Pollen monitoring have become data-intensive in recent years as real-time detectors are deployed to classify airborne pollen grains. Machine learning models with a focus on deep learning, have an essential role in the pollen classification task. Within this study we developed an explainable framework to unveil a deep learning model for pollen classification. Model works on data coming from single particle detector (Rapid-E) that records for each particle optical fingerprint with scattered light and laser induced fluorescence. Morphological properties of a particle are sensed with the light scattering process, while chemical properties are encoded with fluorescence spectrum and fluorescence lifetime induced by high-resolution laser. By utilizing these three data modalities, scattering, spectrum, and lifetime, deep learning-based models with millions of parameters are learned to distinguish different pollen classes, but a proper understanding of such a black-box model decisions demands additional methods to employ. Our study provides the first results of applied explainable artificial intelligence (xAI) methodology on the pollen classification model. Extracted knowledge on the important features that attribute to the predicting particular pollen classes is further examined from the perspective of domain knowledge and compared to available reference data on pollen sizes, shape, and laboratory spectrofluorometer measurements.

Groundwater is one of the most important natural resources for reliable and sustainable water supplies in the world. To understand the use of water resources, the fundamental characteristics of groundwater need to be analyzed, but in many cases, in situ data measurements are not available or are incomplete. In this study, we used GIS and fuzzy analytic hierarchy process (FAHP) techniques for delineation of the groundwater potential zones (GWPZ) in the Titel Municipality (northern Serbia) based on quantitative assessment scores by experts (hydrologists, hydrogeologists, environmental and geoscientists, and agriculture experts). Six thematic layers, such as geology, geomorphology, slope, soil, land use/land cover, and drainage density were prepared and integrated into GIS software for generating the final map. The area falls into five classes: very good (25.68%), good (12.10%), moderate (15.18%), poor (41.34%), and very poor (5.70%). The GWPZ map will serve to improve the management of these natural resources to ensure future water protection and development of the agricultural sector, and the implemented method can be used in other similar natural conditions.

The drought in Serbia in the summer of 2017 heavily affected agricultural production, decreasing yields of maize, sunflower, soybean, and sugar beet. Monitoring moisture levels in crops can provide timely information about potential risk within a growing season, thus helping to create an early warning system for various stakeholders. The purpose of this study was to quantify the level of moisture stress in crops during summer and the consequences that it can have on yields. For that, maize and sunflower yield data provided by an agricultural company were used at specific parcels in the Backa region of Vojvodina province (Serbia) for 2017, 2018, 2019, and 2020. The crop moisture level was estimated at each parcel by calculating the normalized difference moisture index (NDMI) from Sentinel-2 data during the summer months (June–July–August). Based on the average NDMI value in July, the new crop moisture stress (CMS) index was introduced. The results showed that the CMS values at a specific parcel could be used for within-season estimation of maize and sunflower yield and the assessment of drought effects. The CMS index was tested for the current growing season of 2022 as an early warning system for yield reduction, demonstrating the potential to be included in a platform for digital agriculture, such as AgroSens, which is operational in Serbia.

CDR (Call Detail Record) data are one type of mobile phone data collected by operators each time a user initiates/receives a phone call or sends/receives an sms. CDR data are a rich geo-referenced source of user behaviour information. In this work, we perform an analysis of CDR data for the city of Milan that originate from Telecom Italia Big Data Challenge. A set of graphs is generated from aggregated CDR data, where each node represents a centroid of an RBS (Radio Base Station) polygon, and each edge represents aggregated telecom traffic between two RBSs. To explore the community structure, we apply a modularity-based algorithm. Community structure between days is highly dynamic, with variations in number, size and spatial distribution. One general rule observed is that communities formed over the urban core of the city are small in size and prone to dynamic change in spatial distribution, while communities formed in the suburban areas are larger in size and more consistent with respect to their spatial distribution. To evaluate the dynamics of change in community structure between days, we introduced different graph based and spatial community properties which contain latent footprint of human dynamics. We created land use profiles for each RBS polygon based on the Copernicus Land Monitoring Service Urban Atlas data set to quantify the correlation and predictivennes of human dynamics properties based on land use. The results reveal a strong correlation between some properties and land use which motivated us to further explore this topic. The proposed methodology has been implemented in the programming language Scala inside the Apache Spark engine to support the most computationally intensive tasks and in Python using the rich portfolio of data analytics and machine learning libraries for the less demanding tasks.

Tomato is an important commercial product which is perishable by nature and highly susceptible to fungal incidence once it is harvested. Not all tomatoes are equally vulnerable to pathogenic fungi, and an early detection of the vulnerable ones can help in taking timely preventive actions, ranging from isolating tomato batches to adjusting storage conditions, but also in making right business decisions like dynamic pricing based on quality or better shelf life estimate. More importantly, early detection of vulnerable produce can help in taking timely actions to minimize potential post-harvest losses. This paper investigates Near-infrared (NIR) hyperspectral imaging (1000–1700 nm) and machine learning to build models to automatically predict the susceptibility of sepals of recently harvested tomatoes to future fungal infections. Hyperspectral images of newly harvested tomatoes (cultivar Brioso) from 5 different growers were acquired before the onset of any visible fungal infection. After imaging, the tomatoes were placed under controlled conditions suited for fungal germination and growth for a 4-day period, and then imaged using normal color cameras. All sepals in the color images were ranked for fungal severity using crowdsourcing, and the final severity of each sepal was fused using principal component analysis. A novel hyperspectral data processing pipeline is presented which was used to automatically segment the tomato sepals from spectral images with multiple tomatoes connected via a truss. The key modelling question addressed in this research is whether there is a correlation between the hyperspectral data captured at harvest and the fungal infection observed 4 days later. Using 10-fold and group k-fold cross-validation, XG-Boost and Random Forest based regression models were trained on the features derived from the hyperspectral data corresponding to each sepal in the training set and tested on hold out test set. The best model found a Pearson correlation of 0.837, showing that there is strong linear correlation between the NIR spectra and the future fungal severity of the sepal. The sepal specific predictions were aggregated to predict the susceptibility of individual tomatoes, and a correlation of 0.92 was found. Besides modelling, focus is also on model interpretation, particularly to understand which spectral features are most relevant to model prediction. Two approaches to model interpretation were explored, feature importance and SHAP (SHapley Additive exPlanations), resulting in similar conclusions that the NIR range between 1390–1420 nm contributes most to the model’s final decision.

Habitat degradation, mostly caused by human impact, is one of the key drivers of biodiversity loss. This is a global problem, causing a decline in the number of pollinators, such as hoverflies. In the process of digitalizing ecological studies in Serbia, remote-sensing-based land cover classification has become a key component for both current and future research. Object-based land cover classification, using machine learning algorithms of very high resolution (VHR) imagery acquired by an unmanned aerial vehicle (UAV) was carried out in three different study sites on Mt. Stara Planina, Eastern Serbia. UAV land cover classified maps with seven land cover classes (trees, shrubs, meadows, road, water, agricultural land, and forest patches) were studied. Moreover, three different classification algorithms—support vector machine (SVM), random forest (RF), and k-NN (k-nearest neighbors)—were compared. This study shows that the random forest classifier performs better with respect to the other classifiers in all three study sites, with overall accuracy values ranging from 0.87 to 0.96. The overall results are robust to changes in labeling ground truth subsets. The obtained UAV land cover classified maps were compared with the Map of the Natural Vegetation of Europe (EPNV) and used to quantify habitat degradation and assess hoverfly species richness. It was concluded that the percentage of habitat degradation is primarily caused by anthropogenic pressure, thus affecting the richness of hoverfly species in the study sites. In order to enable research reproducibility, the datasets used in this study are made available in a public repository.