講座主題：Analysis of SENTINEL-2 image series based on neural networks
專(zhuān)家姓名：Fedorov Roman K
工作單位：Russian Academy of Sciences
講座時(shí)間：2025年06月04日 10:30-11:30
講座地點(diǎn)：數(shù)學(xué)院大會(huì)議室341
主辦單位：煙臺(tái)大學(xué)數(shù)學(xué)與信息科學(xué)學(xué)院
內(nèi)容摘要：
This study presents a neural network-based analysis of Sentinel-2 satellite image series, focusing on land cover classification and temporal change detection. The methodology involves six key stages: (1) image markup, where satellite images are cataloged and annotated using polygon objects; (2) image preparation, including resolution standardization (10x10 m) and terrain feature extraction; (3) sample creation, generating 64x64-pixel tensor samples from non-background pixels; (4) clustering and balancing to homogenize sample distribution across classes; (5) model training using Random Forest, ConvNet, ResNet50, and LSTM architectures; and (6) large-scale classification of over 22,000 images (May–September, multi-year) on a GPU cluster (3090/4090/A100). The output, saved in GeoTIFF format, enables analysis of crop dynamics and land transitions. The workflow emphasizes automation, parallel processing, and multi-temporal evaluation, demonstrating scalable applications for environmental monitoring.
主講人介紹：
Fedorov R.K. 的全名為 Roman K. Fedorov（羅曼·K·費(fèi)多羅夫），是俄羅斯科學(xué)院西伯利亞分院計(jì)算中心（IDSCT SB RAS）的研究人員，主要研究方向?yàn)檫b感圖像處理與人工智能應(yīng)用。其工作聚焦衛(wèi)星影像的智能解析與大規(guī)模計(jì)算，成果直接服務(wù)于農(nóng)業(yè)監(jiān)測(cè)及生態(tài)變遷分析，體現(xiàn)了較強(qiáng)的工程落地能力。Fedorov R.K. 作為主要作者之一，參與開(kāi)發(fā)了基于神經(jīng)網(wǎng)絡(luò)的哨兵-2（Sentinel-2）衛(wèi)星影像時(shí)序分析技術(shù)。該技術(shù)通過(guò)六階段流程實(shí)現(xiàn)高效處理，包括影像標(biāo)注、多模態(tài)特征融合（光譜+紋理+地形）、動(dòng)態(tài)樣本生成、聚類(lèi)平衡、多模型訓(xùn)練（如ResNet50、LSTM）及大規(guī)模并行分類(lèi)，最終應(yīng)用于農(nóng)作物動(dòng)態(tài)監(jiān)測(cè)與地表覆蓋變遷分析。
