研究目的
To propose a novel approach based on a fully convolutional network with pyramid pooling (FCN-PP) for landslide inventory mapping (LIM) to overcome the limitations of traditional methods in detecting landslide areas due to complexity and spatial uncertainty.
研究成果
The proposed FCN-PP method effectively addresses the challenges of landslide inventory mapping by combining MMR for noise reduction and a deep convolutional network with pyramid pooling for improved feature representation. It outperforms state-of-the-art methods in terms of Precision, Recall, Overall Error, F-score, and Accuracy, demonstrating its superiority for automatic and accurate LIM without extensive parameter tuning.
研究不足
The approach relies on a small dataset of bitemporal images, which may limit generalizability to other regions or landslide types. The computational complexity of deep learning models requires significant hardware resources, and the method's performance may be affected by image quality and environmental factors.
1：Experimental Design and Method Selection:
The methodology involves image preprocessing using multivariate morphological reconstruction (MMR) to filter noise, followed by a deep convolutional neural network (FCN-PP) for semantic segmentation of landslide areas from bitemporal images. The FCN-PP incorporates pyramid pooling to handle multiscale features and improve localization accuracy.
2：Sample Selection and Data Sources:
Five pairs of bitemporal high-resolution remote sensing images from Hong Kong, captured by the Zeiss RMK TOP 15 Aerial Survey Camera System in December 2007 and November 2014, were used. Images were cropped into areas A-E, with sizes ranging from 750x950 to 1252x2199 pixels. Training data consisted of 139 overlapping images augmented to 1112 pairs through rotations, flips, shears, and scaling; testing data included 3 non-overlapping image pairs of size 473x
3：List of Experimental Equipment and Materials:
4 Equipment includes a workstation with Intel Xeon CPU E5-1620v4,
4：5 GHz, four cores, 64-GB RAM, and double NVIDIA GTX 1080 GPU. Software used:
MATLAB 2017b for unsupervised methods and PyTorch for deep learning implementations. The structuring element for MMR is a disk of size 1x
5：Experimental Procedures and Operational Workflow:
Preprocessing: Apply MMR to difference images of bitemporal images to remove noise. Network Training: Initialize FCN-PP with parameters from VGG-16's first four convolutional layers, use stochastic gradient descent with learning rate 1e-4, weight decay
6：0005, momentum 99, minibatch size 4, and 30 epochs. Testing:
Evaluate on cropped test images using the trained network.
7：Data Analysis Methods:
Quantitative evaluation using five metrics: Precision, Recall, Overall Error, F-score, and Accuracy, computed by comparing detected landslide areas with ground truth data.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
