Accurate information on crop spatial distribution is essential for modern agriculture, which faces increasing pressure from global food demand and climate change. Advances in Earth observation, satellite data availability, and machine learning (ML) have enabled effective large-scale crop monitoring. Traditional crop mapping relies on large labeled datasets, which are often costly and difficult to obtain, necessitating the use of alternative strategies. Transfer learning (TL) addresses this limitation by enabling models trained in data-rich regions to be adapted to data-scarce regions. In this study, we investigate TL for crop type classification using Random Forest (RF) and encoder-only Transformer with a classification head across three data modalities: Sentinel-1, Sentinel-2, and their combination. Satellite time series were interpolated to regular 15-day intervals to ensure consistent model input. Models were trained on a large labeled dataset from Slovenia and transferred to Serbia, where labeled data is limited. We evaluated off-the-shelfmodel transfer and fine-tuningstrategies. The bestperformance was achieved by a Transformer-based TL approach with a frozen feature extractor and fine-tuned classifier head, reaching an F1 score of 91% and outperforming the in-region supervised benchmark by 7%. These results demonstrate the potential of TL for cross-regional crop mapping in studied regions while substantially reducing reliance on large labeled datasets.
The performance of a low-cost GNSS receiver (ZED F9P) is evaluated using a zero-baseline configuration and compared with a geodetic-grade receiver through the use of different geodetic-grade and low-cost antenna models, as well as different GNSS constellation combinations. The study focuses on the analysis of ENU coordinates, bias, and total standard deviation, considering both single-constellation and multi-constellation configurations. The results demonstrate that low-cost GNSS receivers can achieve precision and accuracy comparable to those of geodetic-grade equipment when appropriately calibrated antennas are used. The GPS+GALILEO combination improves performance only when calibrated antennas are used, whereas the GLONASS and GALILEO constellations, processed independently or in combination, exhibit significant limitations when applied to low-cost receivers, preventing the attainment of robust solutions under the evaluated conditions. The results indicate that low-cost GNSS receivers constitute a viable alternative for topographic–geodetic applications, provided that the GPS constellation is available and that the project requirements and observation conditions allow it. Furthermore, it is shown that the inclusion of multiple constellations does not, by itself, guarantee an improvement in positioning precision when low-cost receivers are used.
Vietnam, a tropical nation, is frequently exposed to the risk of forest fires, driven by a combination of extreme weather events and anthropogenic activities. Forest fires present significant threats to ecosystems, human life, and infrastructure. Early detection and forecasting of forest fire risks play a pivotal role in mitigating the impact and facilitating effective responses to such disasters. This study investigates the application of advanced machine learning models, specifically Random Forest (RF), Convolutional Neural Network (CNN), and Long Short-Term Memory (LSTM), in conjunction with geospatial data, to construct a forest fire risk prediction map for Dak Lak province in the Central Highlands of Vietnam. The input dataset comprises 12 variables: vegetation cover, surface temperature, average monthly precipitation, wind speed, evapotranspiration, population density, land use, elevation, slope, aspect, proximity to roads, and proximity to hydrological features. These data were derived using remote sensing and GIS technologies, and integrated with historical fire occurrence points (6,088 fire locations) to model fire risk within the study area. The prediction performance of the models was evaluated using several metrics, including Precision, Recall, F1-score, Accuracy, and AUC, with a test dataset consisting of 122 fire points and 170 non-fire points. The results demonstrate that the RF algorithm (Precision: 0.84, Accuracy: 0.88, AUC: 0.95) outperforms both CNN (Precision: 0.80, Accuracy: 0.80, AUC: 0.84) and LSTM (Precision: 0.81, Accuracy: 0.83, AUC: 0.89) models in terms of forecasting forest fire risks in Dak Lak province.. The findings underscore the potential of leveraging advanced machine learning techniques and spatial data analysis to enhance the accuracy of forest fire risk prediction. This study’s outcomes are expected to assist policymakers and forest management authorities in developing more effective forest management strategies and fire prevention measures, thereby contributing to the protection of natural resources and the welfare of local communities.
This study presents a geodetic approach for monitoring structural deformations and sediment accumulation at the Plovdivtsi Dam in the Rhodope Mountains, southern Bulgaria. The monitoring system includes a geodetic network of observation pillars and benchmarks designed to detect horizontal and vertical displacements of the dam. Horizontal displacements are determined using trigonometric measurements, while vertical movements are monitored through precise geometric leveling. Sediment accumulation in the reservoir is assessed through periodic geodetic and hydrographic surveys using sediment profiling and echo-sounding techniques. The integrated methodology provides reliable geodetic data for detecting structural displacements and supporting the evaluation of dam behavior and changes in reservoir capacity. The proposed approach supports early detection of structural changes, improves the assessment of sediment dynamics, and contributes to effective water management and the long-term operational safety of hydraulic infrastructure.
