Linas Petkevičius

A major limitation is the scarcity of geospatial datasets that simultaneously provide multi-spectral imagery and descriptive captions. In particular, datasets containing aligned RGB, multispectral, and caption information remain highly limited. Therefore, we propose a full-circle pipeline to unify triplets of RGB images, image captions, and Sentinel-2–like multispectral data. To accomplish this, we combine a fine-tuned Stable Diffusion model with a Cycle GAN trained on generated images and the EuroSAT dataset. First, we use Qwen2-VL-2B as a zero shot method to generate captions for 675,993 images from the SkyScript dataset. We then fine-tune the Stable Diffusion 2-1 Base model on these image–caption pairs and generate randomly selected 123,081 RGB images conditioned on the Qwen2-VL-2B captions. Finally, we train a Cycle GAN on roughly 27,000 paired RGB and multispectral images and use it to translate synthetic RGB images into multispectral counterparts. In this way, textual prompts produce synthetic satellite imagery that can be converted to multispectral Sentinel-2 data. The pipeline enables unifying datasets that contain only captions or only RGB images by producing complete triplets (caption, RGB, multispectral). Quantitative evaluations support the credibility of the approach: generated captions achieve a SkyCLIP Score of 0.7312, the fine-tuned Stable Diffusion model achieves a CMMD of 0.245, and the Cycle GAN multispectral outputs reach a SAM of 10.16° in our synthetic dataset versus 13.94° on EuroSAT. The code, models and the dataset links are available at GitHub and Hugging Face.

This study evaluates the performance of the MiniCPM-V model on four distinct satellite image datasets: MAI, RSICD, RSSCN7, and a newly created merged dataset that combines these three. The merged dataset was developed to expand the generalization and variation of data distribution associated with the labeling and training processes inherent in satellite image analysis. We systematically collected prediction results for each individual dataset and conducted a comparative analysis against results reported in previous studies to benchmark the model's effectiveness. The findings indicate that large language models, such as MiniCPM-V, exhibit promising capabilities in the realm of satellite image recognition. On the RSSCN7 dataset, MiniCPMV achieved an accuracy of 70.57%, while on RSICD it reached 62.19%, on MAI 7.01%, and on the merged dataset 43.49%. Specifically, the model demonstrated mostly high accuracy (more than 80%) in identifying a majority of object classes across the datasets. Also we identified, it underperformed in accurately classifying certain object categories and recognizing all objects in multi labeled images, which suggests that while the model is robust overall, there are specific areas where its performance can be enhanced. Despite these limitations, the successful recognition of most objects underscores the potential of large language models in advancing satellite imagery analysis. These results highlight the significant potential of integrating large language models into remote sensing applications, offering a foundation for future research aimed at improving classification accuracy and expanding the range of detectable object classes by having caption level textual information.

Algal blooms are a common problem in inland waters, which raise growing awareness on monitoring lakes’ conditions. The on site monitoring is expensive and requires large human resources efforts. This work proposes remote monitoring techniques using satellite images and machine learning algorithms to predict chlorophyll concentration in water bodies and identify algal blooms. The training and test dataset used in this study includes diverse range of lakes in Baltic countries. The lake spectral features obtained from Sentinel-2 satellite images are used as predictors for proposed deep neural network models. The prediction of chlorophyll concentration with MAE 7.97 mg/m ³ and bloom vs. non-bloom classification with 71.6 % accuracy was achieved. The use of Bèzier curves for smoothing the point-wise prediction is proposed for identification of algal bloom characteristics: the bloom start date, end date, and duration. The results showed lake type impact on the blooming time. The experimental data and code are released with paper.

In breast cancer (BC), pathologists visually score ER, PR, HER2, and Ki67 biomarkers to assess tumor properties and predict patient outcomes. This does not systematically account for intratumoral heterogeneity (ITH) which has been reported to provide prognostic value. This study utilized digital image analysis (DIA) and computational pathology methods to investigate the prognostic value of ITH indicators in ER-positive (ER+) HER2-negative (HER2−) BC patients. Whole slide images (WSIs) of surgically excised specimens stained for ER, PR, Ki67, and HER2 from 254 patients were used. DIA with tumor tissue segmentation and detection of biomarker-positive cells was performed. The DIA-generated data were subsampled by a hexagonal grid to compute Haralick’s texture indicators for ER, PR, and Ki67. Cox regression analyses were performed to assess the prognostic significance of the immunohistochemistry (IHC) and ITH indicators in the context of clinicopathologic variables. In multivariable analysis, the ITH of Ki67-positive cells, measured by Haralick’s texture entropy, emerged as an independent predictor of worse BC-specific survival (BCSS) (hazard ratio (HR) = 2.64, p-value = 0.0049), along with lymph node involvement (HR = 2.26, p-value = 0.0195). Remarkably, the entropy representing the spatial disarrangement of tumor proliferation outperformed the proliferation rate per se established either by pathology reports or DIA. We conclude that the Ki67 entropy indicator enables a more comprehensive risk assessment with regard to BCSS, especially in cases with borderline Ki67 proliferation rates. The study further demonstrates the benefits of high-capacity DIA-generated data for quantifying the essentially subvisual ITH properties.

In this paper, we introduce the first use of symbolic integration that leverages the machine learning infrastructure, like automatic differentiation, for finding analytical approximations of ordinary and partial differential equations. Analytical solutions of differential equations are at the core of fundamental mathematical models, which, due to model complexity or non-linearity, very often cannot be found analytically. Traditionally, methods for solving them use hand-designed strategies, numerical methods, or iterative methods. We propose a method that is an application of the differentiable architecture search architecture for finding solutions to differential equations. We demonstrate our proposed method on a set of equations while simultaneously comparing it to numerical solutions from corresponding problems. We demonstrate that the proposed framework allows for solutions to various problems.

In this paper, long short-term memory (LSTM) and Transformer neural network models were developed for classification of different conveyor belt conditions (loaded and unloaded). Comparative shallow models such as logistic regression, support vector machine and random forest were also developed and summarized. Six different-length belt pressure signals were analyzed: 0.2, 0.4, 0.8, 1.6, 3.2, and 5.0 s. Both LSTM and Transformer models achieved 100% accuracy using pressure raw signal. Furthermore, LSTM model reached the highest classification level with the shortest signals. Accuracy and F1-score of 98% and 100% were reached using only 0.8 and 1.6 s-length signals, respectively. Also, LSTM model performed training and testing procedures faster than Transformer. Random forest model demonstrated the best classification level using aggregated signal data with accuracy of 85% and F1-score for loaded and unloaded conditions of 85% and 69%, respectively. Loaded conveyor belt condition was significantly easier to classify than the unloaded one in all models. Only LSTM showed better classification recall for unloaded conveyor belt condition using short signal. Experimental research dataset CORBEL (Conveyor belt pressure signal dataset) and models are open-sourced and accessible on GitHub https://github.com/TadasZvirblis/CORBEL .

The continued spread of electric vehicles raises new challenges for the supporting digital infrastructure. For example, long-distance route planning for such vehicles relies on the prediction of both the expected travel time as well as energy use. We envision a two-tier architecture to produce such predictions. First, a routing and travel-time-prediction subsystem generates a suggested route and predicts how the speed will vary along the route. Next, the expected energy use is predicted from the speed profile and other contextual characteristics, such as weather information and slope. To this end, the paper proposes deep-learning models that are built from EV tracking data. First, as the speed profile of a route is one of the main predictors for energy use, different simple ways to build speed profiles are explored. Next, eight different deep-learning models for energy-use prediction are proposed. Four of the models are probabilistic in that they predict not a single-point estimate but parameters of a probability distribution of energy use on the route. This is particularly relevant when predicting EV energy use, which is highly sensitive to many input characteristics and, thus, can hardly be predicted precisely. Extensive experiments with two real-world EV tracking datasets validate the proposed methods. The code for this research has been made available on GitHub.

Skills gained from vocational or higher education form an essential component of country’s economy, determining the structure of the national labor force. Therefore, knowledge on how people’s education converts to jobs enables data-driven choices concerning human resources within an ever-changing job market. Moreover, the relationship between education and occupation is also relevant in times of global crises, such as the COVID-19 pandemic. Healthcare system overload and skill shortage on one hand, and job losses related to lock-downs on the other, have exposed a necessity to identify target groups with relevant education backgrounds in order to facilitate their occupational transitions. However, the relationship between education and employment is complex and difficult to model. This study aims to propose the methodology that would allow us to model education-to-skill mapping. Multiple challenges arising from administrative datasets, namely imbalanced data, complex labeling, hierarchical structure and textual data, were addressed using six neural network-based algorithms of incremental complexity. The final proposed mathematical model incorporates the textual data from descriptions of education programs that are transformed into embeddings, utilizing transformer neural networks. The output of the final model is constructed as the hierarchical classification task. The effectiveness of the proposed model is demonstrated using experiments on national level data, which covers whole population of Lithuania. Finally, we provide the recommendations for the usage of proposed model. This model can be used for practical applications and scenario forecasting. Some possible applications for such model usage are demonstrated and described in this article. The code for this research has been made available on GitHub.