Land Use Classification of the Deep Convolutional Neural Network Method Reducing the Loss of Spatial Features

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Land use classification is crucial for remote sensing data analysis.
• Deep convolutional neural networks (DCNNs) excel at semantic segmentation but struggle with spatial feature loss.
• Existing DCNNs often compromise spatial details, impacting accuracy in high-resolution imagery.

Purpose of the Study:

• To introduce a novel network, the dense-coordconv network (DCCN), to mitigate spatial feature loss in semantic segmentation.
• To enhance object boundary definition and improve overall classification accuracy.
• To validate the DCCN's effectiveness on a standard remote sensing benchmark dataset.

Main Methods:

• Developed the dense-coordconv network (DCCN) by integrating the coordconv module into an improved DenseNet architecture.
• Incorporated coordinate information into feature maps to preserve and enhance spatial details.
• Evaluated the DCCN on the ISPRS 2D semantic labeling benchmark dataset.

Main Results:

• The DCCN achieved significant performance improvements compared to other DCNNs like U-net, SegNet, and Deeplab-V3.
• The proposed DCCN reached an overall accuracy (OA) of 89.48% and a mean F1 score of 86.89%.
• Demonstrated a notable reduction in spatial feature loss and enhanced boundary delineation.

Conclusions:

• The DCCN effectively addresses the spatial feature loss issue in DCNN-based semantic segmentation.
• The proposed method significantly improves classification accuracy for high-resolution remote sensing imagery.
• DCCN offers a promising approach for advanced land use classification tasks.