Research on Land Use and Land Cover Information Extraction Methods for Remote Sensing Images Based on Improved Convolutional Neural Networks DOI Creative Commons

Xue Ding,

Wang Zhaoqian,

Shuangyun Peng

et al.

ISPRS International Journal of Geo-Information, Journal Year: 2024, Volume and Issue: 13(11), P. 386 - 386

Published: Oct. 31, 2024

To address the challenges that convolutional neural networks (CNNs) face in extracting small objects and handling class imbalance remote sensing imagery, this paper proposes a novel spatial contextual information multiscale feature fusion encoding–decoding network, SCIMF-Net. Firstly, SCIMF-Net employs an improved ResNeXt-101 deep backbone significantly enhancing extraction capability of object features. Next, PMFF module is designed to effectively promote features at different scales, deepening model’s understanding global local information. Finally, introducing weighted joint loss function improves performance LULC under conditions. Experimental results show compared other CNNs such as Res-FCN, U-Net, SE-U-Net, U-Net++, PA by 0.68%, 0.54%, 1.61%, 3.39%, respectively; MPA 2.96%, 4.51%, 2.37%, 3.45%, MIOU 3.27%, 4.89%, 4.2%, 5.68%, respectively. Detailed comparisons locally visualized indicate can accurately extract from imbalanced classes objects.

Language: Английский

Temporal Optimisation of Satellite Image‐Based Crop Mapping: A Comparison of Deep Time Series and Semi‐Supervised Time Warping Strategies DOI Creative Commons
Robert Finnegan,

Julian D. Metcalfe,

Sara Sharifzadeh

et al.

IET Computer Vision, Journal Year: 2025, Volume and Issue: 19(1)

Published: Jan. 1, 2025

ABSTRACT This study presents a novel approach to crop mapping using remotely sensed satellite images. It addresses the significant classification modelling challenges, including (1) requirements for extensive labelled data and (2) complex optimisation problem selection of appropriate temporal windows in absence prior knowledge cultivation calendars. We compare lightweight Dynamic Time Warping (DTW) method with heavily supervised Convolutional Neural Network ‐ Long Short‐Term Memory (CNN‐LSTM) high‐resolution multispectral optical imagery (3 m/pixel). Our integrates effective practical preprocessing steps, augmentation data‐driven strategy window, even presence numerous classes. findings demonstrate that DTW, despite its lower demands, can match performance CNN‐LSTM through our steps while significantly improving runtime. These results both DTW achieve deployment‐level accuracy underscore potential as viable alternative more resource‐intensive models. The also prove effectiveness windowing runtime study, no planting timeframes.

Language: Английский

Citations

0
FarmSeg_VLM: A farmland remote sensing image segmentation method considering vision-language alignment DOI
Haiyang Wu, W. Mu, Dandan Zhong

et al.

ISPRS Journal of Photogrammetry and Remote Sensing, Journal Year: 2025, Volume and Issue: 225, P. 423 - 439

Published: May 13, 2025

Language: Английский

Citations

0
Orthophoto-Based Vegetation Patch Analyses—A New Approach to Assess Segmentation Quality DOI Creative Commons
Witold Maćków, Malwina Bondarewicz, Andrzej Łysko

et al.

Remote Sensing, Journal Year: 2024, Volume and Issue: 16(17), P. 3344 - 3344

Published: Sept. 9, 2024

The following paper focuses on evaluating the quality of image prediction in context searching for plants a single species, using example Heracleum sosnowskyi Manden, given area. This process involves simplified classification that ends with segmentation step. Because particular characteristics environmental data, such as large areas plant occurrence, significant partitioning population, or individual, use standard statistical measures Accuracy, Jaccard Index, Dice Coefficient does not produce reliable results, shown later this study. issue demonstrates need new method assessing betted adapted to unique vegetation patch detection. main aim study is provide metric and demonstrate its usefulness cases discussed. Our proposed introduces two coefficients, M+ M−, which, respectively, reward true positive regions penalise false regions, thus providing more nuanced assessment quality. effectiveness has been demonstrated different scenarios focusing variations spatial distribution fragmentation theoretical patches, comparing traditional metrics. results indicate our offers flexible accurate quality, especially involving complex data. aims applicability real-world detection tasks.

Language: Английский

Citations

0
Research on Land Use and Land Cover Information Extraction Methods for Remote Sensing Images Based on Improved Convolutional Neural Networks DOI Creative Commons

Xue Ding,

Wang Zhaoqian,

Shuangyun Peng

et al.

ISPRS International Journal of Geo-Information, Journal Year: 2024, Volume and Issue: 13(11), P. 386 - 386

Published: Oct. 31, 2024

To address the challenges that convolutional neural networks (CNNs) face in extracting small objects and handling class imbalance remote sensing imagery, this paper proposes a novel spatial contextual information multiscale feature fusion encoding–decoding network, SCIMF-Net. Firstly, SCIMF-Net employs an improved ResNeXt-101 deep backbone significantly enhancing extraction capability of object features. Next, PMFF module is designed to effectively promote features at different scales, deepening model’s understanding global local information. Finally, introducing weighted joint loss function improves performance LULC under conditions. Experimental results show compared other CNNs such as Res-FCN, U-Net, SE-U-Net, U-Net++, PA by 0.68%, 0.54%, 1.61%, 3.39%, respectively; MPA 2.96%, 4.51%, 2.37%, 3.45%, MIOU 3.27%, 4.89%, 4.2%, 5.68%, respectively. Detailed comparisons locally visualized indicate can accurately extract from imbalanced classes objects.

Language: Английский

Citations

0