Comparing Object-Based and Pixel-Based Machine Learning Models for Tree-Cutting Detection with PlanetScope Satellite Images: Exploring Model Generalization

International Journal of Applied Earth Observation and Geoinformation, Journal Year: 2023, Volume and Issue: 125, P. 103555 - 103555

Published: Nov. 8, 2023

Despite utilizing various remote sensing datasets, precise tree-cutting detection remains challenging due to spatial and spectral resolution constraints in satellite imagery, complex landscapes, data integration issues, the need for accurate multi-temporal reference datasets. This study investigates utilization of PlanetScope (PS) images, along with pixel-based (PBIA) object-based (OBIA) image analysis, mapping forest cover tree cuttings. Detailed datasets were collected based on airborne laser scanning (ALS)-derived canopy height models (CHM) very high-resolution (VHR) aerial orthomosaics. Reference used train three machine learning (ML) models: random (RF), support vector (SVM), feed-forward neural network (Nnet) two districts located Western Northern Poland. The also assessed generalization capabilities best model both local temporal contexts. Regarding mapping, OBIA RF classifier outperformed all other an overall accuracy (OA) 99.27 % Kappa 98.18 %, while PBIA SVM showed lowest (OA = 97.18 94.35 %). testing model's confirmed performance model, Dice Coefficient ranging from 95.86 96.74 %. methodology's effectiveness was demonstrated, rate 96.20 99.39 total number cuttings, 99.45 99.86 volume. In conclusion, PS spectral-textural features, generalized ML proves be effective detection.

Language: Английский

---

Deforestation detection using deep learning-based semantic segmentation techniques: a systematic review

Frontiers in Forests and Global Change, Journal Year: 2024, Volume and Issue: 7

Published: Feb. 2, 2024

Deforestation poses a critical global threat to Earth’s ecosystem and biodiversity, necessitating effective monitoring mitigation strategies. The integration of deep learning with remote sensing offers promising solution for precise deforestation segmentation detection. This paper provides comprehensive review methodologies applied analysis through satellite imagery. In the face deforestation’s ecological repercussions, need advanced surveillance tools becomes evident. Remote sensing, its capacity capture extensive spatial data, combined learning’s prowess in recognizing complex patterns enable assessment. Integration these technologies state-of-the-art models, including U-Net, DeepLab V3, ResNet, SegNet, FCN, has enhanced accuracy efficiency detecting patterns. underscores pivotal role imagery capturing information highlights strengths various architectures analysis. Multiscale feature fusion emerge as strategies enabling networks comprehend contextual nuances across scales. Additionally, attention mechanisms combat overfitting, while group shuffle convolutions further enhance by reducing dominant filters’ contribution. These collectively fortify robustness models techniques into applications serves an excellent tool identification monitoring. synergy between fields, exemplified reviewed presents hope preserving invaluable forests. As technology advances, insights from this will drive development more accurate, efficient, accessible detection methods, contributing sustainable management planet’s vital resources.

Language: Английский

---

A unified 2D medical image segmentation network (SegmentNet) through distance-awareness and local feature extraction

Journal of Applied Biomedicine, Journal Year: 2024, Volume and Issue: 44(3), P. 431 - 449

Published: June 13, 2024

Language: Английский

---

A Survey of Computer Vision Techniques for Forest Characterization and Carbon Monitoring Tasks

Remote Sensing, Journal Year: 2022, Volume and Issue: 14(22), P. 5861 - 5861

Published: Nov. 19, 2022

Estimation of terrestrial carbon balance is one the key tasks in understanding and prognosis climate change impacts development tools policies according to mitigation adaptation strategies. Forest ecosystems are major pools stocks affected by controversial processes influencing stability. Therefore, monitoring forest a proper inventory management resources planning their sustainable use. In this survey, we discuss which computer vision techniques applicable most important aspects actions, considering wide availability remote sensing (RS) data different resolutions based both on satellite unmanned aerial vehicle (UAV) observations. Our analysis applies occurring such as estimation areas, tree species classification, resources. Through also provide necessary technical background with description suitable sources, algorithms’ descriptions, corresponding metrics for evaluation. The implementation provided into routine workflows significant step toward systems continuous actualization data, including real-time monitoring. It crucial diverse purposes local global scales. Among improved strategies offset projects, enhancement prediction accuracy system changes under land-use scenarios.

Language: Английский

---

Learning spectral-spatial representations from VHR images for fine-scale crop type mapping: A case study of rice-crayfish field extraction in South China

ISPRS Journal of Photogrammetry and Remote Sensing, Journal Year: 2023, Volume and Issue: 199, P. 28 - 39

Published: April 1, 2023

Language: Английский

---

Mapping intertidal topographic changes in a highly turbid estuary using dense Sentinel-2 time series with deep learning

ISPRS Journal of Photogrammetry and Remote Sensing, Journal Year: 2023, Volume and Issue: 205, P. 1 - 16

Published: Oct. 1, 2023

Intertidal mudflats are an important component of the coastal geomorphological system at interface between ocean and land. Accurate up-to-date mapping intertidal topography high spatial resolution, tracking its changes over time, essential for habitat protection, sustainable management vulnerability analysis. Compared with ground-based or airborne terrain mapping, satellite-based waterline method is more cost-effective constructing large-scale topography. However, accuracy affected by extraction waterlines calibration height. The blurred boundary turbid water in tide-dominated estuary brings enormous challenges accurate extraction, errors estuarine level simulations prevent direct heights. To address these issues, this paper developed a novel deep learning using parallel self-attention mechanism boundary-focused hybrid loss to extract accurately from dense Sentinel-2 time series. UAV photogrammetric surveys were employed calibrate heights rather than simulated levels, such that error propagation constrained effectively. Annual topographic maps Yangtze China generated 2020 2022 optimized method. Experimental results demonstrate proposed could achieve excellent performance land segmentation time-varying tidal environments, better generalization capability compared benchmark U-Net, U-Net++ U-Net+++ models. comparison observations resulted RMSE 13 cm, indicating effectiveness monitoring morphological mudflats. successfully identified hotspots mudflat erosion deposition. Specifically, connected predominantly experienced deposition 10–20 cm two-year period, whereas offshore sandbars exhibited instability significant 20–60 during same period. These serve as valuable datasets providing scientific baseline information support decisions.

Language: Английский

---

Predicting the temperature field of thermal cloaks in homogeneous isotropic multilayer materials based on deep learning

International Journal of Heat and Mass Transfer, Journal Year: 2023, Volume and Issue: 219, P. 124849 - 124849

Published: Oct. 31, 2023

Language: Английский

---

An Improved U-Net Model Based on Multi-Scale Input and Attention Mechanism: Application for Recognition of Chinese Cabbage and Weed

Sustainability, Journal Year: 2023, Volume and Issue: 15(7), P. 5764 - 5764

Published: March 26, 2023

The accurate spraying of herbicides and intelligent mechanical weeding operations are the main ways to reduce use chemical pesticides in fields achieve sustainable agricultural development, an important prerequisite for achieving these is identify field crops weeds accurately quickly. To this end, a semantic segmentation model based on improved U-Net proposed paper address issue efficient identification vegetable weeds. First, simplified visual group geometry 16 (VGG16) network used as coding model, then, input images continuously naturally down-sampled using average pooling layer create feature maps various sizes, laterally integrated from into model. Then, number convolutional layers decoding cut channel attention (ECA) introduced before fusion network, so that jump connection encoding up-sampled pass through ECA module together fusion. Finally, study uses obtained Chinese cabbage weed dataset compare with original current commonly models PSPNet DeepLab V3+. results show mean intersection over union pixel accuracy increased comparison by 1.41 0.72 percentage points, respectively, 88.96% 93.05%, processing time single image 9.36 points 64.85 ms. In addition, has more effect close overlap compared other three models, which necessary condition weeding. As result, can offer strong technical support development robots robots.

Language: Английский

---

Decreased and Fragmented Greenspaces in and around Rural Residential Areas of Eastern China in the Process of Urbanization

Remote Sensing Applications Society and Environment, Journal Year: 2025, Volume and Issue: unknown, P. 101518 - 101518

Published: March 1, 2025

Language: Английский

---

HyFormer: Hybrid Transformer and CNN for Pixel-Level Multispectral Image Land Cover Classification

International Journal of Environmental Research and Public Health, Journal Year: 2023, Volume and Issue: 20(4), P. 3059 - 3059

Published: Feb. 9, 2023

To effectively solve the problems that most convolutional neural networks cannot be applied to pixelwise input in remote sensing (RS) classification and adequately represent spectral sequence information, we propose a new multispectral RS image framework called HyFormer based on Transformer. First, network combining fully connected layer (FC) (CNN) is designed, 1D sequences obtained from layers are reshaped into 3D feature matrix for of CNN, which enhances dimensionality features through FC as well increasing expressiveness, can problem 2D CNN achieve pixel-level classification. Secondly, three levels extracted combined with linearly transformed information enhance expression capability, also used transformer encoder improve using powerful global modelling capability Transformer, finally skip connection adjacent encoders fusion between different information. The pixel results by MLP Head. In this paper, mainly focus distribution eastern part Changxing County central Nanxun District, Zhejiang Province, conduct experiments Sentinel-2 images. experimental show overall accuracy study area 95.37% Transformer (ViT) 94.15%. District 95.4% 94.69%, performance dataset better than

Language: Английский

---