Pylyp Prystavka

The application of the ecosystem concept to a field that is fundamentally different from ecology necessitates the identification of appropriate analogies. This is particularly relevant in relation to landscapes, chains of energy and material exchange (such as trophic chains), and nutrient cycles. Without establishing such parallels within software engineering, ecosystem-related studies risk being reduced to traditional systems analysis, leaving the ecosystem notion as little more than an appealing metaphor. The purpose of this article is to draw the attention of the software engineering community to ecosystem research and to show that software engineering ecosystems exist. Based on the concept of ecosystem in biology and taking into account the concept of software ecosystem, the previously introduced concept of software engineering ecosystem is reviewed. According to the nature of the landscape, all software engineering ecosystems can be divided into two types - ecosystems of human activity landscapes and ecosystems of software landscapes. In this article ecosystems of human activity landscapes are examined. The anthropos as new biotic component and the anthropogenic artifacts as new abiotic components in the software engineering ecosystem are introduced. The ecosystems of human activity landscapes should be considered as a population of organisms and, therefore, as a system of the supra-organismal level. In the context of the software engineering habitats of anthropos their activity leads to processes that can be a linear (chains), nonlinear (networks), or repetitive (cycles). The following processes are looking: software engineering life cycle; cycles of data, information and knowledge; software reuse cycle (this cycle is analogous to nutritional cycles); value chain/web - an analogue of energy chain; waste cycles (Reuse, Rework, and Recovery). These processes must be objects for the ecosystems study. The literature review of the state of the ecosystem research for the ecosystems of human activity landscapes has been carried out.

The aim of this study is to experimentally evaluate the performance of contemporary deep learning models under constrained computational resources and to establish formalized dependencies between inference speed and input batch size. The Raspberry Pi 4 Model B single-board computer was used as the computing platform, representing a typical example of low-power, resource-limited hardware employed onboard unmanned aerial vehicles (UAVs). In contrast to existing YOLO-based approaches that remain ill-suited for low-resource UAV platforms, this study proposes an alternative method focused on modeling classification time using linear regression. The results provide a framework for the development of onboard vision subsystems for domestic UAVs, capable of near-real-time image analysis under batch-processing constraints. This work presents, for the first time, linear regression models that describe the classification time of neural networks with varying architectures—specifically, mobile, balanced, and deep-level models— depending on the number of images in the input batch. The statistical significance of the resulting regression models has been experimentally validated, as well as their consistency with real-world measurements obtained during actual single-board computer operation. These regression equations enable inference time prediction without the need for repeated empirical testing, thus significantly improving the efficiency of neural system configuration in embedded environments. Based on the developed models, boundary conditions were identified for the applicability of each considered architecture to ensure near-real-time data processing. The results demonstrate that some models can handle relatively large data batches without exceeding critical time thresholds, whereas others, despite offering higher classification accuracy, exhibit excessive computational complexity and require hardware acceleration or optimization. The relevance of this research is driven by the growing demand for autonomous image analysis systems, particularly in the context of UAV deployment for military operations, reconnaissance, search and rescue missions, and monitoring applications. The proposed approach can be integrated into hardware-software systems to enable adaptive selection of neural architectures according to operational conditions and resource constraints. This creates a foundation for the further development of intelligent UAV systems with enhanced autonomy.

The object of the study is the process of aerial imagery data processing under limited computational resources, particularly onboard unmanned aerial vehicles (UAVs) using classification models. One of the most challenging issues is the adaptation of classification models to scale variations and perspective distortions that occur during UAV maneuvers. Additionally, the high computational complexity of traditional methods, such as sliding window approaches, significantly limits their applicability on resource-constrained devices. The study utilized state-of-the-art neural network classifiers, including ResNet50v2, DenseNet121, and MobileNetV2, which were fine-tuned on a specialized aerial imagery dataset. An experimental evaluation of the proposed neural network classifiers was conducted on Raspberry Pi 4 Model B and OrangePi 5 Pro platforms with limited computational power, simulating the constrained resources of UAV systems. To optimize performance, a stripe-based processing approach was proposed for streaming video, ensuring a balance between processing speed and the amount of analyzed data for surveillance applications. Specific execution time evaluations were obtained for different types of classifiers running on single-board computers suitable for UAV deployment. This approach enables real-time aerial imagery processing, significantly enhancing UAV system autonomy. Compared to traditional methods, the proposed solutions offer advantages such as reduced power consumption, accelerated computations, and improved classification accuracy. These results demonstrate high potential for implementation in various fields, including military operations, reconnaissance, search-and-rescue missions, and agricultural technology applications.

The object of this work is the recognition algorithms of aerial photography objects, namely, the analysis of recognition accuracy based on data sets with different aggregation classes. To solve this problem, an information system for object recognition based on aerial photography data has been developed. An architecture based on neural network architectures of the ConvNets group with structural modifications was chosen and used to create the information system. The use of a convolutional neural network of the ConvNets group in the architecture of the information system for the recognition of objects of aerial photography gives high accuracy rates when training the information system and validating its results. But the authors did not find any studies on the learning of the neural network of the ConvNets group. Therefore, it was decided to conduct an analysis in which case the ConvNets network will provide validation results with higher accuracy when the training takes place on datasets with or without class aggregation. The authors performed an analysis of the accuracy of recognition of aerial photography objects based on data sets with different aggregation classes. The dataset used for neural network training consisted of 3-channel labeled images of 64x64 pixels size. Based on the analysis, the optimal number of epochs for training is selected, which makes it possible to recognize aerial photography objects with greater accuracy and speed. It was concluded that greater accuracy in image classification is achieved for sampling without crossing data from different classes (without aggregation of classes). The result of the work is recommended for use in the automation of dataset filling and information filtering of visual images

The information technology that implements evaluation of redundant information using the methods of preprocessing and segmentation of digital images has been devised. The metrics for estimating redundant information containing a photo image using the approach based on texture variability were proposed. Using the example of aerial photography data, practical testing and research into the proposed assessment were carried out. Digital images, formed by various optoelectronic facilities, are distorted under the influence of obstacles of various nature. These obstacles complicate both the visual analysis of images by a human and their automatic processing. A solution to the problem can be obtained through preprocessing, which will lead to an increase in the informativeness of digital image data at a general decrease in content. An experimental study of the dependence of image informativeness on the results of overlaying previous filters for processing digital images, depending on the values of parameters of methods, was carried out. It was established that the use of algorithms sliding window analysis can significantly increase the resolution of analysis in the time area while maintaining a fairly high ability in the frequency area. The introduced metrics can be used in problems of computer vision, machine and deep learning, in devising information technologies for image recognition. The prospect is the task of increasing the efficiency of processing the monitoring results by automating the processing of the received data in order to identify informative areas. This will reduce the time of visual data analysis. The introduced metrics can be used in the development of automated systems of air surveillance data recognition.

Investigation of informative representations of complex data is a rapidly developing field of research in machine learning. In this work we present a process of production and analysis of informative low-dimensional latent representations of real-world image data with neural network models of unsupervised generative learning. A model of convolutional autoencoder based on VGG-16 architecture was used to produce low-dimensional latent representations of aerial image data and the characteristics of distributions of several higher-level classes of terrain types were studied. The analysis of distributions demonstrated a landscape of compact concept clusters for most studied types of terrain with good separation between concept regions. The results of this work can be used in developing methods of effective learning with minimal labeled data based on the emergent concept-sensitive structure in the latent representations.

In the paper it has been propounded and experimentally researched the linear operators that can be used to sharpen digital images or video frames distorted by a micro-motion of fixation chamber. It has been assumed that the cause of the problem, which leads to deterioration of sharpness, is a low- frequency interference, which is exemplified by a random non-recursive filter in the form of a discrete convolution. It has been experimentally proven that the proposed stabilizer filters allow for significant visual enhancement of distorted images, with a peak-to-peak signal-to-noise ratio for the improved images high- er than if using similar sharpening filters implemented in Adobe PhotoShop CS6. The corresponding masks of the proposed operators and the examples of application are given in the paper.

In military combat missions and/or rescue operations intelligence da- ta have a significant role. Last time using unmanned aircraft vehicles (UAV) is an effective way for its obtaining. As practice shows, the main aerial intelli- gence tasks are object detection and object tracking, and these tasks are desira- ble for automation. Existing UAVs either have no admitted task automation functionality at all or have particular functionality for civil purposes. This pub- lication describes automated complex with implemented potentially interesting objects search and the object-of-interest tracking functionality.