Prof. Dubravko Ćulibrk, PhD
Scalable Soil Carbon Platform
Organic carbon in soil (SOC) represents an important component of soil, as it plays a crucial role in maintaining soil fertility, promoting healthy plant growth, and mitigating climate change by serving as a carbon reservoir. Depleting SOC stocks can have negative consequences for the environment, agriculture, and food security. Therefore, it is important to manage land use practices and conservation methods to enable monitoring, improvement, and maintenance of SOC levels in the soil.
While monitoring changes in SOC is very relevant, it is also a costly and impractical procedure, as it involves conducting labor- and resource-intensive soil sampling campaigns, additional laboratory testing, and data processing.
To obtain an efficient tool for assessing SOC in the soil, we apply a standard segmentation artificial neural network (ANN) with U-net architecture, which is trained to assess land use based on multispectral satellite images from the Sentinel-2 satellite. Features derived from the ANN are then used as input for a machine learning model used to estimate SOC.
This approach allows us to harness the power of ANN to identify complex and nonlinear relationships in the data and, in combination with a limited number of field observations, train a machine learning model that enables us to achieve a high level of accuracy.
Assessment of Soil Moisture Content Based on Soil Properties and Weather Conditions
Soil moisture is a crucial parameter in plant growth, hydrological cycles, and climate dynamics, and it is of essential importance for many fields such as agriculture, meteorology, and disaster management. Understanding soil moisture in both extent and depth is key to improved irrigation planning in agriculture, accurate weather forecasting, and precise flood and/or drought risk management.
While the most accurate measurement of soil moisture content is achieved directly using hardware sensors, there are also indirect methods. The goal of this work is to create virtual sensors for predicting soil moisture content at greater depths (deeper than 5cm) based on deep learning models and data from meteorological stations.
For this purpose, exclusively the International Soil Moisture Network database was used, which includes soil moisture content, weather conditions (precipitation and temperature), and soil properties (saturation, clay content, organic carbon, sand content, and silt content).
Vegetation Control on and near Railway Tracks
Controlling vegetation on and near railway tracks is essential to ensure safe and timely train operations. This paper presents a real-time automated system developed for the needs of Serbian Railways, used for maintaining vegetation on railway tracks. This system enables precise herbicide spraying only where it is truly needed, unlike older vegetation control systems where herbicides are evenly sprayed along the railway tracks.
The system consists of a camera mounted on a locomotive, which records the track ahead of the train. The camera’s video feed is sent to a standard computer, where it is processed using convolutional neural network-based software. This software allows for the detection of weeds and shrubs on and near the railway tracks and sends signals to a PLC (Programmable Logic Controller) to activate the sprayers that disperse herbicide.
Active Learning with Self-Correcting Neural Networks (ALSCN)
Data labeling represents a major obstacle when training new models because the performance of machine learning models directly depends on the quality of the data. Labeling the dataset used for training such models often requires a significant amount of manual work. Labeling the entire training dataset is not always necessary because individual items in the dataset do not equally contribute to the training process.
The new active learning algorithm (ALSCN) consists of two networks, a convolutional neural network and a Self-Correcting neural network, which enables efficient selection of the best candidates for labeling from the unlabeled dataset. Experiments show that networks trained with a smaller number of such selected items outperform networks trained with the same number of randomly selected items from the complete dataset. Items from complete datasets are chosen in several iterations. The efficiency of the presented algorithm was tested on three datasets (MNIST, Fashion MNIST, and CIFAR 10). The final results show that manual labeling is only needed for 6.11% (3667/60000) of MNIST, 23.92% (14353/60000) of Fashion MNIST, and 59.4% (29704/50000 items) of CIFAR 10 datasets.
SCyDia - OCR for Serbian Accented Cyrillic
In the process of retrodigitizing Serbian dialectal dictionaries, the biggest obstacle is the lack of machine-readable versions of paper editions. Therefore, one essential step is needed before embarking on the process of creating dictionaries in a digital environment – OCR (Optical Character Recognition) processing of pages with the highest possible accuracy. OCR processing is not a new technology; there are many open-source and commercial software solutions that can reliably convert scanned images of paper documents into digital documents.
Available software solutions are usually efficient enough for processing scanned contracts, invoices, financial reports, newspapers, and books. In cases where documents containing accented text need to be processed, and each character with diacritics accurately extracted, such software solutions are not sufficiently efficient. ‘SCyDia’ is a web-based software solution that relies on the open-source ‘Tesseract’ software in the background. ‘SCyDia’ also includes a module for semi-automatic text correction. The results of processing 13 dialectal dictionaries show that accuracy ranges from 97.19% to 99.87%.
Forecasting Bitcoin with technical analysis: A not-so-random forest?
This paper uses data sampled at hourly and daily frequencies to predict Bitcoin returns. We consider various advanced non-linear models based on a multitude of popular technical indicators that represent market trend, momentum, volume, and sentiment. We run a robust empirical exercise to observe the impact of forecast horizon, model type, time period, and the choice of inputs (predictors) on the forecast performance of the competing models. We find that Bitcoin prices are weakly efficient at the hourly frequency. In contrast, technical analysis combined with non-linear forecasting models becomes statistically significantly dominant relative to the random walk model on a daily horizon. Our comparative analysis identifies the random forest model as the most accurate at predicting Bitcoin. The estimated measures of the relative importance of predictors reveal that the nature of investing in the Bitcoin market evolved from trend-following to excessive momentum and sentiment in the most recent time period.
A Hybrid Approach to Estimation of Soil Organic Carbon Based on Satellite Imagery in Agriculture
Servitization 4.0 as a Trigger for Sustainable Business: Evidence from Automotive Digital Supply Chain
The COVID-19 pandemic strengthens the use of digital services in the supply chains of manufacturers and suppliers in the automotive industry. Furthermore, the digitalization of the production process changed how manufacturing firms manage their value chains in the era of Industry 4.0. The automotive sector represents the ecosystem with rapid digital transformation, which provides a strong relationship between manufacturing firms in supply chains. However, there are many gaps in understanding how digital technologies and services could better shape relations between manufacturers and suppliers in the automotive industry. Accordingly, this study investigates the relations in deliveries of digital services in supply chains of the automotive industry. The data set was obtained through annual reports of the automotive firms, both from suppliers and manufacturers, between 2018 and 2020. From the network perspective, throughout the years, authors have used Social Network Analysis (SNA) method. SNA evaluates the relationship between actors (i.e., manufacturers and suppliers) in the use of services in their business models. The research results demonstrate how suppliers influence car manufacturers to deliver digital services to their customers. Finally, this study provides information that the combination of digital technologies with product-related services enables a stronger relationship between manufacturers and suppliers in the manufacturing ecosystem. These relations support the manufacturing ecosystem to survive the influence of different environments.
Tuning the configuration of a convolutional neural network to produce sharper land use/land coverage maps based on satellite imagery
Unlocking the black box: Non-parametric option pricing before and during COVID-19
This paper addresses the interpretability problem of non-parametric option pricing models by using the explainable artificial intelligence (XAI) approach. We study call options written on the S&P 500 stock market index across three market regimes: pre-COVID-19, COVID-19 market crash, and post-COVID-19 recovery. Our comparative option pricing exercise demonstrates the superiority of the random forest and extreme gradient boosting models for each market regime. We also show that the model’s pricing accuracy has worsened from the pre-COVID-19 to the recovery period. Moneyness was the most important price determinants across the market regimes, while the implied volatility and time-to-maturity inputs contributed intermittently to a lesser extent. During the COVID-19 crash, open interest gained more economic importance due to the increased behavioral tendencies of traders consistent with market distress.
Monitoring the Impact of Large Transport Infrastructure on Land Use and Environment Using Deep Learning and Satellite Imagery
Large-scale infrastructure, such as China–Europe Railway Express (CER-Express), which connects countries and regions across Asia and Europe, has a potentially profound effect on land use, as evidenced by changes in land cover along the railway. To ensure sustainable development of such infrastructure and appropriate land administration, effective ways to monitor and assess its impact need to be developed. Remote sensing based on publicly available satellite imagery represents an obvious choice. In the study presented here, we employ a state-of-the-art deep-learning-based approach to automatically detect different types of land cover based on multispectral Sentinel-2 imagery. We then use these data to conduct and present a study of the changes in land use in two geopolitically diverse regions of interest (in Serbia and China and with and without CER-Express infrastructure) for the period of the last three years. Our results show that the standard image-patch-based land cover classification approaches suffer a significant drop in performance in our target scenario in which each pixel needs to be assigned a cove class, but still, validate the applicability of the proposed approach as a remote sensing tool to support the sustainable development of large infrastructure. We discuss the technical limitations of the proposed approach in detail and potential ways in which it can be improved.
Quality Assessment of DIBR-synthesized views based on Sparsity of Difference of Closings and Difference of Gaussians
Images synthesized using depth-image-based-rendering (DIBR) techniques may suffer from complex structural distortions. The goal of the primary visual cortex and other parts of brain is to reduce redundancies of input visual signal in order to discover the intrinsic image structure, and thus create sparse image representation. Human visual system (HVS) treats images on several scales and several levels of resolution when perceiving the visual scene. With an attempt to emulate the properties of HVS, we have designed the no-reference model for the quality assessment of DIBR-synthesized views. To extract a higher-order structure of high curvature which corresponds to distortion of shapes to which the HVS is highly sensitive, we define a morphological oriented Difference of Closings (DoC) operator and use it at multiple scales and resolutions. DoC operator nonlinearly removes redundancies and extracts fine grained details, texture of an image local structure and contrast to which HVS is highly sensitive. We introduce a new feature based on sparsity of DoC band. To extract perceptually important low-order structural information (edges), we use the non-oriented Difference of Gaussians (DoG) operator at different scales and resolutions. Measure of sparsity is calculated for DoG bands to get scalar features. To model the relationship between the extracted features and subjective scores, the general regression neural network (GRNN) is used. Quality predictions by the proposed DoC-DoG-GRNN model show higher compatibility with perceptual quality scores in comparison to the tested state-of-the-art metrics when evaluated on four benchmark datasets with synthesized views, IRCCyN/IVC image/video dataset, MCL-3D stereoscopic image dataset and IST image dataset.
Vegetation suppression system on and near the railway tracks based on PLC and deep learning
Vegetation suppression on and near the railway tracks is very important to keep trains running safely and on time. This paper presents an automated real-time system, developed for the Serbian Railways company, that is used to maintain vegetation on and near railway tracks. This system allows precise spraying of herbicides only in locations where it is really needed, unlike older systems for controlling vegetation where herbicides are sprayed evenly along the railways tracks. The system consists of a camera mounted on a locomotive, which records the railway in front of the train, the video stream from the camera is sent to the standard PC and processed using software based on convolutional neural networks. This software allows the detection of weed and bushes on and near the railway tracks and sends signals to the PLC when it is the right moment to activate the herbicide sprayers.
Publication title: SCyDia - OCR for Serbian Cyrillic with Diacritics
In the currently ongoing process of retro-digitization of Serbian dialectal dictionaries, the biggest obstacle is the lack of machine-readable versions of paper editions. Therefore, one essential step is needed before venturing into the dictionary-making process in the digital environment -OCRing the pages with the highest possible accuracy. Successful retro-digitization of Ser-bian dialectal dictionaries, currently in progress, has shown a dire need for one basic yet necessary step, lacking until now-OCRing the pages with the highest possible accuracy. OCR processing is not a new technology, as many open-source and commercial software solutions can reliably convert scanned images of paper documents into digital documents. Available software solutions are usually efficient enough to process scanned contracts, invoices, financial statements, newspapers , and books. In cases where it is necessary to process documents that contain accented text and precisely extract each character with diacritics, such software solutions are not efficient enough. This paper presents the OCR software called “SCyDia”, developed to overcome this issue. We demonstrate the organizational structure of the OCR software “SCyDia” and the first results. The “SCyDia” is a web-based software solution that relies on the open-source software “Tes-seract” in the background. “SCyDia” also contains a module for semi-automatic text correction. We have already processed over 15000 pages, 13 dialectal dictionaries, and five dialectal monographs. At this point in our project, we have analyzed the accuracy of the “SCyDia” by processing 13 dialectal dictionaries. The results were analyzed manually by an expert who examined a number of randomly selected pages from each dictionary. The preliminary results show great promise, spanning from 97.19% to 99.87%.
Machine learning-based system for weed control on railways
In order to establish safe operation of the railway, it is very important to maintain vegetation on and near the railway. In this paper, a solution based on deep learning for detection of weeds and bushes on the railways and the control of intelligent herbicide sprayers using PLC, is presented. Hardware setup, software implementation, the configurations of the convolutional neural networks, and the datasets used to train the neural networks are described. A video stream with recording of the railway tracks in front of the train, is sent to a PC computer located in the cabin. This video stream is processed using software for detection of weeds and bushes based on multiple convolutional neural networks. This software sends signals to the PLC that controls the opening of the nozzles on the herbicide sprayer. The presented solution is developed for Serbian Railways company.