eigenvalue, is separated, this process being equivalent to a filtering operation. A new data vector is constructed from the principal eigenvalue and eigenvector, being further processed for the identification of the equivalent PS along with the estimation of its parameters, using a SL-GLRT based detection rule. An adaptive multi-looking strategy has been adopted, based on the identification of Statistical Homogeneous Pixels (SHP) along the 3x11 search windows. The variable number of looks represents an optimal balance between the noise filtering (required in areas with DS) and preservation of spatial resolution and pixels’ statistics (useful for areas with PS). The SHP identification has been conducted based on Kolmogorov-Smirnov test. The CAESAR-based detector, CAESAR-D, has been implemented in the 4D space, along the residual topography and mean deformation rates dimensions. The search grids have been set according to the Rayleigh resolution values along both directions. The detection threshold has been set by analyzing the false alarm probability curves, derived with Monte Carlo simulations.

The poster presentation will include the maps of the estimated elevation and linear deformation rates in the detected PS.

Urban expansion is identified as a pivotal factor in the dynamics of the urban thermal environment. This study presents a novel machine learning application, within the Google Earth Engine (GEE) platform to assess urban expansion and its subsequent impact on thermal risk in two major cities: Beijing, China, and Athens, Greece. Utilizing Earth Observation data, the research identifies urban growth patterns. Neural networks are selected for their proven efficacy in classification tasks in complex environmental datasets. Urban expansion introduces significant thermal risks, including heat islands and altered local climates, which necessitate robust adaptation and mitigation strategies. The findings illustrate the relationship between urbanization and its thermal implications, underscoring the urgency of sustainable urban planning practices to combat these challenges. For this purpose, a tool designed to accurately map urban heat risk, previously developed during earlier phases of the DRAGON 5 project, will be employed to associate the dynamics of urban areas with their associated heat risks. The research highlights the replication potential of the employed methodology on a global scale across various urban settings to better understand and mitigate the thermal risks associated with urban expansion. This study contributes to more informed decision-making processes in urban planning and environmental management using advanced remote sensing techniques and machine learning algorithms.

Ports play a pivotal role in urban economies, serving as vital hubs for international trade and contributing significantly to the prosperity of cities. To monitor port changes, we need to retrieve similar ship images in these areas. Thus, we design a cross-modal hashing method based on a feature semi-interaction module and a semantic ranking objective function. Our method not only captures intricate correlations between different ship image modalities but also enables the construction of hash tables for large-scale retrieval. The semi-interaction module utilizes clustering centers from one modality to learn the correlations between two modalities and generate robust shared representations. The objective function optimizes these representations in a common Hamming space, consisting of a shared semantic alignment loss and a margin-free ranking loss. The alignment loss employs a shared semantic layer to preserve label-level similarity, while the ranking loss incorporates hard examples to establish a margin-free loss that captures similarity ranking relationships. We evaluate the performance of our method on benchmark datasets and demonstrate its effectiveness for cross-modal remote sensing ship image retrieval.

Building height is one of the most crucial vertical representations of urban morphology, intimately linked to the level of urbanization. However, the accurately mapping of large-scale distribution of building heights at a fine scale is still challenging, due to the prevalent saturation effects in height retrieval methods. This study proposes the Tri-Modal Height Estimation Network (TMHEN), which effectively integrates Sentinel-1 SAR, Sentinel-2 multispectral, and 10m panchromatic nighttime light data from SDGSAT-1 to achieve precise accurate estimation of building heights at a 10m resolution . Trained on a dataset collected from 62 cities nationwide, the model attains attained an accuracy of RMSE=4.77m and MAE=1.70m, outperforming existing studies with richer spatial details. We emphasize the significant contribution of nighttime light data to building height estimation, especially in enhancing the accuracy of predicting heights for high-rise buildings. Further research results indicate that the involvement of high-resolution nighttime light data of SDGSAT-1 effectively mitigates the underestimation issues for in high-rise buildings, reducing residuals by 17.98% and 13.01% compared to models without nighttime light data and with coarse-resolution nighttime light data, respectively. The map of building heights revealed vertical form differences and diverse urban development patterns and regional characteristics in five largest urban agglomerations in China. This study offers new perspectives for a comprehensive understanding of global urbanization levels and sustainable development.

Urbanization, indisputably acts as the impetus of global change, influencing facets of societal, economic, and environmental life. Monitoring the influence of urban human activities on the city entity is critical for sustainable management and planning. Historically, multispectral imaging applications were restricted by cloud cover weather conditions and thus applied only to low frequency or long-cycle urbanization studies, such as annual urban land-use changes, seasonal, cyclical urban activities, etc.

The advent of Synthetic Aperture Radar (SAR) sensors, offering round-the-clock terrestrial observational capabilities, has sparked interest in high-frequency urban change activities. This includes the particular observation of non-cyclical, high-frequency, abrupt urban activities such as building collapses, house demolitions and rebuilding, and commuting traffic, among other changes. Although current free and commercial satellites provide a wealth of SAR imaging data, they fail to perform cross-modal continuous observation and capture higher frequency change information. To elucidate, SAR image data from various sensors obtained at the same time demonstrate unique backscattering characteristics for the same target, making change detection difficult.

In light of this, our research proposes a framework for detecting urban change using SAR images acquired at different time points from heterogeneous sensors, further boosting SAR imagery's ability to capture high-frequency urban change information. Despite variances in imaging systems between different sensors causing scatterers without actual changes to display different scattering properties on two images, our innovative approach leverages deep learning models' capacity to learn and express high-level semantic features, creating cross-modal technologies. The primary goal hinges on obtaining authentic urban activity changes based on aligned heterogeneous modality information.

Initially, we collect observational images pairs with as short a sampling interval as possible to serve as co-occurring impact pairs and generate co-occurring data patch pairs for training and testing sets. Sequentially, image-to-image models such as Cycle-GAN, Conditional GAN, and Style-GAN are employed for image translation adversarial network models, translating one SAR image into another sensor's SAR image. Considering topographic differences and weather conditions could affect the backscattering characteristics of ground objects in various urban areas, we guide the cross-modal translation process by introducing environmental control parameters. Eventually, higher spatio-temporal granularity of urban observational information is obtained through the translation model, enabling higher frequency change information extraction utilising change detection methods.

The urban activity change information extracted by the proposed model deepens our understanding of urban activity change processes based on information decryption. We aim to capture the duration, intensity, and different stages of urban construction activities, such as construction stoppages, and resumptions, demolition, and reconstruction. Extracting such diverse construction activities reflects both the intensity and scope of urbanization activities and indirectly discloses the vitality of underlying economic activities. In summary, given that current multi-source free and commercial SAR satellites provide abundant multimodal image data but fail to directly perform change detection with images from different sensors, our study proposes using cross-modal image translation models to align modalities between different sensors, obtain higher temporal resolution image sequences and apply them to monitor urban activity changes.

Urbanization advances at unprecedented rates. The rapid expansion of urban land, i.e., urban sprawl, is associated with multiple negative effects on the environment and human well-being. To mitigate urban sprawl, informed and sustainable urban development strategies are crucial. However, these strategies are currently hampered by a lack of accurate information on the expansion of human settlements.

Remote sensing is an efficient tool to monitor urbanization at large scales. In recent years, algorithms based on deep learning techniques have achieved promising urban mapping and urban change detection results. For example, urban mapping methods have made significant advancements through the integration of multi-modal Earth observation data from Synthetic Aperture Radar (SAR) and optical sensors [1]. The combination of multi-modal Earth observation data with novel deep learning architectures has also led to improvements in urban change detection [2]. However, despite these promising results, many challenges remain. In particular, the presence of clouds often leads to a partial loss of optical data, which poses challenges for multi-temporal urban mapping. Furthermore, most urban mapping and change detection methods rely heavily on labeled data for supervised training. While labeled satellite data are time-consuming or costly to obtain, a plethora of unlabeled satellite data is freely available. Finally, although urban mapping and change detection methods are frequently combined using multi-task learning, the effective integration of urban maps and urban change detection outputs remains largely unaddressed.

The EO-AI4Urban project developed several methods addressing these limitations. First, we presented a multi-modal urban mapping method that utilizes a reconstruction network to approximate the features of the optical modality when only SAR data is available [3]. The method was evaluated on a multi-temporal urban mapping dataset featuring Sentinel-1 SAR and Sentinel-2 MSI data and achieved performance improvements over two baselines. Furthermore, we proposed a semi-supervised urban change detection method that exploits unlabeled Sentinel-1 SAR and Sentinel-2 MSI data [4]. Specifically, the proposed multi-modal Siamese network predicts urban changes and performs urban mapping for both timestamps. Additionally, a consistency loss penalizes inconsistent urban maps across sensor modalities on unlabeled data, leading to more robust features. The effectiveness of the proposed method was demonstrated under label-scarce conditions using Sentinel-1 SAR and Sentinel-2 MSI data and labels from the SpaceNet 7 dataset. Finally, our most recent work addresses the integration of urban mapping and urban change detection outputs in multi-task learning. To that end, we propose a novel multi-task integration module to find an optimal building map time series based on segmentation and change outputs. Preliminary results demonstrate that the proposed method detects urban changes accurately in satellite image time series acquired by the PlanetScope constellation (SpaceNet 7) and Gaofeng-2. Furthermore, we obtain accurate and up-to-date building maps for both platforms. Our experiments also demonstrate the effectiveness of the proposed method compared to related segmentation, change detection, and multi-task methods.

[1] Hafner, S., Ban, Y. and Nascetti, A., 2022. Unsupervised domain adaptation for global urban extraction using Sentinel-1 SAR and Sentinel-2 MSI data. Remote Sensing of Environment, 280, p.113192.

[2] Hafner, S., Nascetti, A., Azizpour, H. and Ban, Y., 2021. Sentinel-1 and Sentinel-2 data fusion for urban change detection using a dual stream U-Net. IEEE Geoscience and Remote Sensing Letters, 19, pp.1-5.

[3] Hafner, S. and Ban, Y., 2023, July. Multi-Modal Deep Learning for Multi-Temporal Urban Mapping with a Partly Missing Optical Modality. In IGARSS 2023-2023 IEEE International Geoscience and Remote Sensing Symposium (pp. 6843-6846). IEEE.

[4] Hafner, S., Ban, Y. and Nascetti, A., 2023. Semi-Supervised Urban Change Detection Using Multi-Modal Sentinel-1 SAR and Sentinel-2 MSI Data. Remote Sensing, 15(21), p.5135.