Research on Improved DeepLab v3+ Model for Terrace Remote Sensing Extraction

A study has proposed an improved DeepLab v3+ model for the automated extraction of terraced fields from high-resolution remote sensing imagery. Terraced fields are a critical component of agricultural production, and accurate area estimation is essential for policy-making, land planning, and resource management. Traditional remote sensing methods struggle with complex terrain and planting conditions, prompting the exploration of deep learning techniques for precise extraction.

The research utilized GaoFen-6 satellite imagery of fallow terraces to construct a semantic segmentation dataset. The improved DeepLab v3+ model replaces the Atrous Spatial Pyramid Pooling (ASPP) module with a Multi-scale Feature Fusion (MSFF) module, employing cascaded dilated convolutions with progressively increasing dilation rates to mitigate information loss. Additionally, a coordinate attention mechanism was applied to both shallow and deep features to enhance target recognition.

Results showed that combining red, green, and near-infrared bands yielded the highest accuracy in terrace extraction. Compared to the original DeepLab v3+, the improved model achieved increases of 4.62% in precision, 2.61% in recall, 3.81% in F1-score, and 2.81% in intersection-over-union (IoU). Furthermore, the refined model demonstrated superior computational efficiency, with parameters reduced to 28.6% of UNet and 19.5% of the original DeepLab v3+, while floating-point operations were only one-fifth of both models. This makes the model particularly suitable for resource-constrained environments.

The study concludes that deep learning offers high precision in identifying terraced fields from high-resolution remote sensing data, providing valuable insights for refined monitoring and management.

Key terms: Terrace extraction, remote sensing, convolutional neural network, GaoFen-6 satellite, DeepLab v3+.

Figures included in the study illustrate the location and sample distribution in Yuanyang County, dataset construction, model architecture, extraction results under different band combinations, identification outcomes, and spatial distribution across slopes and elevations.

The research was conducted by a team from Yunnan University’s School of Earth Sciences and Institute of International Rivers and Eco-Security, with findings published in *Smart Agriculture* (2024, Vol. 6, No. 3).