Dual-Encoder Network Improves Mountainous Terrace Field Mapping Precision in Remote Sensing

Terraced agricultural fields play a vital role in mitigating surface runoff and intercepting sediment across upland landscapes, making large-scale and high-precision spatial identification essential for environmental monitoring and soil erosion control. In a study published in Frontiers in Ecology and Evolution, researchers introduced DETS-Net, a dual-encoder deep learning framework engineered specifically to address longstanding obstacles in automated terrace extraction from high-resolution remote sensing imagery. Mountainous terrace identification routinely faces severe technical difficulties, including extreme scale variability among individual farm plots, pervasive shadow interference caused by rugged terrain, and highly irregular geometric configurations that confound standard segmentation models.

To overcome these environmental and visual complexities, the research team designed DETS-Net by coupling convolutional neural network capabilities with vision transformer mechanisms. The network employs ResNet-34 as the CNN encoder for local feature detail extraction. Working in parallel, a Transformer encoder models long-range spatial dependencies across the broader landscape, capturing contextual relationships across wider terrain formations. The dual representations generated by these branches are dynamically integrated using an adaptive feature fusion module, which balances local detail against wide spatial context during semantic segmentation.

Benchmarking the architecture required a specialized geographical dataset capable of reflecting realistic upland topography. The researchers constructed the THDIR dataset, assembling 9,336 standardized images measuring 256 by 256 pixels.

Experimental evaluations on the THDIR dataset demonstrated that DETS-Net established superior segmentation precision when systematically compared against eight representative deep learning models. In experimental comparisons, DETS-Net outperformed current mainstream methods in segmentation precision.

Following validation on the benchmark dataset, the investigators deployed the optimized model onto 2023 remote sensing imagery covering Luotian County, situated in the Dabie Mountains of Hubei Province, China. The practical deployment produced a high-precision, county-level spatial distribution map documenting the extent and configuration of terraced land across the region.

The framework directly supports soil and water conservation assessments by generating precise boundary delineations, overall surface area statistics, and slope distribution data. The framework provides a reproducible technical pathway for automated terrace mapping using high-resolution remote sensing imagery in complex mountainous regions.

Subject of Research: Automated remote sensing extraction of terraced fields in complex mountainous terrain using deep learning

Article Title: A comprehensive benchmark and dual-encoder baseline for terraced field extraction in complex mountainous regions

Article References: Tian, P., Sun, Y., Chen, Y., Cai, Y., Deng, R., Zhang, Z., Ma, B., & Xiao, R. (2026). A comprehensive benchmark and dual-encoder baseline for terraced field extraction in complex mountainous regions. Frontiers in Ecology and Evolution, 14, Article 2011654. https://doi.org/10.3389/fevo.2026.2011654

Image Credits: AI Generated

DOI: 10.3389/fevo.2026.2011654

Keywords: DETS-Net, remote sensing, semantic segmentation, terraced fields, soil erosion control, ResNet-34, Transformer encoder, adaptive feature fusion, THDIR dataset, Luotian County, Dabie Mountains, soil and water conservation, deep learning, mountainous terrain, plateau mountains, hills, mountain-basin transition zones