• Rural Digital Europe
• Publications close
• Компьютерная оптика close

This paper considers an issue of delineating agricultural fields in satellite images. In this task we follow a multi-temporal data approach. We show that on such data, good quality can be achieved using a simple low-parameter method. The method consists of a combination of a field detector and an edge detector. The field detection is based on an Otsu thresholding technique and for the edge detection we use a Canny detector. Facing a lack of available datasets and aiming to estimate the proposed algorithm, we prepared and published our dataset consisting of 18,859 expertly annotated fields using Sentinel-2 data. We implement one of the state-of-the-art deep-learning approaches and compare it with the proposed method on our dataset. The experiment shows the proposed simple multi-temporal algorithm to outperform the state-of-the-art instant data approach. This result confirms the importance of using multi-temporal data and for the first time demonstrates that the delineation problem can be solved at a lower cost without loss of quality. Our approach requires a significantly less amount of training data when compared with the NN-based one. The dataset of agricultural fields used in the work and the proposed algorithm implementation in Python are published in open access.

Various combinations of neural network parameters and sets of input data for satellite image classification are considered in the article. The training set is completed with a NDVI (normalized difference vegetation index) and local binary patterns. Testing of classifiers created on a different number of epochs and samples is carried out. Values of the neural network hyperparameters are determined that allow a classification accuracy of 0.70 and an F-measure of 0.65 to be achieved. Separation into classes with similar spectral characteristics is shown to offer low classification quality at different parameters and input data sets. Additional information is required. For example, for forests to be divided into more detailed classes, one needs to employ classifiers that use images from different seasons and vegetation periods. In addition, the training set needs to be extended to take into account various natural zones, soils, etc.

The paper considers a problem of classifying Sentinel-2 multispectral satellite images for environmental monitoring of the Baikal Natural Territory (BNT). The specificity of the BNT required the creation of a new set of 12 classes, which takes into account current problems. The set was formed in such a way that the areas corresponding to these classes completely covered the BNT. A training dataset was formed using a web interface based on Sentinel-2 satellite images. The classification of satellite images was carried out using Random Forest algorithms and the ResNet50 neural network. The accuracy of the calculations showed that the classification results can be used to solve actual problems of the Baikal natural territory, in particular, to analyze changes in the forestland, assess the impact of climate change on the landscape, analyze the dynamics of development activities, create farmland inventory, etc.

Autonomous navigation of unmanned vehicles (UVs) is currently one of the most interesting scientific and technical problems, and this is even more true for UVs moving across rough terrain. As a rule, such vehicles are equipped with different sensors operating simultaneously, and sophisticated software is being developed to collect and analyze miscellaneous data. For the joint use of data from multiple cameras and lidars, all of them need to be mutually referenced in a single coordinate system. This problem is solved in a process of system calibration. The general idea is to place an object of special type in the field of view of the sensors, so that its characteristic points can be automatically detected from different points of view by all sensors. Then the repeated survey of the object allows you to create the required number of tie points for mutual alignment of the sensors. This work presents a technique for automatic calibration of a system of cameras and lidars using an original calibration object. The presented results of the experiments show that the calibration accuracy is sufficiently high.

Classification of the land cover types from multi- and hyperspectral (HS) imagery is traditionally carried out on the basis of analysis of scatter plots of pixel values in a multidimensional feature space, which are used as brightness in individual channels. To increase the reliability of HS image classification, approaches are used based on simultaneously accounting for the characteristics of each pixel and the nearest-neighbor pixels, i.e., on the joint analysis of spectral and spatial features. The pixel neighborhood analysis is performed at various stages of the classification process. In this work, using a test hyperspectral image, the efficiency of spectral-spatial data classification methods that take into account spatial information at various stages of processing is studied. Special attention is paid to selecting the size of the spatial processing core. It is shown that the best results are obtained by combining pre-processing of raw data before performing the procedures of pixel-by-pixel spectral classification and post-processing of the resulting maps. Prospects of multi-scale smoothing of initial images, with the increase of the number of spectral-spatial features being multiple of the number of the scales, are shown.

A digital image processing algorithm based on sampling aerosol inhomogeneities was developed in the applied problem of laser remote sensing for measuring the velocity of wind. Tests of the developed algorithm were conducted for synthetic data from numerical simulations and data measured by a lidar. The algorithm developed performs processing of the field of aerosol backscattering coefficient in “Range-Time” coordinates and sufficiently increases the measurement accuracy in comparison with correlation methods.

A method for calculating the geomagnetic activity index Dst (Dst-index) based on a wavelet model of geomagnetic field variations is proposed. The method allows values of the Dst-index to be automatically obtained with a 1-minute resolution. The method is tested using data from equatorial stations [1]. The paper describes a calculation algorithm and presents estimation results. The calculation results are compared with the classical approach and the Kyoto method [2]. It is shown that the proposed method allows values of the Dst index to be obtained in the on-line mode with an admissible error.

Earth remote sensing data fusion is intended to produce images of higher quality than the original ones. However, the fusion impact on further thematic processing remains an open question because fusion methods are mostly used to improve the visual data representation. This article addresses an issue of the effect of fusion with increasing spatial and spectral resolution of data on thematic classification of images using various state-of-the-art classifiers and features extraction methods. In this paper, we use our own algorithm to perform multi-frame image fusion over optical remote sensing images with different spatial and spectral resolutions. For classification, we applied support vector machines and Random Forest algorithms. For features, we used spectral channels, extended attribute profiles and local feature attribute profiles. An experimental study was carried out using model images of four imaging systems. The resulting image had a spatial resolution of 2, 3, 4 and 5 times better than for the original images of each imaging system, respectively. As a result of our studies, it was revealed that for the support vector machines method, fusion was inexpedient since excessive spatial details had a negative effect on the classification. For the Random Forest algorithm, the classification results of a fused image were more accurate than for the original low-resolution images in 90% of cases. For example, for images with the smallest difference in spatial resolution (2 times) from the fusion result, the classification accuracy of the fused image was on average 4% higher. In addition, the results obtained for the Random Forest algorithm with fusion were better than the results for the support vector machines method without fusion. Additionally, it was shown that the classification accuracy of a fused image using the Random Forest method could be increased by an average of 9% due to the use of extended attribute profiles as features. Thus, when using data fusion, it is better to use the Random Forest classifier, whereas using fusion with the support vector machines method is not recommended.