An efficient parallel DCNN algorithm in big data environment.

Area of Science:

• Geospatial data analysis
• Remote sensing technology
• Machine learning applications

Background:

• Big data is crucial for remote sensing and landslide prediction.
• Deep convolutional neural networks (DCNNs) improve prediction accuracy but face challenges with large datasets.
• Existing methods struggle with redundant features, slow computations, and poor convergence.

Purpose of the Study:

• To develop an efficient parallel DCNN algorithm (PDCNN-MI) for processing vast satellite imagery and geospatial data.
• To address challenges of redundant features, slow convolution operations, and poor loss function convergence in DCNNs.
• To enhance the accuracy and speed of landslide prediction using big data.

Main Methods:

• Implemented a parallel feature extraction strategy (PFE-MHO) using the Marr-Hildreth operator to reduce data redundancy.
• Developed a parallel model training strategy (PMT-IM) with the Im2col method to accelerate convolution operations.
• Introduced an improved small batch gradient descent strategy (IMBGD) to enhance loss function convergence.

Main Results:

• The PDCNN-MI algorithm demonstrated superior classification accuracy compared to existing methods.
• The proposed methods effectively reduced data redundancy and improved convolution operation speed.
• The IMBGD strategy successfully addressed poor convergence issues in the loss function.

Conclusions:

• PDCNN-MI is well-suited for fast and large-scale image dataset processing in remote sensing.
• The integration of MapReduce and Im2col algorithms significantly enhances DCNN performance for geospatial applications.
• This approach offers a robust solution for improving landslide prediction accuracy and efficiency.