Nikita Firsov

Mr. Nikita Firsov

orcid

Research Focus & Metrics

Academic Career Length

3Years (2024-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
30 November 2024 in Sensors
Verified
Journal Article
Open access
PDF file
NFNikita Firsov
Nikita Firsov
EMEvgeny Myasnikov
Evgeny Myasnikov
VLValeriy Lobanov
Valeriy Lobanov
RKRoman Khabibullin
Roman Khabibullin
Nikolay KazanskiyNikolay Kazanskiy
Nikolay Kazanskiy
Artem NikonorovArtem Nikonorov
Artem Nikonorov
In traditional neural network designs, a multilayer perceptron (MLP) is typically employed as a classification block following the feature extraction stage. However, the Kolmogorov–Arnold Network (KAN) presents a promising alternative to MLP, offering the potential to enhance prediction accuracy. In this paper, we studied KAN-based networks for pixel-wise classification of hyperspectral images. Initially, we compared baseline MLP and KAN networks with varying numbers of neurons in their hidden layers. Subsequently, we replaced the linear, convolutional, and attention layers of traditional neural networks with their KAN-based counterparts. Specifically, six cutting-edge neural networks were modified, including 1D (1DCNN), 2D (2DCNN), and 3D convolutional networks (two different 3DCNNs, NM3DCNN), as well as transformer (SSFTT). Experiments conducted using seven publicly available hyperspectral datasets demonstrated a substantial improvement in classification accuracy across all the networks. The best classification quality was achieved using a KAN-based transformer architecture.

2 June 2025 in IEEE Access
Verified
Journal Article
Open access
AMAndrey Makarov
Andrey Makarov
IMIslombek Mirpulatov
Islombek Mirpulatov
NFNikita Firsov
Nikita Firsov
VLValeriy Lobanov
Valeriy Lobanov
Svetlana IllarionovaSvetlana Illarionova
Svetlana Illarionova
Artem NikonorovArtem Nikonorov
Artem Nikonorov
Hyperspectral imaging serves as a powerful tool for environmental studies, enabling the capture of significant properties in the objects being analyzed. These hidden properties may not be discernible through traditional RGB analysis, thus enhancing the distinguishability of target classes. Machine learning (ML) algorithms are widely utilized for processing hyperspectral data in tasks such as image classification or segmentation. Despite the rich hyperspectral feature space, ML-based approaches often struggle with the challenges posed by varying outdoor lighting conditions, leading to instability in the developed solutions. Furthermore, traditional segmentation methods designed for RGB images need to be adapted to ensure robust performance when applied to hyperspectral images (HSI). This work proposes a robust method for the classification of HSI in the outdoor environment under varying lighting conditions. A calibration procedure for hyperspectral sensor sensitivity has been proposed, ensuring improved quality of input data and, consequently, enhanced performance of models in solving hyperspectral image classification tasks. The paper also proposes a novel architecture, the Deep Spectral Spatial Transformer (DSST), which is employed as the classifier. This architecture leverages a deeper feature extractor to ensure robustness in various lightning conditions. Among these, standardization within individual channels yielded the best results. The applicability of the proposed approach is validated in the weed detection task in an agricultural field. The data collected were captured using a pushbroom hyperspectral sensor. A series of comparative experiments were conducted on a range of classification algorithms, encompassing both traditional ML algorithms and contemporary neural network architectures developed in recent years. The proposed DSST architecture demonstrated superior performance metrics, with values of 0.911 and 0.907 on F1-Score and Accuracy, respectively.

22 June 2026 in Preprints.org
Verified
Preprint
Open access
PDF file
NFNikita Firsov
Nikita Firsov
OTOlga Terekhova
Olga Terekhova
NONikita Odinets
Nikita Odinets
AFAlexey Fedotov
Alexey Fedotov
AMArtem Muzyka
Artem Muzyka
DSDmitry Sidorchuk
Dmitry Sidorchuk
Vision-language models (VLMs) are increasingly used in remote sensing (RS), yet their ability to perform abstract visual reasoning remains poorly understood. Existing RS benchmarks primarily evaluate perception-oriented capabilities, such as scene classification, object detection, image captioning, and visual question answering, providing limited insight into higher-level reasoning. To address this gap, we introduce BMRS, the first remote-sensing benchmark based on the Bongard problem paradigm. The benchmark comprises 122 problems constructed from satellite and aerial imagery and organized into seven categories: Shape, Semantic, Presence, Spatial, Size, Number, and Same. Unlike conventional perception benchmarks, BMRS requires models to infer abstract concepts from multiple images, thereby evaluating few-shot learning, concept induction, analogy making, and relational reasoning. To establish a human baseline, we conducted a study involving 113 participants, who achieved an average accuracy of 74.5%. We then evaluated three groups of models: proprietary large-scale VLMs, open-source general-purpose VLMs, and remote sensing vision-language models (RSVLMs). The strongest proprietary models, ChatGPT and Gemini, achieved accuracies of 89.3% and 86.9%, respectively, surpassing human performance. In contrast, the best open-source general-purpose model achieved 42.6% accuracy, while the strongest RSVLM reached only 20.5%. Analysis across problem categories revealed that Semantic problems were the easiest for both humans and models, indicating that object-level semantic understanding transfers effectively to the remote-sensing domain: several RSVLMs achieved performance comparable to substantially larger general-purpose models. Conversely, Spatial problems proved the most challenging, highlighting spatial-relational reasoning as a key limitation of current RSVLMs. BMRS provides a challenging benchmark for measuring progress toward reasoning-capable RSVLMs and offers new insights into the strengths and limitations of current model adaptation strategies.