Comparison of RFE-DL and stacking ensemble learning algorithms for classifying mangrove species on UAV multispectral images

International Journal of Applied Earth Observation and Geoinformation, Journal Year: 2022, Volume and Issue: 112, P. 102890 - 102890

Published: June 30, 2022

Mangroves are highly productive wetland ecosystems, located at the interlocking area of tropical and subtropical coastal zones. Accurately mapping distribution, quality quantity species crucial for mangrove management, protection, restoration. This study proposed a approach by combining recursive feature elimination (RFE) with deep learning (DL) algorithms, further assess effectiveness selection DL (DeeplabV3+ PSPNet) algorithm to improve classification accuracy under high dimensional UAV image datasets. We constructed an ensemble models (SEL) stacking five base (Random Forest, XGBoost, LightGBM, CatBoost, AdaBoost), evaluate ability between SEL RFE-DL algorithms. Comparison classifications was differences models. Results indicated that: (1) RFE could using optimal features achieved 94.8% overall (OA), which 0.2%-8.5% higher than only original multispectral bands; (2) produced better performance 1.6%-12.7% accuracy. Mcnemar's test showed were significant three algorithms; (3) had strong stable classifying species. The OA six scenarios from 75.5% 92.2%, highest 0.8%-4.2% models; (4) XGBoost importance, while AdaBoost lowest importance in SEL-based classifications.

Language: Английский

---

Urban Tree Classification Based on Object-Oriented Approach and Random Forest Algorithm Using Unmanned Aerial Vehicle (UAV) Multispectral Imagery

Remote Sensing, Journal Year: 2022, Volume and Issue: 14(16), P. 3885 - 3885

Published: Aug. 11, 2022

Timely and accurate information on the spatial distribution of urban trees is critical for sustainable development, management planning. Compared with satellite-based remote sensing, Unmanned Aerial Vehicle (UAV) sensing has a higher temporal resolution, which provides new method identification trees. In this study, we aim to establish an efficient practical tree by combining object-oriented approach random forest algorithm using UAV multispectral images. Firstly, image was segmented multi-scale segmentation based scale determined Estimation Scale Parameter 2 (ESP2) tool visual discrimination. Secondly, spectral features, index texture features geometric were combined form schemes S1–S8, S9, consisting selected recursive feature elimination (RFE) method. Finally, classification performed nine (RF), support vector machine (SVM) k-nearest neighbor (KNN) classifiers, respectively. The results show that RF classifier performs better than SVM KNN, achieves highest accuracy in overall (OA) 91.89% Kappa coefficient (Kappa) 0.91. This study reveals have negative impact classification, other three types positive impact. importance ranking map shows are most important type followed features. Most species high accuracy, but Camphor Cinnamomum Japonicum much lower species, suggesting cannot accurately distinguish these two so it necessary add such as height future improve accuracy. illustrates combination images powerful classification.

Language: Английский

---

The Effects of Spatial Resolution and Resampling on the Classification Accuracy of Wetland Vegetation Species and Ground Objects: A Study Based on High Spatial Resolution UAV Images

Drones, Journal Year: 2023, Volume and Issue: 7(1), P. 61 - 61

Published: Jan. 15, 2023

When employing remote sensing images, it is challenging to classify vegetation species and ground objects due the abundance of wetland high fragmentation objects. Remote images are classified primarily according their spatial resolution, which significantly impacts classification accuracy However, there still some areas for improvement in study effects resolution resampling on results. The area this paper was core zone Huixian Karst National Wetland Park Guilin, Guangxi, China. aerial (Am) with different resolutions were obtained by utilizing UAV platform, resampled (An) pixel aggregation method. In order evaluate impact accuracy, Am An utilized based geographic object-based image analysis (GEOBIA) method addition various machine learning classifiers. results showed that: (1) multi-scale both optimal scale parameter (SP) processing time decreased as diminished multi-resolution segmentation process. At same SP greater than that Am. (2) case An, appropriate feature variables different, spectral texture features more significant those (3) classifiers exhibited similar trends ranging from 1.2 5.9 cm, where overall increased then accordance decrease resolution. Moreover, higher An. (4) at scales, differed between

Language: Английский

---

Monitoring detailed mangrove hurricane damage and early recovery using multisource remote sensing data

Journal of Environmental Management, Journal Year: 2022, Volume and Issue: 320, P. 115830 - 115830

Published: Aug. 6, 2022

Language: Английский

---

A Survey on Applications of Unmanned Aerial Vehicles Using Machine Learning

IEEE Access, Journal Year: 2023, Volume and Issue: 11, P. 117582 - 117621

Published: Jan. 1, 2023

Unmanned Aerial Vehicles (UAVs) play an important role in many applications, including health, transport, telecommunications and safe rescue operations. Their adoption can improve the speed precision of applications when compared to traditional solutions based on handwork. The use UAVs brings scientific technological challenges. In this context, Machine Learning (ML) techniques provide several problems concerning civil military applications. An increasing number papers ML context have been published academic journals. work, we present a literature review UAVs, outlining most recurrent areas commonly used UAV results reveal that environment, communication security are among main research topics.

Language: Английский

---

Deep and Machine Learning Image Classification of Coastal Wetlands Using Unpiloted Aircraft System Multispectral Images and Lidar Datasets

Remote Sensing, Journal Year: 2022, Volume and Issue: 14(16), P. 3937 - 3937

Published: Aug. 13, 2022

The recent developments of new deep learning architectures create opportunities to accurately classify high-resolution unoccupied aerial system (UAS) images natural coastal systems and mandate continuous evaluation algorithm performance. We evaluated the performance U-Net DeepLabv3 convolutional network two traditional machine techniques (support vector (SVM) random forest (RF)) applied seventeen land cover types in west Florida using UAS multispectral imagery canopy height models (CHM). Twelve combinations spectral bands CHMs were used. Our results showed that (83.80–85.27% overall accuracy) DeepLabV3 (75.20–83.50% outperformed SVM (60.50–71.10% RF (57.40–71.0%) algorithms. addition CHM slightly increased accuracy as a whole models, while notably improved results. Similarly, outside three bands, namely, near-infrared red edge, classifiers but had minimal impact on classification difference accuracies produced by UAS-based lidar SfM point clouds, supplementary geometrical information, process was across all techniques. highlight advantage networks highly diverse landscapes. also found low-cost, three-visible-band produces comparable do not risk significant reduction when adopting models.

Language: Английский

---

Karst vegetation coverage detection using UAV multispectral vegetation indices and machine learning algorithm

Plant Methods, Journal Year: 2023, Volume and Issue: 19(1)

Published: Jan. 23, 2023

Abstract Background Karst vegetation is of great significance for ecological restoration in karst areas. Vegetation Indices (VIs) are mainly related to plant yield which helpful understand the status Recently, surveys have gradually shifted from field remote sensing-based methods. Coupled with machine learning methods, Unmanned Aerial Vehicle (UAV) multispectral sensing data can effectively improve detection accuracy and extract important spectrum features. Results In this study, UAV image at flight altitudes 100 m, 200 400 m were collected be applied a area. The resulting ground resolutions 5.29, 10.58, 21.16 cm/pixel, respectively. Four models, including Random Forest (RF), Support Vector Machine (SVM), Gradient Boosting (GBM), Deep Learning (DL), compared test performance coverage detection. 5 spectral values (Red, Green, Blue, NIR, Red edge) 16 VIs selected perform variable importance analysis on best models. results show that model each altitude has highest detecting its training (over 90%), GBM constructed based all yields covering data, an overall 95.66%. variables significantly correlated not Modified Soil Adjusted Index (MSAVI) Anthocyanin Content (MACI), Finally, was used invert complete images different altitudes. Conclusions general, GBM_all imaging feasible accurately detect coverage. prediction models had certain similarity distribution index importance. Combined method visual interpretation, green predicted by good agreement truth, other land types hay, rock, soil well predicted. This study provided methodological reference eastern China.

Language: Английский

---

Fusion Classification of HSI and MSI Using a Spatial-Spectral Vision Transformer for Wetland Biodiversity Estimation

Remote Sensing, Journal Year: 2022, Volume and Issue: 14(4), P. 850 - 850

Published: Feb. 11, 2022

The rapid development of remote sensing technology provides wealthy data for earth observation. Land-cover mapping indirectly achieves biodiversity estimation at a coarse scale. Therefore, accurate land-cover is the precondition estimation. However, environment wetlands complex, and vegetation mixed patchy, so recognition based on full challenges. This paper constructs systematic framework multisource image processing. Firstly, hyperspectral (HSI) multispectral (MSI) are fused by CNN-based method to obtain with high spatial-spectral resolution. Secondly, considering sequentiality spatial distribution spectral response, vision transformer (SSViT) designed extract sequential relationships from images. After that, an external attention module utilized feature integration, then pixel-wise prediction achieved mapping. Finally, benthos sites analyzed consistently reveal rule benthos. Experiments ZiYuan1-02D Yellow River estuary wetland conducted demonstrate effectiveness proposed compared several related methods.

Language: Английский

---

Coastal Wetland Vegetation Classification Using Pixel-Based, Object-Based and Deep Learning Methods Based on RGB-UAV

Land, Journal Year: 2022, Volume and Issue: 11(11), P. 2039 - 2039

Published: Nov. 14, 2022

The advancement of deep learning (DL) technology and Unmanned Aerial Vehicles (UAV) remote sensing has made it feasible to monitor coastal wetlands efficiently precisely. However, studies have rarely compared the performance DL with traditional machine (Pixel-Based (PB) Object-Based Image Analysis (OBIA) methods) in UAV-based wetland monitoring. We constructed a dataset based on RGB-based UAV data PB, OBIA, methods classification vegetation communities wetlands. In addition, our knowledge, OBIA method was used for first time this paper Google Earth Engine (GEE), ability GEE process confirmed. results showed that comparison PB methods, achieved most promising results, which capable reflecting realistic distribution vegetation. Furthermore, paradigm shifts from terms feature engineering, training reference explained considerable by method. suggested combination UAV, DL, cloud computing platforms can facilitate long-term, accurate monitoring at local scale.

Language: Английский

---

Comparison of Different Transfer Learning Methods for Classification of Mangrove Communities Using MCCUNet and UAV Multispectral Images

Remote Sensing, Journal Year: 2022, Volume and Issue: 14(21), P. 5533 - 5533

Published: Nov. 2, 2022

Mangrove-forest classification by using deep learning algorithms has attracted increasing attention but remains challenging. The current studies on the transfer of mangrove communities between different regions and sensors are especially still unclear. To fill research gap, this study developed a new deep-learning algorithm (encoder–decoder with mixed depth-wise convolution cascade upsampling, MCCUNet) modifying encoder decoder sections DeepLabV3+ presented three transfer-learning strategies, namely frozen (F-TL), fine-tuned (Ft-TL), sensor-and-phase (SaP-TL), to classify MCCUNet high-resolution UAV multispectral images. This combined recursive feature elimination principal component analysis (RFE–PCA), high-dimensional dataset map communities, evaluated their performance. results showed following: (1) outperformed original for classifying achieving highest overall accuracy (OA), i.e., 97.24%, in all scenarios. (2) RFE–PCA dimension reduction improved performance algorithms. OA species from was 7.27% after adding dimension-reduced texture features vegetation indices. (3) Ft-TL strategy enabled achieve better stability than F-TL strategy. improvement F1–score Spartina alterniflora 19.56%, (4) SaP-TL produced classifications images phases sensors. Aegiceras corniculatum 19.85%, (5) All strategies achieved high mean 84.37~95.25%.

Language: Английский

---