论文信息

Pei Tian / Yujie Sun / Yan Chen 等 8 人

Frontiers in Ecology and Evolution

论文解读

AI 辅助整理
方法
CNN与Transformer双编码器加自适应特征融合,构建9336张影像的TH
发现
DETS-Net精度优于8种主流模型,OA达89.91%,并生成县级梯田分布图。
机会点
可延伸至梯田撂荒与退化动态监测,或迁移到其他复杂地形农业设施遥感提取。

论文摘要

梯田在削减径流和拦截泥沙方面发挥着至关重要的作用。从遥感影像中大规模、高精度地提取梯田对于环境监测和水土流失控制具有重要意义。然而,梯田的精确空间识别仍面临诸多挑战,包括尺度变化、阴影干扰和复杂几何构型等。为解决这些问题,我们提出了DETS-Net,一种融合CNN和Transformer架构的双编码器网络。该网络采用ResNet-34作为CNN编码器以提取局部特征细节,利用Transformer编码器增强长程依赖关系建模,并引入自适应特征融合(AFF)模块实现两个特征分支的动态整合。我们构建了THDIR数据集,涵盖四种典型地形类型,即高原山地、丘陵和山盆过渡带,包含9,336幅标准化的256×256像素图像。我们与八种代表性深度学习模型进行了系统比较。实验结果表明,DETS-Net在THDIR数据集上取得了89.91%的总体精度(OA)、81.28%的平均交并比(MIoU)和87.99%的F1分数,其分割精度优于当前主流方法。我们将优化后的模型应用于湖北省大别山区罗田县2023年遥感影像,生成了高精度的县级梯田空间分布图。该框架通过面积统计、坡度分布数据和高精度边界提取,为水土保持效益评估提供支持,并为复杂山区高分辨率遥感影像的梯田自动化制图提供了可复现的技术路径。

全文

Frontiers in Ecology and Evolution

A comprehensive benchmark
and dual-encoder baseline
for terraced field extraction in
complex mountainous regions
Pei Tian1*, Yujie Sun1, Yan Chen1, Yuanzhe Cai2, Ruoxuan Deng1,
Zheng Zhang1, Biao Ma3,4 and Ruihan Xiao1
1College of Urban and Environmental Sciences, Central China Normal University, Wuhan, China,
2College of Big Data and Internet, Shenzhen Technology University, Shenzhen, China, 3Changjiang
Survey, Planning, Design and Research Co., Ltd., Wuhan, China, 4Hubei Key Laboratory of Basin Water
Security, Wuhan, China
Terraces play a vital role in mitigating runoff and intercepting sediment. Large-
scale, high-precision extraction of terraces from remote sensing imagery is
essential for environmental monitoring and soil erosion control. However, accu-
rate spatial identification of terraces still presents signi ficant challenges, including
scale variability, shadow interference, and complex geometric con figurations. To
address these issues, we propose DETS -Net, a dual-encoder network that
integrates CNN and Transformer architectures. The network employs ResNet-
34 as the CNN encoder for local feature detail extraction, utilizes a Transformer
encoder to strengthen long-range dependency modeling, and incorporates an
adaptive feature fusion (AFF) module for the dynamic integration of the two
feature branches. We develop the THDIR dataset, which covers four typical terrain
types, namely, plateau mountains, hills, and mountain-basin transition zones, and
contains 9,336 standardized 256 × 256 pixel images. We conduct systematic
comparisons with eight representative deep learning models. Experimental re-
sults demonstrate that DETS-Net achieves an overall accuracy (OA) of 89.91%, a
mean intersection over union (MIoU) of 81.28%, and an F1-score of 87.99% on the
THDIR dataset, outperforming current mainstream methods in segmentation
precision. We apply the optimized model to 2023 remote sensing imagery of
Luotian County in the Dabie Mountains, Hubei Province, and generate a high-
precision county-level terrace spatial distribution map. This framework supports
the assessment of soil and water conservation bene fits through area statistics,
slope distribution data, and high-precision boundary extraction and provides a
reproducible technical pathway for automated terrace mapping using high-
resolution remote sensing imagery in complex mountainous regions.
KEYWORDS
deep learning, remote sensing, semantic segmentation, soil erosion, terrace dataset,
terraced field identi fication
Frontiers in Ecology and Evolution frontiersin.org01
OPEN ACCESS
EDITED BY
Ren Geng,
Nanjing Hydraulic Research Institute,
China
REVIEWED BY
Hai Xiao,
China Three Gorges University, China
Wenzhao Guo,
Northwest A&F University, China
*CORRESPONDENCE
Pei Tian
tianpei@mail.ccnu.edu.cn
RECEIVED 27 September 2026
REVISED 15 October 2026
ACCEPTED 19 October 2026
PUBLISHED 13 November 2026
CITATION
Tian P, Sun Y, Chen Y, Cai Y, Deng R,
Zhang Z, Ma B and Xiao R (2026)
A comprehensive benchmark and
dual-encoder baseline for terraced
field extraction in complex
mountainous regions.
Front. Ecol. Evol. 14:2011654.
doi: 10.3389/fevo.2026.2011654
COPYRIGHT
©2 0 2 6T i a n ,S u n ,C h e n ,C a i ,D e n g ,
Zhang, Ma and Xiao. This is an open-
access article distributed under the terms
of the Creative Commons Attribution
License (CC BY). The use, distribution or
reproduction in other forums is
permitted, provided the original
author(s) and the copyright owner(s) are
credited and that the original publication
in this journal is cited, in accordance
with accepted academic practice. No
use, distribution or reproduction is
permitted which does not comply with
these terms.
TYPE Original Research
PUBLISHED 13 November 2026
DOI 10.3389/fevo.2026.2011654

1 Introduction
For millennia, terraces have served as a crucial agricultural
engineering measure for adapting slopes to human needs, playing a
pivotal role in mitigating soil erosion, expanding arable land areas,
ensuring food security, and preserving ecosystem services ( Chen
et al., 2017a; Qamar et al., 2025 ; Wei et al., 2016 ). Globally, terraces
are widely distributed across tropical, subtropical, arid, and semi-
arid mountain regions, formin gd i v e r s el a n d s c a p ep a t t e r n s
(Ackermann et al., 2019 ). Studies demonstrate that terraced fields
significantly reduce runoff and soil erosion —with their anti-erosion
efficacy exceeding that of conventional non-terraced systems by
more than 11-fold ( Wei et al., 2016 )—while enhancing soil organic
carbon storage, promoting water recharge, and facilitating nutrient
accumulation ( Cheng et al., 2025 ; Shi and Wei, 2025 ). Under
extreme rainfall conditions, the synergy between terraces and
vegetation effectively attenuates flood peak flows ( Fu et al., 2020 )
and reduces landslide risks ( Sibomana et al., 2025 ; Yang et al.,
2025b). Additionally, terraced agriculture contributes substantial
cultural ecosystem services, including aesthetic landscape value and
heritage signi ficance ( Chen et al., 2025b ). However, accelerated
urbanization and the loss of agricultural labor have led to wide-
spread abandonment of terraces worldwide, resulting in embank-
ment collapse, soil d egradation, and declining biodiversity
(Cammeraat et al., 2024 ; Long et al., 2025 ; Song et al., 2026 ).
Therefore, accurate mapping of terraces ’ spatial distribution and
dynamic changes is essential for scienti fic soil and water conserva-
tion planning, ecological bene fit assessment, and sustainable man-
agement strategy development (Amodio et al., 2025; Fan et al., 2026;
Lu et al., 2025a ).
The advancement of remote sensing technology has provided
an effective tool for large-scale monitoring of terraced fields (Patric
et al., 2026 ; Zhang et al., 2017 ). In recent years, semantic segmen-
tation methods leveraging high-resolution remote sensing imagery
and deep learning have demonstrated signi ficant potential in
terraced field detection ( Li et al., 2025 ; Lv et al., 2023 ; Zhao et al.,
2026b). However, existing research faces multiple challenges: ter-
raced fields exhibit complex edge textures and multi-scale features
in remote sensing images, making them easily confused with natural
topography or other landforms; moreover, existing terraced field
datasets often suffer from regional limitations, insuf ficient annota-
tion precision, or a lack of diverse geographic coverage, hindering
the development of high-precision, large-scale semantic segmenta-
tion models ( Liu et al., 2024 ). Some approaches rely on manually
drawn terraced field boundaries, resulting in limited generalization
capabilities ( Zhang et al., 2017 ), while deep learning-based studies
predominantly employ generic semantic segmentation architectures
(e.g., U-Net, DeepLabv3+) that lack specificity for the unique spatial
structure of terraced fields ( Li et al., 2025 ).
The rapid advancement of deep learning technologies has
revolutionized the intelligent interpretation of remote sensing
imagery ( Hankui K Zhang et al., 2024 ; Kattenborn et al., 2021 ).
Deep learning models, particularly convolutional neural networks
(CNNs), can automatically extract semantic features across multiple
layers, signi ficantly outperforming traditional manual extraction
methods in efficiency (LeCun et al., 2015 ). The introduction of fully
convolutional networks (FCNs) enabled dense prediction at the
pixel level ( Long et al., 2015 ); U-Net preserves spatial details
through skip connections ( Ronneberger et al., 2015 ), while the
DeepLab series employs hollow convolutions to expand the recep-
tive field for capturing multi-scale contextual information and
refines object edge regions to improve segmentation accuracy
(Chen et al., 2017b , 2018; Sun et al., 2025 ). These architectures
have been widely adopted in high-resolution remote sensing feature
extraction tasks ( Ma et al., 2019 ; Zhu et al., 2017 ). However, the
inherent local receptive fields of CNNs limit their ability to model
long-range dependencies and global context ( Luo et al., 2017 ). The
emergence of Vision Transformers (ViTs) offers a novel solution to
this challenge ( Vaswani et al., 2017 ), as they capture global
dependencies by calculating correlation weights between all posi-
tions in sequences, demonstrating strong contextual modeling
capabilities in remote sensing segmentation ( Dosovitskiy et al.,
2021; Wang et al., 2022 ). The Swin Transformer reduces compu-
tational complexity while maintaining global modeling through a
shifting window mechanism (Liu et al., 2021), whereas SETR applies
a pure Transformer architecture speci fically to semantic segmenta-
tion ( Zheng et al., 2021 ). Building on this foundation, hybrid
architectures combining CNNs and Transformers have become a
mainstream approach in remote sensing segmentation —where the
former extract local details and boundary textures, while the latter
capture global context and long-range dependencies ( Kattenborn
et al., 2021 ; Zhang et al., 2020 ). Representative works such as
TransUNet, MCANet, and CTSeg ( Chen et al., 2021 ; Chen et al.,
2025a; Yan et al., 2026 ) have all demonstrated the effectiveness of
dual-branch collaborative modeling. Furthermore, the integration of
multi-scale feature fusion techniques ( Hu et al., 2018 ; Lin et al.,
2017) and attention mechanisms (including channel attention
(Hu et al., 2018 ), spatial attention ( Woo et al., 2018 ), and self-
attention ( Vaswani et al., 2017 )) signi ficantly enhances models ’
adaptability to scale variations and complex backgrounds. In agri-
cultural remote sensing, researchers have developed multiple farm-
land datasets and proposed specialized networks for precise cultivated
land boundary extraction ( Fang et al., 2026 ; Xu et al., 2024 ; Zhao
et al., 2026a , 2025). However, dedicated datasets and segmentation
architectures tailored to speci fic objects such as terraced fields—a
distinct type of arti ficial terrain—remain scarce in existing research.
The precise spatial information of terraced fields holds critical
value across multiple application scenarios. In soil erosion assess-
ment, the spatial distribution of terraces, one of the most important
water conservation engineering measures, directly determines the
accuracy of soil retention calculations (Saiz et al., 2016; Yaseen et al.,
2025). Stability monitoring of terraces serves as a key early warning
system for preventing embankment collapses and landslide disas-
ters ( Ajith et al., 2024 ; Sun et al., 2026 ). For accurate agricultural
subsidy distribution and delineation of cultivated land protection
redlines, high-precision boundary mapping of terraced fields pro-
vides fundamental data for determining arable land areas ( Sui et al.,
2026; Zhang et al., 2024). Furthermore, as an integral component of
globally important agricultural heritage sites, spatial mapping of
terraces is indispensable for both cultural preservation and tourism
planning (Cao et al., 2024 ; Yang et al., 2025a ). However, traditional
information acquisition methods, including manual field surveys,
Tian et al. 10.3389/fevo.2026.2011654
Frontiers in Ecology and Evolution frontiersin.org02

digital terrain mapping, and low-to-medium resolution remote
sensing classi fication, exhibit systemic limitations in coverage
scope, update frequency, and spatial accuracy ( Pijl et al., 2021 ;
Wang et al., 2026), failing to meet practical demands for large-scale,
high-precision, and rapid updates. The integration of high-resolu-
tion remote sensing imagery with deep learning-based semantic
segmentation offers a pivotal technical approach for automated and
refined extraction of terraced field information ( Attri et al., 2023 ;
Wang et al., 2026 ).
The evolution of remote sensing methods for terraced field
identification can be divided into three stages. In the early stage,
indirect identi fication approaches based on DEM topographic
attributes (such as slope and curvature) ( Cucchiaro et al., 2021 ;
Sofia et al., 2014) were predominantly used; however, these methods
struggled to distinguish between arti ficial terraces and natural
steppe landscapes, exhibiting high misclassi fication rates.
Additionally, pixel-level classi fication using medium-to-low-reso-
lution remote sensing imagery (e.g., Landsat, MODIS) faced chal-
lenges due to mixed-pixel issues, failing to accurately delineate
terrace boundaries (T. Liu et al., 2025 ). During the mid-stage,
object-based image analysis (OBIA) methods were introduced for
high-resolution terraced field recognition (Lu et al., 2025b; Xu et al.,
2022), leveraging geometric, textural, and contextual information
for segmentation and classification, and achieving promising results
in localized areas. However, these methods heavily relied on
manually designed features, demonstrated poor transferability,
and exhibited limited generalization capabilities ( Blaschke, 2010 ).
Currently, deep learning approaches have been widely applied in
terraced field identification (Li et al., 2025 ; Zhang et al., 2017 ; Zhao
et al., 2026b ). Although these studies have demonstrated the
feasibility of deep learning in terraced field extraction, several
critical shortcomings remain: first, there is a lack of network
architectures speci fically designed for capturing the geometric
features of terraced field steps, with existing methods predomi-
nantly relying on direct applications of general segmentation
models; second, segmentation accuracy at terraced field boundaries
is inadequate, characterized by prominent issues such as blurred
boundaries and broken lines ( Lu et al., 2024 ); third, robustness to
complex scenarios, including shadow occlusion, vegetation cover-
age, and fragmented field plots, remains insufficient (Li et al., 2026);
fourth, existing terraced field datasets cover only a single type of
terrain ( Zhao et al., 2026b ), lacking benchmark datasets across
diverse terrains, which hinders systematic evaluation of models ’
generalization capabilities.
In summary, current research on remote sensing semantic seg-
mentation for terracedfields faces the following key gaps: (1) there is a
lack of dedicated segmentation network architectures tailored to
terracedfield geometric features; existing methods struggle to simulta-
neously capture both local details and global context, resulting in
insufficient segmentation integrity of terracedfield boundaries; (2) the
absence of high-resolution semantic segmentation dat asets covering
diverse terrain types hinders mode l training and the evaluation of
cross-regional generalization; (3)coordination between global context
modeling and local detail extraction remains inadequate; and (4) the
large-scale applicability of models in unknown complex terrains has
not been adequately validated.
To address these gaps, we construct the high-resolution terraced
field dataset THDIR, which covers three typical terrain types:
plateau mountains, hills, and mountain-basin transition zones. It
includes 9,336 standardized samples of 256 × 256 pixels, providing
ample complex-scene samples for model training. A dual-encoder
segmentation network, DETS-Net, is proposed, integrating CNN
and Transformer architectures: ResNet-34 captures local details of
terraced fields, while the Transformer models long-range depen-
dencies across the entire landscape, with dynamic feature integra-
tion achieved through an adaptive feature fusion (AFF) module.
The optimal model was applied to high-resolution remote sensing
images of Luotian County in the Dabie Mountain region of China ’s
Hubei Province, demonstrating its large-scale extraction capability
and fine spatial representation in unknown complex terrains.
Experimental results show that DETS-Net achieves an overall
accuracy (OA) of 89.91%, a mean intersection over union (MIoU)
of 81.28%, and an F1-score of 87.99% on the THDIR dataset,
outperforming eight mainstream benchmark models. The extrac-
tion results for Luotian County reveal a spatial pattern of terraced
field distribution characterized by “high density in the south, low
density in the north, and higher density in the west than in the east,”
with a validation accuracy of 90.00% achieved through
random sampling.
The main contributions of this paper are threefold. First, the
dual-encoder architecture with adaptive fusion improves terraced
field boundary segmentation accuracy under complex conditions
such as shadow occlusion and field fragmentation, offering a
transferable methodological reference for automated mapping in
challenging geographical environments. Second, we publish the
THDIR high-resolution semantic segmentation dataset covering
diverse terrain types, filling the data gap for high-precision terraced
field mapping in complex mountainous areas. Third, we produce a
detailed county-level terraced field distribution map for Luotian
County, Hubei, China, revealing the slope-elevation differentiation
patterns of terrace distribution and providing critical data for
regional soil and water conservation bene fit evaluation and dy-
namic monitoring of terrace resources.
2 Materials and methods
2.1 Study area
Luotian County is located in the northeastern part of Hubei
Province. The terrain consists mainly of medium-to-low moun-
tains, followed by hills; the northeastern region contains concen-
trated ranges of high mountains, while the central area represents a
transitional zone of low mountains and hills at elevations between
200 and 500 meter, and the southern part comprises shallow hills
and alluvial plains below 200 meter. The country thus exhibits
highly complex terrain and diverse landforms. Agricultural pro-
duction remains one of the key industries in the region. Figure 1
illustrates the elevation and geographical layout of Luotian County.
Luotian County possesses extensive, diverse, and historically
significant terraces, making it a representative example of terrace
Tian et al. 10.3389/fevo.2026.2011654
Frontiers in Ecology and Evolution frontiersin.org03

landscapes in the mountainous regions of eastern Hubei. As a key
area for soil and water conservation in the Dabie Mountains, the
county faces severe soil erosion challenges. High-precision analysis
of the spatial distribution of terraces and evaluation of their soil and
water conservation bene fits hold signifi cant scienti fic value and
regional relevance; therefore, we select Luotian County as the
study area.
2.2 Dataset
2.2.1 Dataset construction
To address the challenges posed by existing terrace datasets, we
develop the Terrace High-Resolution Image Dataset (THDIR),
which is speci fically designed to enhance the performance of
remote sensing image segmentation models in the complex task
of large-scale terrace extraction. The dataset comprises 9,336
standardized images with dimensions of 256 × 256 pixels, covering
terrace distribution areas across three major terrain types: plateau
mountains, hills, and mountain-basin transition zones in China,
including the Loess Plateau, the Yunnan-Guizhou Plateau, the
central and southeastern hills of China, and the mountain-basin
transition zone around the Sichuan Basin. Each image has been
carefully annotated manually to ensure precise delineation of
terrace boundaries and features, as illustrated in Figure 2.
2.2.2 Image acquisition range and parameters
China is renowned for its vast geographical diversity and serves
as an ideal site for research on terrace resource development.
Therefore, we selected high-resolution satellite remote sensing
imagery from the Google Earth platform and downloaded it using
BIGEMAP software. The image acquisition area covers terrace
distribution regions across thr ee typical topographic zones in
China: plateau mountains, hills, and mountain-basin transition
zones, including the Loess Plateau, the Yunnan-Guizhou Plateau,
central and southeastern hilly regions, and the mountain-basin
transition zones around the Sichuan Basin. The complex interac-
tions of geographical factors such as terrain and climate result in
significant variations in the surface coverage and morphological
characteristics of terraces across different regions of China, provid-
ing diverse conditions for enhancing the generalization capability
and adaptability of automatic terrace extraction models.
Based on this, we selected core areas within the aforementioned
three major terrain categories to acquire high-resolution satellite
imagery, each with a spatial resolution of 0.5 m and covering a total
area of approximately 78 km 2. This approach not only ensures the
uniformity and standardization of the image data but also provides
comprehensive and representative data for training, validating, and
testing the DETS-Net model through diverse terrace scenarios,
effectively overcoming the limitation of restricted terrain coverage
in existing datasets.
2.2.3 Sample labeling and processing
We provide a comprehensive overview of the annotation and
preprocessing work flow implemented for the THDIR dataset. By
employing visual interpretation for sample labeling and standard-
ized preprocessing of sample data, we ensure that the dataset quality
meets the requirements of model training while effectively prevent-
ing over fitting and enhancing sample diversity. Using the ArcGIS
Pro 3.2 software platform, precise annotations were performed on
remote sensing images through manual visual interpretation: pixel
values in terrace areas were marked as 1, corresponding to an RGB
fill of (255, 255, 255) (white), whereas those in non-terrace areas
were marked as 0, corresponding to an RGB fill of (0, 0, 0) (black).
Upon completion, all annotated samples were uniformly cropped to
256 × 256 pixels to standardize dimensions and satisfy subsequent
model training requirements. During data screening, images with
significant annotation defects or excessively small terrace coverage
areas were excluded to maintain dataset quality and a balanced
distribution between positive and negative samples. To prevent data
leakage, the original samples were divided into training, validation,
and test sets in an 8:1:1 ratio, with data augmentation applied only
to the training and test sets. To mitigate over fitting and increase
diversity, high-quality sub-images were augmented using five tech-
niques, namely, rotations of 90, 180, and 270 degrees, in addition to
horizontal and vertical flipping, resulting in a final dataset of 9,336
samples, with the augmentation effects demonstrated in Figure 3 .
FIGURE 1
Schematic diagram of elevation and geographical location of Luotian County.
Tian et al. 10.3389/fevo.2026.2011654
Frontiers in Ecology and Evolution frontiersin.org04

2.2.4 Statistics and analysis
Statistical analyses of sample counts and proportions were
conducted for the four representative terrace distribution regions,
namely, the Loess Plateau, the Yunnan-Guizhou Plateau, the
southeastern hills, and the eastern mountain-basin transition zone
of the Sichuan Basin, to characterize their respective distribution
patterns, as shown in Figure 4. The sample proportions are 31% for
the Loess Plateau, 12% for the Yunnan-Guizhou Plateau, 26% for
the southeastern hills, and 31% for the eastern mountain-basin
transition zone of the Sichuan Basin. This sample allocation
supports the DETS-Net model in performing terrace extraction
tasks across diverse regions and terrains, thereby enhancing the
model’s generalization performance.
2.3 Method
2.3.1 DETS-Net framework and training process
In this work, the network input consists exclusively of the three
visible RGB bands from the high-resolution imagery, without incor-
porating ancillary topographic data (e.g., DEM or slope), to evaluate
the model’s capability for visual feature extraction alone. The network
architecture of DETS-Net is illustrated in Figure 5, which shows a
parallel dual-encoder structure that extracts both local details and
global semantic features. The left CNN encoder features a four-layer
convolutional pathway (Conv Layer 1 –Conv Layer 4) that captures
low-level spatial details such as terraced field edges and textures. The
right Transformer encoder adopts a hierarchical design: itfirst divides
t h ei m a g ei n t o2×2p a t c h e s ,flattens and linearly projects them, and
then feeds them into the encoding layers. Each layer contains three
Transformer blocks, each comprising a multi-scale multi-head self-
attention (M2SA) module and an efficient discriminative frequency-
domain feed-forward network (EDFFN) module. These blocks cor-
respond one-to-one with the feature dimensions of the CNN
branches, and an adaptive feature fusion (AFF) module is incorpo-
rated at each layer to dynamically combine feature representations
from both branches at different scales, balancing fine-grained details
with global semantics. The DETS-Net decoder employs a deep-to-
shallow hierarchical upsampling strategy, using the fused features of
the final layer as the initial input for three decoding stages. At each
level, images are first upsampled using sub-pixel convolution (Pixel
Shuffle)—which outperforms traditional deconvolution in detail
preservation—and are then fused with corresponding-scale feature
channels. Local features are enhanced via inverse depth-separable
convolution (IDSC), followed by cross-scale aggregation through
Transformer blocks. Finally, the image is resampled to the original
dimensions, and a 1 × 1 convolution outputs a pixel-wise segmen-
tation mask, enabling precise terraced field recognition.
2.3.2 Accuracy assessment
The evaluation metrics include overall accuracy (OA) (Equation
1), precision (Equation 2), recall rate ( Equation 3 ), F1-score
(Equation 4 ), and mean intersection over union (MIoU)
(Equation 5). All metrics are calculated as follows:
FIGURE 2
Schematic diagram of the geographical distribution of samples in the THDIR dataset.
FIGURE 3
Data enhancement effects. The first row shows the original images, and the second row shows their corresponding labels. (a–e) High-quality sub-
images augmented by five transformation operations: 90°, 180°, 270° rotation, horizontal flip and vertical flip for sample expansion.
Tian et al. 10.3389/fevo.2026.2011654
Frontiers in Ecology and Evolution frontiersin.org05

OA = TP + TNTP + TN + FN + FP (1)
Precision = TPTP + FP (2)
Recall = TPTP + FN (3)
F1 = 2 /C2 Precision /C2 RecallPrecision + Recall (4)
MIoU = 1k+1 oki=0TPiTPi +F Pi + FNi (5)
Here, k represents the total number of categories, and i denotes
the i-th category. TP i indicates the number of pixels belonging to
the i-th category that are correctly predicted; FP i refers to the
number of pixels belonging to other categories that are incorrectly
predicted as the i-th category; and FN i denotes the number of pixels
belonging to the i-th category that are incorrectly classi fied as
other categories.
In addition, a confusion matrix was constructed to quantita-
tively evaluate the reliability and accuracy of the terraced field
semantic segmentation model, providing a comprehensive assess-
ment of its performance, as shown in Table 1.
Furthermore, to evaluate the computational ef ficiency and
lightweight performance of the model, we adopt two metrics:
parameter count (Params, in millions [M]) and floating-point
operations (FLOPs, in giga flops [G]). Parameter count represents
FIGURE 4
Sample distribution and proportions in the THDIR dataset.
FIGURE 5
Architecture of the dual encoder terrace segmentation network (DETS-Net). Gray blocks represent the ResNet34 branch, while blue-gray blocks
represent the Transformer branch. C denotes channel dimension concatenation.
Tian et al. 10.3389/fevo.2026.2011654
Frontiers in Ecology and Evolution frontiersin.org06

the total number of learnable parameters and indicates the memory
requirements of the model; FLOPs denote the total number of
floating-point operations required for a single forward pass and
measure the computational complexity of the model. Together,
these metrics assess the deployment potential and operational
efficiency of the model in practical applications.
2.3.3 Model training details
For model training, this study employs a server platform
equipped with a 16-core Intel Xeon(R) Platinum 8481C processor,
80 GB of RAM, and an RTX 4090D graphics card with 24 GB of
dedicated video memory. The software environment is con figured
using Anaconda 3 (64-bit) development suite on a Windows 11
Professional 64-bit operating system, with Python 3.8 as the
programming language and PyTorch 1.11.0 as the deep learning
framework. During training, the batch size is set to 32, and the
maximum number of epochs is set to 50. The loss function is binary
cross-entropy, the optimizer is Adam, and the learning rate is
dynamically adjusted using a multi-stage policy. Speci fically, the
Poly policy is employed for dynamic learning rate adjustment,
which effectively accelerates model convergence and improves
final performance; its formula is as follows ( Equation 6):
lr = base_lr /C2 1 − epochnum_epoch/C18/C19 power (6)
Here, lr denotes the adjusted learning rate, base_lr represents
the base learning rate, epoch is the current iteration number,
num_epoch is the maximum iteration number, and power controls
the shape of the decay curve. In this experiment, the base learning
rate is initially set to 0.001, and the curve adjustment coef ficient is
set to 0.9. All models in each experimental group are trained until
convergence. For different experimental groups, the parameters are
optimized and adjusted according to the available computational
resources and training conditions.
3 Results and analysis
3.1 Experimental evaluation
3.1.1 Accuracy assessment
To comprehensively evaluate the performance of the proposed
model, DETS-Net, this study conducts standardized benchmark
tests on the THDIR dataset against eight representative state-of-the-
art semantic segmentation algorithms: DBBANet ( Li, 2025 ),
DeepLabv3+ ( Chen et al., 2018 ), REAUNet (Lu et al., 2024 ),
Buildformer ( Wang et al., 2022 ), PSPNet ( Zhao et al., 2017 ),
SegNet ( Badrinarayanan et al., 2016 ), LANet ( Ding et al., 2021 ),
and ABCNet ( Li et al., 2021 ).
The quantitative assessment results are shown in Figure 6.
Experimental results demonst rate that DETS-Net achieves
stong performance across five metrics: 89.91% OA, 81.28% MIoU,
87.99% F1-score, 86.51% precision, and 89.52% recall. The high
recall is particularly noteworthy for terrace mapping, as it indicates
the successful identi fication of the vast majority of terrace pixels,
minimizing omission errors that could lead to systematic underes-
timation of terrace extent. The strong precision score con firms that
false-positive detections remain well controlled, which is equally
important for ensuring the reliability of area statistics derived from
the extraction results.
Compared with DBBANet, DETS-Net improves overall accu-
racy by 1.32%, the F1-score by 1.41%, and MIoU by 2.19%. The
disproportionate MIoU gain relative to OA and F1-score improve-
ments suggests that DETS-Net ’s performance advantage is most
pronounced in challenging boundary regions where precise pixel-
level classi fication is dif ficult and MIoU is particularly sensitive to
errors. Relative to ABCNet, the best-performing competitor, DETS-
Net improves accuracy by 0.85%, the F1-score by 0.98%, and MIoU
by 1.41%, achieving consistent gains without compromising recog-
nition balance. The margin over ABCNet, though modest in
absolute terms, is meaningful given that ABCNet itself is a strong
TABLE 1 Confusion matrix.
Terraced fields Background
Terraced fields TP (true positive) FN (false negative)
Background FP (false positive) TN (true negative)
Predicted value
Ground truth
FIGURE 6
Accuracy evaluation of terraced field recognition results for the THDIR dataset.
Tian et al. 10.3389/fevo.2026.2011654
Frontiers in Ecology and Evolution frontiersin.org07

baseline with an attention-based boundary re finement mechanism
specifically designed for capturing fine structural details.
Against classical general-purpose segmentation models, DETS-
Net outperforms DeepLabv3+ by 0.99% in recall, 2.18% in F1-score,
and 3.21% in MIoU. This substantial MIoU gap underscores the
limitations of applying general-purpose architectures directly to
terrace segmentation without domain-speci fic adaptation.
Compared with SegNet and PSPNet, DETS-Net improves MIoU
by 4.10% and 3.93%, respectively, with corresponding F1-score
gains of 2.90% and 2.70%. These large performance differentials
against older encoder-decoder architectures highlight the rapid
evolution of segmentation methodology and the tangible bene fits
that modern design choices —dual-encoder fusion and attention
mechanisms —bring to fine-grained mapping. Compared with
REAUNet, DETS-Net achieves an F1-score that us 5.25% higher
and MIoU that is 7.82% higher. Against LANet, it shows a 1.59%
F1-score improvement and a 2.17% MIoU increase. The contrast
with LANet is instructive: while LANet prioritizes computational
efficiency via a lightweight design, its capacity constraints lead to a
clear accuracy penalty, reinforcing the value of DETS-Net ’s ba-
lanced design. Overall, DETS-Net exhibits superior scene adapta-
bility and strong overall performance in terrace segmentation, with
advantages observed consistently across lightweight models, atten-
tion-enhanced architectures, and classical encoder-decoders.
To visually demonstrate model superiority, we compare its
terrace segmentation results with those of other state-of-the-art
methods and evaluate generalization across various terrain scenar-
ios. Figure 7 illustrates extraction results for three typical terrain
types: hills, plateaus, and mountain-basin transitions. These terrain
types represent the dominant geomorphic settings in the study area
and test the model under conditions varying substantially in slope
gradient, terrain ruggedness, and terrace spatial arrangement,
providing a cross-terrain robustness assessment.
Comparative analysis across diverse terrain samples shows that
DETS-Net achieves superior segmentation performance in hilly
areas, plateaus, and mountain-basin transitions. PSPNet and
REAUNet suffer a signi ficant loss of terrace textures and structural
integrity across all terrains, while DeepLabv3+ and SegNet com-
monly exhibit gaps in line segmentation, jagged edges, and artifacts.
Although ABCNet, Buildformer, DBBANet, and LANet maintain
basic structural features, they sti ll exhibit over-segmentation,
under-segmentation, and boundary cohesion issues. In contrast,
DETS-Net achieves clearer line reconstruction and more precise
region segmentation across diverse terrains, including dense, wind-
ing terrace patterns in hilly areas, smooth, expansive contours on
plateaus, and complex mountain-basin transitions, with greater
consistency between segmentation results and ground-truth
labels. However, there is still room for improvement: slight discon-
tinuities persist at some texture edges, and certain segmentation
details appear slightly simpli fied relative to the labels.
3.1.2 Ablation experiment
To investigate the impact of core modules, network architec-
ture, and component count on DETS-Net ’s terrace segmentation
performance, this study conducts multiple ablation experiments to
evaluate the effectiveness of the dual-encoder fusion architecture,
M2SA, EDFFN, the number of Transformer blocks, CNN encoder
selection, and the AFF module. All experiments are performed
using the THDIR dataset.

  1. Validation of the dual-encoder architecture
    To demonstrate the necessity of integrating CNN and
    Transformer encoders, we develop single-branch models retaining
    only the CNN or Transformer encoder and compare them with the
    full DETS-Net model. Table 2 presents the results.
    The results demonstrate that the complete model achieves the
    highest values across all metrics, with overall accuracy, MIoU, F1-
    score, precision, and recall reaching 89.91%, 81.28%, 87.99%,
    86.51%, and 89.52%, respectively. While the CNN-only encoder
    achieves higher recall, its precision is suboptimal; the Transformer-
    only encoder exhibits slightly lower precision and recall. The
    integrated DETS-Net model, which combines the CNN and
    Transformer encoders, achieves the best performance across all
    evaluation metrics, indicating that dual-branch fusion effectively
    leverages the complementary advantages of local details and global
    information to significantly enhance terrace segmentation accuracy.
  2. Validation of the functions of the M2SA, EDFNN modules,
    and the sub-components of the AFF module
    For M2SA and EDFFN, we design three experiments: replacing
    M2SA with standard MHSA; removing both M2SA and EDFFN;
    and replacing EDFFN with a conventional FFN. For the AFF
    subcomponents T2C and C2T, we conduct three experiments:
    removing T2C, removing C2T, and swapping their interaction
    order. Table 3 presents the results.
    Replacing M2SA with conventional multi-head self-attention
    significantly degrades performance: MIoU drops to 76.94% and the
    F1-score to 84.81%, indicating that multi-scale multi-head attention
    is crucial for capturing hierarchical features and long-range depen-
    dencies in terraces. Removing both M2SA and EDFFN reduces
    accuracy further, with MIoU falling to 79.59%, highlighting the
    frequency-domain-enhanced feedforward network’s effectiveness in
    improving feature representation and detail capture. Excluding
    EDFFN alone lowers MIoU to 78.81% and the F1-score to
    86.09%, con firming its positive impact on overall performance. In
    the AFF ablation experiment s, removing either T2C or C2T
    significantly degrades accuracy; retaining only a single interaction
    path fails to achieve effective feature fusion, while swapping the
    module order yields slight improvement but still falls short of the
    full model. The complete DETS-Net model attains optimal preci-
    sion, recall, F1-score, and MIoU through the coordinated operation
    of its modules, suggesting that M2SA, EDFFN, and AFF are
    important contributors to performance improvement. Only
    through their synergistic integration can the dual advantages of
    CNN and Transformer be fully leveraged to enhance terrace
    segmentation accuracy.
  3. Veri fication of the number of Transformer blocks
    Four comparative experiments were conducted by setting the
    number of Transformer blocks to 1, 2, 3, and 4 to determine the
    optimal configuration; the results are shown in Table 4 .
    As the number of Transformer blocks increases from 1 to 3,
    overall accuracy, MIoU, the F1-score, and recall all exhibit a
    continuous upward trend. With 2 blocks, precision reaches
    Tian et al. 10.3389/fevo.2026.2011654
    Frontiers in Ecology and Evolution frontiersin.org08

FIGURE 7
Comparison of terraced field identi fication results across different models. Panels (A, B) represent samples from hilly regions; (C, D) represent
samples from plateau regions; and (E, F) represent samples from the mountain-basin transition zone.
TABLE 2 Comparison of CNN-only, Transformer-only, and dual-encoder architectures on the THDIR dataset.
Model OA (%) MIoU (%) F1-score (%) Precision (%) Recall (%)
CNN only 88.47 78.85 86.16 85.36 86.98
Transformer only 88.25 78.51 85.94 84.96 86.94
Full DETS-Net 89.91 81.28 87.99 86.51 89.52
Tian et al. 10.3389/fevo.2026.2011654
Frontiers in Ecology and Evolution frontiersin.org09

88.01%, and with 3 blocks, all metrics attain optimal performance.
However, when the number of blocks increases to 4, all metrics
decline signi ficantly: overall accuracy drops by 1.41 percentage
points, MIoU by 2.38, the F1 score by 1.79, the precision by 1.11,
and the recall by 2.50. This indicates that an appropriate number of
Transformer blocks effectively captures the global semantic features
of terraces and enhances the model ’s representation capacity,
whereas an excessive number of blocks leads to over fitting and
degrades segmentation performance, making three blocks the
optimal con figuration.
4. CNN encoder validation
To identify the optimal CNN encoder, we compare
MobileNetv2, MobileNetv3, ResNet18, and ResNet34 with the
number of Transformer blocks fixed at three. Table 5 presents
the results.
When ResNet34 is used as the CNN encoder, the model
achieves optimal segmentation performance, with OA, MIoU, F1-
score, precision, and recall r eaching 89.91%, 81.28%, 87.99%,
86.51%, and 89.52%, respectively. The performance metrics of the
lightweight networks MobileNetv2 and MobileNetv3 remain at
relatively low levels. Speci fically, MobileNetv3 achieves only
87.29% accuracy, an 84.79% F1-score, and a 76.95% MIoU,
making it the worst-performing encoder across all tests. While
MobileNetv2 shows a slight improvement in recall (89.18%), its F1-
score remains at 85.58% and its MIoU at 77.58%, indicating
inadequate extraction of fine local features in terraces. ResNet18
demonstrates signi ficant improvements, with an F1-score and
MIoU of 86.43% and 79.24%, respectively, yet still lags behind
ResNet34. Deeper architectures such as ResNet50 and ResNet101
fail to achieve further gains: ResNet50 attains an F1-score of 86.13%
and an MIoU of 78.72%, while ResNet101 achieves an F1-score of
86.08% and an MIoU of 78.73%, all lower than those of ResNet34,
suggesting pronounced over fitting in these architectures.
3.1.3 Analysis
We adopt MIoU as the primary accuracy metric and Params
(M) and FLOPs (G) as ef ficiency metrics to evaluate the accuracy-
efficiency balance of DETS-Net and the comparison models in
terrace segmentation. Table 6 shows that DETS-Net achieves
optimal accuracy with a reasonable computational cost, striking a
good balance.
DETS-Net achieves an MIoU of 81.28%, outperforming
DBBANet (79.09%), DeepLabv 3+ (78.27%), and ABCNet
(79.87%). The 1.41 –3.01-percentage-point gain over these three
baselines is relevant in terrace mapping. Despite its higher accuracy,
DETS-Net maintains FLOPs under control, so ef ficiency does
not suffer.
TABLE 3 Ablation experiment results of subcomponents in the M2SA, EDFNN, and AFF modules.
Model OA (%) MIoU (%) F1-score (%) Precision (%) Recall (%)
w/o M2SA replace MHSA 87.27 76.94 84.81 83.57 86.09
w/o M2SA and EDFFN 88.93 79.59 86.83 84.70 89.08
w/o EDFFN replace FFN 88.46 78.81 86.09 85.64 86.55
w/o T2C in AFF 87.85 77.88 85.56 83.97 87.20
w/o C2T in AFF 88.71 79.31 86.66 84.57 88.85
swap T2C and C2T 89.32 80.33 87.43 85.02 89.98
Our Model 89.91 81.28 87.99 86.51 89.52
TABLE 4 Comparison of different numbers of Transformer blocks on the THDIR dataset.
Number of transformer blocks OA (%) MIoU (%) F1-score (%) Precision (%) Recall (%)
×1 89.06 79.78 86.74 86.80 86.68
×2 89.33 80.16 86.91 88.01 85.84
×3 89.91 81.28 87.99 86.51 89.52
×4 88.50 78.90 86.20 85.40 87.02
TABLE 5 Comparison of different CNN encoders on the THDIR dataset.
CNN encoder OA (%) MIoU (%) F1-score (%) Precision (%) Recall (%)
MobileNetv2 87.59 77.58 85.58 82.25 89.18
MobileNetv3 87.29 76.95 84.79 83.77 85.83
ResNet50 88.37 78.72 86.13 84.88 87.41
ResNet34 89.91 81.28 87.99 86.51 89.52
ResNet18 88.71 79.24 86.43 85.81 87.05
ResNet101 88.39 78.73 86.08 85.19 86.99
Tian et al. 10.3389/fevo.2026.2011654
Frontiers in Ecology and Evolution frontiersin.org10

Other models show clearer accuracy-ef ficiency trade-offs
shaped by architectural choices. LANet is parameter-ef ficient, but
its 79.11% MIoU suggests limitations in capturing terraces’ complex
geometries and fragmented patterns. Its intentionally limited re-
ceptive field hinders the resolution of irregular boundaries and
narrow ridge lines. DeepLabv3+ takes a different approach, adding
network complexity and atrous spatial pyramid pooling for multi-
scale context. However, accuracy gains remain modest, suggesting
that the added complexity is not well-matched to the task demands.
ABCNet relies on an attention-based boundary re finement strategy,
achieving a relatively high MIoU.
DETS-Net addresses this trade-off by pairing CNN and
Transformer encoders, raising accuracy while containing computa-
tional complexity. The dual-encoder setup allows each stream to
focus on its strengths: the CNN branch preserves fine spatial details
for precise boundary localization, while the Transformer branch
models long-range dependencies to resolve ambiguities arising
from shadows, spectral overlap with vegetation, and topographic
variation. The AFF module fuses these complementary representa-
tions through bidirectional cross-modal interaction, yielding features
with both local precision and global coherence. The broader principle
is that architectural components should be chosen to address task-
specific challenges, not simply scaled up to increase model capacity.
From the perspective of the accuracy-ef ficiency trade-off, DETS-Net
provides the highest accuracy among the compared models for high-
precision terrace extraction tasks in complex mountainous areas. For
application scenarios with different requirements, different model
selection strategies may be appropriate. For county-level fine-scale
terrace mapping and soil erosion bene fit accounting that demand
high boundary accuracy, DETS-Net is the preferred solution. For
routine daily monitoring tasks that allow a moderate reduction in
accuracy and prioritize computational efficiency, lightweight models
such as ABCNet may be more cost-effective, achieving approximately
80% MIoU with only 3.9 GFLOPs.
3.2 Extraction results for Luotian County
3.2.1 Analysis of extraction results
After training the network proposed in this study, the optimal
model was employed to extract terraces across Luotian County
using 2023 imagery. For county-level analysis, we adopted a sliding-
window approach with multi-batch prediction. Input images were
divided into 256 × 256-pixel overlapping blocks with a 128-pixel
stride to balance context and ef ficiency, and were processed in
batches of eight. For end-to-end vectorization, rasterio extracted
terrace contours from segmentation results, followed by patch
removal, morphological optimization, and manual correction tai-
lored to the study area. The final high-precision vector map is
presented in Figure 8 .
The terraces in Luotian County (red areas) exhibit strip-like and
patch-like patterns corresponding to the mountainous and hilly
terrain. They are concentrated in the southern and western low-
mountainous and hilly regions, while their distribution is sparser in
the northern mountains, showing an overall pattern of higher
density in the south and west than in the north and east. In the
zoomed images A and B, the model-derived terrace boundaries
closely match the textural and contour features in the remote
sensing imagery, enabling accurate identi fication of field bound-
aries, ridge orientations, and precise terrace plot extraction even
under complex topographic conditions.
To further analyze terrace distribution in Luotian County,
slopes were categorized into six grades following the Land
Erosion Classi fication and Grading Standards (SL190 –2007), and
terrace distribution across slope gradients was statistically analyzed,
as summarized in Table 7 . The total terrace area is approximately
37,875.69 ha. Terraces show signi ficant gradient variations, with
gentle slopes predominant; as the slope angle increases, both area
and proportion decline markedly. Areas below 5° account for
14,541.39 ha (38.39%), constituting the primary terrace zone.
Terraces are also widely distributed in the 5 –8° and 8 –15° zones,
covering 9,070.02 and 10,330.65 ha (23.95% and 27.28%), with the
combined proportion of these three gentle slope zones approaching
90%. This pattern indicates that terrace development prioritizes flat
terrain, facilitating both construction and agricultural activities. The
predominance of terraces on slopes below 15° also re flects a long-
term adaptive strategy in which farmers historically selected easily
cultivable land to minimize labor and maximize returns under given
technological conditions. Terrace area on slopes exceeding 15°
decreases signi ficantly; the 15 –25° zone covers 3,557.25 ha
(9.39%). Terraces are scarce above 25°, with the 25 –35° and >35°
zones covering only 355.59 and 20.79 ha, respectively. This sharp
reduction beyond 25° suggests a practical slope threshold for terrace
construction, beyond which engineering feasibility and economic
viability are severely constrained. This pattern re flects the inherent
TABLE 6 Comparison of computational ef ficiency and accuracy across
models.
Model Params (M) FLOPs (G) MIoU (%)
DBBANet 33.65 16.25 79.09
DeepLabv3+ 42.4 11.71 78.27
REAUNet 142.43 254.61 73.46
PSPNet 46.59 45.03 77.35
Buildformer 38.35 29.83 78.81
SegNet 29.44 40.17 77.18
LANet 5.48 23.79 79.11
ABCNet 13.81 3.9 79.87
Our Model 82.07 13.05 81.28
TABLE 7 Area and proportion of terraced fields under different slope
gradients in Luotian County.
Slope gradient ( °) Area/hm 2 Proportion/%
<5 14541.39 38.39
5–8 9070.02 23.95
8–15 10330.65 27.28
15–25 3557.25 9.39
25–35 355.59 0.94

35 20.79 0.05
Tian et al. 10.3389/fevo.2026.2011654
Frontiers in Ecology and Evolution frontiersin.org11

challenges and high costs of construction on steep slopes, along
with soil and water conservation requirements. Moreover, minimal
terrace coverage above 25° underscores the ecological role of these
steep areas, which may be better suited for forest and grassland
restoration to prevent soil erosion than for agricultural use. The
spatial distribution patterns of terraces across slope gradients and
elevation zones are illustrated in Figure 9.
Terrace spatial distribution in Luotian County was analyzed
across six altitude intervals. As shown in Table 8, terrace distribu-
tion exhibits distinct vertical differentiation, with low-altitude areas
forming the core region. Area and proportion decrease progres-
sively with increasing altitude, revealing a strong inverse relation-
ship between elevation and terrace density. Speci fically, areas below
200 m cover 25,652.69 ha (67.53% of the total), constituting the
primary distribution zone. Medium-low altitudes (200 –500 m)
maintain concentrated coverage at 10,724.74 ha (28.23%), with
the low- and medium-low zones together contributing over 95%.
This pronounced concentration re flects the fact that terrace devel-
opment primarily occurs on gentle slopes where favorable topog-
raphy facilitates construction and agricultural management while
meeting production requirements. Moreover, such concentration
suggests that agricultural activities and rural settlements are simi-
larly clustered within lower-elevation bands, given their intrinsic
dependence on cultivated land. In contrast, high-altitude regions
above 500 m show a sharp decline: the 500 –800 m interval accounts
for only 1,563.63 ha (4.12%), while the 800 –1,000 m interval
accounts for only 44.55 ha (0.12%). No terraces are found above
1,000 m, indicating a clear upper altitudinal limit. This pattern
aligns closely with regional topography: high-altitude areas feature
steep slopes, signi ficant topographic variation, challenging con-
struction conditions, and limited soil and water resources.
Notably, the absence of terraces above 1,000 m is not merely due
to construction dif ficulty but also re flects fundamental climatic
constraints, including shorter growing seasons and lower
FIGURE 9
Spatial distribution of terraced fields in Luotian County across different slopes and elevations. (A) Spatial distribution of terraced fields across different
slope gradients; (B) spatial distribution of terraced fields across different elevations.
FIGURE 8
Identification results of terraced fields in Luotian County; insets A and B are magni fied views showing local details.
Tian et al. 10.3389/fevo.2026.2011654
Frontiers in Ecology and Evolution frontiersin.org12

accumulated temperatures that may make crop cultivation increas-
ingly diffi cult at higher elevations. Consequently, these combined
topographic and climatic limitations may make high-altitude zones
unsuitable for terrace development and less suitable for meeting
basic agricultural needs.
3.2.2 Extraction accuracy veri fication
To verify extraction accuracy, we conducted a validation study
using random sampling with visual interpretation. A total of 200
validation points were randomly selected from the extraction map
(Figure 10). Reference land-cover types were determined using remote
sensing imagery. A comparison between the model extraction results
and visual interpretation yielded the following values: TP = 52, TN =
128, FP = 3, and FN = 17. Based on the confusion matrix, we obtained
90.00% accuracy, an F1-score of 83.87%, and an MIoU of 79.36%.
The extraction results achieve 90.00% overall accuracy, dem-
onstrating strong terrace identi fication performance. The low false-
positive count (FP = 3) indicates strong speci ficity in distinguishing
terraces from other land-cover types, minimizing commission
errors. Both the F1-score (83.87%) and MIoU (79.36%) exceed
79%, indicating good precision-recall balance and strong alignment
with actual terrace distribution. The MIoU value of 79.36% suggests
that extracted terrace boundaries maintain satisf actory spatial
overlap with ground truth, which is particularly important for
area-dependent analyses such as soil erosion modeling. These
validation results con firm that extraction accuracy meets the
requirements for subsequent research and that the results provide
reliable foundational data for re fined CSLE engineering factor
evaluation, soil erosion estimation, and the analysis of the soil
and water conservation bene fit of terraces.
4 Discussion
4.1 Topographic constraints and spatial
distribution of terraces
Based on DETS-Net extraction results, the total terrace area in
Luotian County is 37,875.69 ha, w ith the spatial distribution
exhibiting pronounced topographic constraints. Regarding slope
distribution, terraces show clear variation across slope gradients.
The terrace area on slopes below 5° is 14,541.39 ha (38.39% of the
total), making this the primary terrace zone. Terraces are also
widely distributed in the 5–8° and 8–15° intervals, with areas of
9,070.02 ha (23.95%) and 10,330.65 ha (27.28%), respectively. The
combined proportion of these three gentle-slope zones approaches
90%, indicating preferential selection of flat terrain for ter-
race development.
As the slope gradient increases, both terrace distribution area
and proportion decline signifi cantly. The 15 –25° interval contains
FIGURE 10
Distribution of sample points.
TABLE 8 Area and proportion of terraced fields at different altitudes in
Luotian County.
Altitude/m Area/hm 2 Proportion/%
<200 25652.69 67.53
200–500 10724.74 28.23
500–800 1563.63 4.12
800–1000 44.55 0.12
1000–1500 N/A N/A

1500 N/A N/A
Tian et al. 10.3389/fevo.2026.2011654
Frontiers in Ecology and Evolution frontiersin.org13

3,557.25 ha (9.39%); the 25 –35° interval contains 355.59 ha; and on
slopes exceeding 35°, terrace distribution is exceedingly sparse, at
only 20.79 ha. This pattern re flects combined physiographic and
agricultural engineering constraints.
From a geomorphological perspective, the spatial patterns of
terraces in hilly regions are governed by multiple factors. Slope
determines terrace surface width and earthwork volume: as the
slope gradient increases, achieving the same width for contour
farming requires a greater slope length, leading to increased ridge
density and reduced individual plot area. Moreover, terrace con-
struction on steep slopes involves higher engineering requirements
and costs.
From a regulatory perspective, Article 23 of China ’s Soil and
Water Conservation Law requires soil and water conservation
measures when reclaiming slopes between 5° and the prohibited
reclamation limit for crop cultivation. Our results show that, despite
differentiated management requirements across slope gradients,
terraces on slopes exceeding 5° still account for a large proportion
(61.61%) of mapped terrace area. This may re flect that farmers,
based on long-term practical considerations of engineering feasi-
bility and economic returns, tend to actively develop and utilize
terraces on steeper slopes.
4.2 Complementarity of the DETS-Net
dual-encoder architecture
DETS-Net achieves 89.91% overall accuracy and 81.28% MIoU
on the THDIR dataset, substantially outperforming eight main-
stream segmentation models, including DBBANet and DeepLabv3
+. Ablation experiments further reveal each component ’s contribu-
tion to the dual-encoder architecture: MIoU is 78.85% with only the
CNN encoder, 78.51% with only the Transformer encoder, and rises
to 81.28% when both are fused via the AFF module. This 2.77
percentage-point improvement con firms that the CNN and
Transformer are complementary, and that combining them yields
stronger results than either encoder alone.
Visual analysis helps explain this complementarity. The CNN
encoder is particularly effective at capturing edge textures along
terrace boundaries, which is important for identifying narrow and
fragmented ridge lines. This capability follows naturally from the
inductive biases of convolutional operations —locality and transla-
tion equivariance —which make the CNN sensitive to fine-grained
spatial detail. The Transformer encoder, in contrast, excels at
modeling long-range dependencies via self-attention, enabling the
recovery of internal regions that appear as voids due to uneven
illumination or shadow occlusion. Because self-attention aggregates
contextual information across the entire feature map, the
Transformer helps overcome the CNN ’s limited receptive field,
allowing regions with weak local signals to be inferred from global
context. The AFF module bridges these complementary represen-
tations. Removing either the T2C or C2T submodule causes a clear
performance drop, indicating that explicit cross-branch interaction
is important for balancing detail preservation and semantic com-
pleteness. The bidirectional fusion design allows each encoder to
draw on the other ’s strengths, rather than simply concatenating
outputs, which may help explain why straightforward feature
summation cannot match the full model ’s performance.
4.3 Limitations
Although our study yielded satisfactory results, several limita-
tions merit discussion. The vali dation samples were collected
exclusively from Luotian County, which restricts the assessment
of DETS-Net ’s generalization capability across other geomorphic
regions. Moreover, the study does not distinguish between culti-
vated and abandoned terraces; the latter may exhibit reduced
extraction accuracy due to vegetation recovery. Furthermore, the
high cost of 0.5 m-resolution imagery imposes data acquisition
constraints that hinder large-scale application.
5 Conclusion
This study focuses on Luotian County in the Dabie Mountains
of Hubei Province, and develops a dedicated THDIR dataset and the
DETS-Net dual-encoder segmentation network to achieve high-
precision terrace extraction under complex mountainous condi-
tions. The main conclusions are summarized as follows:

  1. Dataset contribution: The THDIR dataset comprises 9,336
    standardized sample images covering typical terrain types,
    including plateaus, mountains, hills, and basin transition
    zones, with standardized annotations and balanced distri-
    bution, thus establishing a benchmark for terrace segmen-
    tation in complex scenes.
  2. Model contribution: DETS-Net attains 89.91% OA, 81.28%
    MIoU, and 87.99% F1-score on the THDIR dataset, outper-
    forming eight mainstream models such as DBBANet. Ablation
    experiments demonstrate the contribution of the dual-encoder
    architecture, M2SA, EDFFN, and AFF. The con figuration
    combining ResNet34 with three Transformer blocks proves
    optimal, striking a balance between accuracy and efficiency.
  3. Application contribution: The extraction results from
    DETS-Net indicate that the total terrace area in Luotian
    County is approximately 37,875.69 ha, exhibiting a spatial
    distribution characterized by higher density in the south and
    west and lower density in the north and east, with over 95%
    of the terrace area concentrated in regions below 500 meter
    in elevation and with slopes under 15°. Random-sampling
    validation yields an extraction accuracy of 90.00%, an F1-
    score of 83.87%, and an MIoU of 79.36%, providing reliable
    data for terrace resource monitoring.
  4. Future work: We propose four targeted directions to further
    promote the practical deployment of the method. First, we
    will extend the THDIR dataset to include samples from
    additional geomorphic regions, enabling rigorous evaluation
    and enhancement of DETS-Net ’s cross-regional generaliza-
    tion. Second, we will integrate multitemporal remote sensing
    imagery for automatic differentiation between cultivated and
    Tian et al. 10.3389/fevo.2026.2011654
    Frontiers in Ecology and Evolution frontiersin.org14

abandoned terraces, thereby supporting dynamic monitoring
of terrace resource utilization. Third, we will investigate
lightweight model compression and its integration with
Chinese high-resolution satellite data (e.g., GF-7). We will
evaluate and optimize the model’s inference performance on
lower-end of fice devices commonly used by local soil and
water conservation institutions, thereby reducing hardware
requirements for routine business applications. In addition,
we will reduce data acquisition costs through the use of
Chinese satellite data, thus facilitating operational deploy-
ment across larger administrative regions. Fourth, we intend
to incorporate topographic and climatic auxiliary variables
into a multitask learning framework to uncover the driving
mechanisms of terrace distribution and abandonment,
thereby offering more scienti fic guidance for soil and water
conservation planning and sustainable land management.
Data availability statement
The original contributions presented in the study are included
in the article/supplementary material, further inquiries can be
directed to the corresponding author/s.
Author contributions
PT: Methodology, Writing – original draft, Conceptualization,
Writing – review & editing. YS: Writing – review & editing, Writing
– original draft. YaC: Writing – original draft, Visualization,
Conceptualization. YuC: Vali d a t i o n ,M e t h o d o l o g y ,W r i t i n g–
review & editing. RD: Writing – review & editing, Supervision,
Visualization. ZZ: Supervision, Writing – review & editing. BM:
Visualization, Supervision, Writing – review & editing. RX: Writing
– review & editing, Supervision.
Funding
The author(s) declared that financial support was received for
this work and/or its publication. This research was jointly funded by
the National Natural Science Foundation of China(42377354), the
“DoubleHundred” Science and Technology Empowerment Project
of Hubei Province(2026EIA033), the project of Hubei Key
Laboratory of Basin Water Security(QTKS0034W26144), and the
self-determined research funds of CCNU from the colleges ’ basic
research and operation of MOE (XJ2026007401).
Conflict of interest
BM was employed by company Changjiang Survey, Planning,
Design and Research Co., Ltd.
The remaining author(s) declared that this work was conducted
in the absence of any commercial or financial relationships that
could be construed as a potential con flict of interest.
Generative AI statement
The author(s) declared that generative AI was used in the
creation of this manuscript. AI was used for language polishing,
English translation and manuscript expression optimization. All
research content, experimental design, data analysis and conclu-
sions were completed by the authors. The authors take full respon-
sibility for the content of this manuscript.
Any alternative text (alt text) provided alongside figures in this
article has been generated by Frontiers with the support of arti ficial
intelligence and reasonable e fforts have been made to ensure
accuracy, including review by the authors wherever possible. If
you identify any issues, please contact us.
Publisher’s note
All claims expressed in this article are solely those of the authors
and do not necessarily represent those of their af filiated organiza-
tions, or those of the publisher, the editors and the reviewers. Any
product that may be evaluated in this article, or claim that may be
made by its manufacturer, is not guaranteed or endorsed by
the publisher.
References
Ackermann, O., Zhevelev, H. M., and Svoray, T. (2019). Agricultural systems and
terrace pattern distribution and preservation along climatic gradient: From sub-humid
mediterranean to arid conditions. Quaternary International Landforms Sediments Soils
Palaeosols as Records Present Former Environ. Conditions Human-Environment
Interact. 502, 319 –326. doi: 10.1016/j.quaint.2018.09.032
Ajith, A., Anto Francis, K., and Pillai, R. J. (2024). Evaluation of pore-pressure variation
and slope stability on terraced cultivation using physics-based landslide susceptibility
model. Geomorphology 450, 109081. doi: 10.1016/j.geomorph.2024.109081
Amodio, A. M., Bonazza, A., Sardella, A., and Gizzi, F. T. (2025). “Semi-automated
human terrace recognition: A case study on Panarea Island, Italy, ” in Computational
Science and its Applications – ICCSA 2025 Workshops , vol. 15894 . Eds. O. Gervasi, B.
Murgante, C. Garau, Y. Karaca, M. N. Faginas Lago and F. Scorza (Springer Nature
Switzerland), 18 –29. doi: 10.1007/978-3-031-97648-3_2
A t t r i ,I . ,A w a s t h i ,L .K . ,S h a r m a ,T .P . ,a n dR a t h e e ,P .( 2 0 2 3 ) .Ar e v i e wo fd e e p
learning techniques used in agriculture. Ecol. Inf. 77, 102217. doi: 10.1016/
j.ecoinf.2023.102217
Badrinarayanan, V., Kendall, A., and Cipolla, R. (2016). “SegNet: A deep convolutional
encoder-decoder architecture for image segmentation (arXiv:1511.00561),” in Arxiv.
(Los Alamitos, California: IEEE Computer Society). doi: 10.48550/arXiv.1511.00561
Blaschke, T. (2010). Object based image analysis for remote sensing. ISPRS J.
Photogramm. Remote Sens. 65, 2 –16. doi: 10.1016/j.isprsjprs.2009.06.004
Tian et al. 10.3389/fevo.2026.2011654
Frontiers in Ecology and Evolution frontiersin.org15

Cammeraat, E., Anders, N., Scholten, R., and Boix-Fayos, C. (2024). Erosion and
sedimentation on an abandoned terraced and gullied hillslope. CATENA 246, 108393.
doi: 10.1016/j.catena.2024.108393
Cao, H., Wu, Z., and Zheng, W. (2024). Impact of touristification and landscape pattern
on habitat quality in the Longji Rice Terrace Ecosystem, southern China, based on
geographically weighted regression models. Ecol. Indic. 166, 112259. doi: 10.1016/
j.ecolind.2024.112259
Chen, D., Wei, W., and Chen, L. (2017a). Effects of terracing practices on water erosion
control in China: A meta-analysis. Earth Sci. Rev. 173, 109 –121. doi: 10.1016/
j.earscirev.2017.08.007
Chen, L.-C., Papandreou, G., Kokkinos, I., Murphy, K., and Yuille, A. L. (2017b).
“DeepLab: Semantic image segmentation with deep convolutional nets, atrous convolu-
tion, and fully connected CRFs (arXiv:1606.00915),” in Arxiv (Los Alamitos, CA: IEEE
Computer Society). doi: 10.48550/arXiv.1606.00915
Chen, L.-C., Zhu, Y., Papandreou, G., Schroff, F., and Adam, H. (2018). “Encoder-
decoder with atrous separable convolut ion for semantic image segmentation
(arXiv:1802.02611),” in Arxiv. (Cham: Springer). doi: 10.48550/arXiv.1802.02611
Chen, J., Lu, Y., Yu, Q., Luo, X., Adeli, E., Wang, Y., et al. (2021). “TransUNet:
Transformers make strong encoders for medical image segmentation (arXiv:2102.04306),”
in Arxiv. doi: 10.48550/arXiv.2102.04306
Chen, J., Chen, G., Zhou, P., He, Y., Yue, L., Ding, M., et al. (2025a). CTSeg: CNN and
ViT collaborated segmentation framework for ef ficient land-use/land-cover mapping
with high-resolution remote sensing images. Int. J. Appl. Earth Obs. Geoinf. 139,
104546. doi: 10.1016/j.jag.2025.104546
Chen, S., Wang, X., Liu, T., Xie, M., and Lin, Q. (2025b). Using geo-data and social
media images to explore the supply and demand of cultural ecosystem services for
terraces in China. Ecosyst. Serv. 76, 101778. doi: 10.1016/j.ecoser.2025.101778
Cheng, H., Feng, H., Zhang, X., Li, Z., Yu, K., Ma, H., et al. (2025). Terracing stabilizes
and enhances soil organic carbon sequestration bene fits of revegetation on the Loess
Plateau, China. CATENA 255, 109018. doi: 10.1016/j.catena.2025.109018
Cucchiaro, S., Paliaga, G., Fallu, D. J., Pears, B. R., Walsh, K., Zhao, P., et al. (2021).
Volume estimation of soil stored in agricultural terrace systems: A geomorphometric
approach. CATENA 207, 105687. doi: 10.1016/j.catena.2021.105687
Ding, L., Tang, H., and Bruzzone, L. (2021). LANet: Local attention embedding to
improve the semantic segmentation of remote sensing images. IEEE Trans. Geosci.
Remote Sens. 59, 426 –435. doi: 10.1109/TGRS.2020.2994150
Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T.,
et al. (2021). “An image is worth 16x16 words: Transformers for image recognition at
scale (arXiv:2010.11929),” in Arxiv. doi: 10.48550/arXiv.2010.11929
Fan, R., Fang, N., Zong, R., Qiang, R., Zeng, Y., Wang, Q., et al. (2026). Spatial
prioritization of terrace construction on the Chinese Loess Plateau using interpretable
machine learning. Int. Soil Water Conserv. Res. 14, 100642. doi: 10.1016/
j.iswcr.2026.100642
Fang, K., Zhang, S., Su, T., Wang, B., Jia, G., and Zhang, X. (2026). MTFormer: Joint
uncertainty-aware abrupt-change distance transform and multi-task transformer for
agricultural parcel segmentation. Expert Syst. Appl. 319, 132068. doi: 10.1016/
j.eswa.2026.132068
Fu, S., Yang, Y., Liu, B., Liu, H., Liu, J., Liu, L., et al. (2020). Peak flow rate response to
vegetation and terraces under extreme rainstorms. Agric. Ecosyst. Environ. 288, 106714.
doi: 10.1016/j.agee.2019.106714
Hankui K Zhang, Shi Qiu, Ji Won Suh, and Dong Luo (2024). Machine learning and
deep learning in remote sensing data analysis.
Hu, J., Shen, L., Albanie, S., Sun, G., and Wu, E. (2018). “Squeeze-and-excitation
networks (arXiv:1709.01507),” in Arxiv. (Los Alamitos, California: IEEE Computer
Society). doi: 10.48550/arXiv.1709.01507
Kattenborn, T., Leitloff, J., Schiefer, F., and Hinz, S. (2021). Review on convolutional
neural networks (CNN) in vegetation remote sensing. ISPRS J. Photogramm. Remote
Sens. 173, 24–49. doi: 10.1016/j.isprsjprs.2020.12.010
LeCun, Y., Bengio, Y., and Hinton, G. (2015). Deep learning. Nature 521, 436 –444.
doi: 10.1038/nature14539
Li, J., Wei, Y., Wei, T., and He, W. (2025). A comprehensive deep-learning framework
for fine-grained farmland mapping from high-resolution images. IEEE Trans. Geosci.
Remote Sens. 63, 1 –15. doi: 10.1109/tgrs.2024.3515157
Li, Y., Tian, F., Zhang, M., Zeng, H., Ahmed, S., Qin, X., et al. (2025). A 10-meter global
terrace mapping using sentinel-2 imagery and topographic features with deep learning
methods and cloud computing platform support. Int. J. Appl. Earth Obs. Geoinf. 139,
104528. doi: 10.1016/j.jag.2025.104528
Li, R., Zheng, S., Zhang, C., Duan, C., Wang, L., and Atkinson, P. M. (2021). ABCNet:
Attentive bilateral contextual network for ef ficient semantic segmentation of fine-
resolution remotely sensed imagery. ISPRS J. Photogramm. Remote Sens. 181, 84 –98.
doi: 10.1016/j.isprsjprs.2021.09.005
Li, H., Zhu, J., Mao, X., Hao, X., Li, S., Yu, Q., et al. (2026). Achieving precise cropland
parcel extraction from remote sensing images through integration of segment anything
model and adaptive mask re finement. Comput. Electron. Agric. 243, 111347.
doi: 10.1016/j.compag.2025.111347
Lin, T.-Y., Dolla ́r, P., Girshick, R., He, K., Hariharan, B., and Belongie, S. (2017).
“Feature pyramid networks for object detection (arXiv:1612.03144),” in Arxiv. (Los
Alamitos, California: IEEE Computer Society). doi: 10.48550/arXiv.1612.03144
Liu, Z., Chen, G., Tang, B., Wen, Q., Tan, R., and Huang, Y. (2024). Regional scale
terrace mapping in fragmented mountainous areas using multi-source remote sensing
data and sample puri fication strategy. Sci. Total Environ. 925, 171366. doi: 10.1016/
j.scitotenv.2024.171366
Liu, Z., Lin, Y., Hu, H., and Ca, Y. (2021). Swin transformer: Hierarchical vision
transformer using shifted windows | IEEE Conference Publication | IEEE Xplore. 9992 –
10002. doi: 10.1109/ICCV48922.2021.00986
Liu, T., Yu, L., Liu, X., and Zhao, Q. (2025). A Global Review of Monitoring Cropland
Abandonment Using Remote Sensing: Temporal –Spatial Patterns, Causes, Ecological
Effects, and Future Prospects | Journal of Remote Sensing . (Washington, DC, USA:
American Association for the Advancement of Science (AAAS)). doi: 10.34133/
remotesensing.0584
Long, J., Shelhamer, E., and Darrell, T. (2015). Fully convolutional networks for
semantic segmentation. In: 2015 IEEE Conference on Computer Vision and Pattern
Recognition (CVPR) (Boston, MA: IEEE), 3431 –40. doi: 10.1109/CVPR.2015.7298965
Long, M., Tan, R., Zhao, Y., Li, X., Zhou, C., and Zhang, D. (2025). The rice terrace
abandonment and its challenge for food security in the rural villages of southern China.
Habitat Int. 160, 103398. doi: 10.1016/j.habitatint.2025.103398
Lu, R., Liao, R., Meng, R., Hu, Y., Zhao, Y., Guo, Y., et al. (2025b). Strategic sampling for
training a semantic segmentation model in operational mapping: Case studies on
cropland parcel extraction. Remote Sens. Environ. 331, 115034. doi: 10.1016/
j.rse.2025.115034
Lu, D., Su, K., Wang, Z., Hou, M., Li, X., Lin, A., et al. (2025a). Patterns and drivers of
terrace abandonment in China: Monitoring based on multi-source remote sensing data.
Land Use Policy 148, 107388. doi: 10.1016/j.landusepol.2024.107388
Lu, R., Zhang, Y., Huang, Q., Zeng, P., Shi, Z., and Ye, S. (2024). A re fined edge-aware
convolutional neural networks for agricultural parcel delineation. Int. J. Appl. Earth
Obs. Geoinf. 133, 104084. doi: 10.1016/j.jag.2024.104084
Luo, W., Li, Y., Urtasun, R., and Zemel, R. (2017). “Understanding the effective receptive
field in deep convolutional neural networks (arXiv:1701.04128),” in Arxiv. (Red Hook,
New York: Curran Associates (NeurIPS 2016 proceedings)). doi: 10.48550/
arXiv.1701.04128
Lv, J., Shen, Q., Lv, M., Li, Y., Shi, L., and Zhang, P. (2023). Deep learning-based
semantic segmentation of remote sensing images: A review. Front. Ecol. Evol. 11,
1201125. doi: 10.3389/fevo.2023.1201125
Ma, L., Liu, Y., Zhang, X., Ye, Y., Yin, G., and Johnson, B. A. (2019). Deep learning in
remote sensing applications: A meta-analysis and review. ISPRS J. Photogramm.
Remote Sens. 152, 166 –177. doi: 10.1016/j.isprsjprs.2019.04.015
Patric, D., Kappas, M., Wyss, D., and Schaab, G. (2026). Semi-automated object-based
image analysis for peri-urban land use/cover classi fication using WorldView-3 imagery
of Mankweng, South Africa. Sci. Rep. 16, 27075. doi: 10.1038/s41598-026-66042-x
Pijl, A., Quarella, E., Vogel, T. A., D ’Agostino, V., and Tarolli, P. (2021). Remote
sensing vs. Field-based monitoring of agricultural terrace degradation. Int. Soil Water
Conserv. Res. 9, 1 –10. doi: 10.1016/j.iswcr.2020.09.001
Q a m a r ,R . ,A s h r a f ,S . ,A b b a s ,F .T . ,J a v e e d ,H .M .R . ,N a d e e m ,M .A . ,A t i q u e - u r - R e h m a n ,
et al. (2025).“Terrace farming for better water management in agriculture,” in Innovations
in Agricultural Water Management.E d s .M .M u b e e n ,W .N .J a t o i ,M .Z .H a s h m ia n dM .
Ahmad (Springer Nature Switzerland), 131–158. doi: 10.1007/978-3-031-91883-4_7
Ronneberger, O., Fischer, P., and Brox, T. (2015). “U-Net: Convolutional networks for
biomedical image segmentation (arXiv:1505.04597), ” in Arxiv. (Cham: Springer).
doi: 10.48550/arXiv.1505.04597
Saiz, G., Wandera, F. M., Pelster, D. E., Ngetich, W., Okalebo, J. R., Ru fino, M. C., et al.
(2016). Long-term assessment of soil and water conservation measures (Fanya-juu
terraces) on soil organic matter in South Eastern Kenya. Geoderma 274, 1 –9.
doi: 10.1016/j.geoderma.2016.03.022
Shi, Y., and Wei, W. (2025). Terracing drives Chinese Loess Plateau toward carbon
neutrality: Spatiotemporal interaction between land use transitions and SOC storage. J.
Environ. Manage. 385, 125760. doi: 10.1016/j.jenvman.2025.125760
Sibomana, P., Vanmaercke, M., Depicker, A., Tychon, B., Hubert, A., and Dewitte, O.
(2025). Effects of agricultural terraces on landslide occurrence: Insights from a tropical
mountainous region (Rwanda, Africa). CATENA 253, 108898. doi: 10.1016/
j.catena.2025.108898
Sofia, G., Marinello, F., and Tarolli, P. (2014). A new landscape metric for the
identification of terraced sites: The Slope Local Length of Auto-Correlation (SLLAC).
ISPRS J. Photogramm. Remote Sens. 96, 123 –133. doi: 10.1016/j.isprsjprs.2014.06.018
Song, H., Li, X., He, Y., Wang, X., Xin, L., Tan, M., et al. (2026). Understanding terraced
cropland marginalization in China ’s mountainous areas: Process, characteristics, and
underlying mechanisms. J. Rural Stud. 125, 104177. doi: 10.1016/
j.jrurstud.2026.104177
Sui, Y., Niu, C., Jia, Y., Tan, X., Wang, D., Wang, X., et al. (2026). Quantifying
anthropogenic impacts on water conservation and exploring enhancement pathways
within a coupled nature-society framework. J. Hydrology: Regional Stud. 64, 103269.
doi: 10.1016/j.ejrh.2026.103269
Tian et al. 10.3389/fevo.2026.2011654
Frontiers in Ecology and Evolution frontiersin.org16

Sun, H., Gong, R., Chu, M., and Wang, Y. (2025). HRC-Deeplab: Improved Deeplabv3

  • for semantic segmentation of road traf fic markings. Signal. Image Video Process. 19,
  1. doi: 10.1007/s11760-025-05008-4
    Sun, W., Li, J., Zhang, Y., Zhao, Y., Fang, N., Wang, L., et al. (2026). Terrace failure and
    soil loss induced by seepage in a coarse-grained brown soil. CATENA 271, 110205.
    doi: 10.1016/j.catena.2026.110205
    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., et al. (2017).
    “Attention is all you need (arXiv:1706.03762),” in Arxiv. (Red Hook, New York: Curran
    Associates (NeurIPS 2017 proceedings)). doi: 10.48550/arXiv.1706.03762
    Wang, L., Fang, S., Li, R., and Meng, X. (2022). Building extraction with vision
    transformer. IEEE Trans. Geosci. Remote Sens. 60, 1 –11. doi: 10.1109/
    TGRS.2022.3186634
    Wang, C., Ge, X., Zhao, Y., Cai, J., Xia, Y., Wen, Z., et al. (2026). A comprehensive
    survey on segmentation-enabled precision agriculture: Methodologies, applications,
    and the road to agriculture 5.0. Comput. Electron. Agric. 251, 111986. doi: 10.1016/
    j.compag.2026.111986
    Wei, W., Chen, D., Wang, L., Daryanto, S., Chen, L., Yu, Y., et al. (2016). Global
    synthesis of the classi fications, distributions, bene fits and issues of terracing. Earth Sci.
    Rev. 159, 388 –403. doi: 10.1016/j.earscirev.2016.06.010
    Woo, S., Park, J., Lee, J.-Y., and Kweon, I. S. (2018). “CBAM: convolutional block
    attention module (arXiv:1807.06521),” in Arxiv. (Cham: Springer). doi: 10.48550/
    arXiv.1807.06521
    Xu, L., Ming, D., Du, T., Chen, Y., Dong, D., and Zhou, C. (2022). Delineation of
    cultivated land parcels based on deep convolutional networks and geographical
    thematic scene division of remotely sensed images. Comput. Electron. Agric. 192,
  2. doi: 10.1016/j.compag.2021.106611
    Xu, F., Yao, X., Zhang, K., Yang, H., Feng, Q., Li, Y., et al. (2024). Deep learning in
    cropland field identi fication: A review. Comput. Electron. Agric. 222, 109042.
    doi: 10.1016/j.compag.2024.109042
    Yan, Y., Zhang, Z., Zan, C., Fang, S., Shao, Z., Wang, J., et al. (2026). HSRRSFPD and
    MEGRDNet: Empowering high-resolution farmland parcel mapping in complex
    terrains. Int. J. Appl. Earth Obs. Geoinf. 150, 105346. doi: 10.1016/j.jag.2026.105346
    Yang, L., Li, Y., and Wei, X. (2025a). Validating social media data reliability for cultural
    ecosystem services mapping in data-scarce GIAHS: A mixed-methods approach in hani
    rice Terraces. Ecol. Indic. 175, 113530. doi: 10.1016/j.ecolind.2025.113530
    Yang, Y., Zhan, J., Zhuang, J., Zhu, W., and Peng, J. (2025b). Space-air-ground-interior
    integrated framework for mapping loess landslide fissures: A case study in the
    Heifangtai terrace, Northwest China. Bull. Eng. Geol. Environ. 84, 459. doi: 10.1007/
    s10064-025-04493-8
    Yaseen, M., Mellios, N., Pagano, A., Portoghese, I., Giordano, R., Iacobellis, V., et al.
    (2025). Quantitative assessment of best management practices for soil and water
    conservation: A case study from the Tarquinia plain. Agric. Water Manage. 322,
  3. doi: 10.1016/j.agwat.2025.109952
    Zhang, Q., Cao, Z., and Liu, Y. (2024). Assessing rural revitalization potential through
    rural transformation degree and sustainability: A quantitative study of 460 case villages
    in Lingbao, China. Habitat Int. 153, 103194. doi: 10.1016/j.habitatint.2024.103194
    Zhang, C., Harrison, P. A., Pan, X., Li, H., Sargent, I., and Atkinson, P. M. (2020). Scale
    Sequence Joint Deep Learning (SS-JDL) for land use and land cover classi fication.
    Remote Sens. Environ. 237, 111593. doi: 10.1016/j.rse.2019.111593
    Zhang, Y., Shi, M., Zhao, X., Wang, X., Luo, Z., and Zhao (2017). Methods for automatic
    identification and extraction of terraces from high spatial resolution satellite data
    (China-GF-1). Int. Soil Water Conserv. Res. 5, 17–25. doi: 10.1016/j.iswcr.2017.02.002
    Zhao, Y., Cheng, H., Peng, S., Liu, S., Xie, Y., and Liu, B. (2026b). High-Resolution (0.9
    m) terrace mapping in Low-Latitude hilly regions using deep learning and Area –Slope
    Denoising: A case study from Guangdong Province, China. Int. J. Appl. Earth Obs.
    Geoinf. 146, 105126. doi: 10.1016/j.jag.2026.105126
    Zhao, H., Shi, J., Qi, X., Wang, X., and Jia, J. (2017). “Pyramid scene parsing network
    (arXiv:1612.01105),” in Arxiv. (Los Alamitos, California: IEEE Computer Society
    (CVPR 2017 proceedings)). doi: 10.48550/arXiv.1612.01105
    Zhao, A., Wang, X., Wang, N., Pan, Y., and Zhong, Y. (2026a). Towards individual
    agricultural parcel vectorization from VHR imagery: A coarse-to- fine and multi-task
    learning method. ISPRS J. Photogramm. Remote Sens. 238, 191 –229. doi: 10.1016/
    j.isprsjprs.2026.05.003
    Zhao, H., Wu, B., Zhang, M., Long, J., Tian, F., Xie, Y., et al. (2025). A large-scale VHR
    parcel dataset and a novel hierarchical semantic boundary-guided network for
    agricultural parcel delineation. ISPRS J. Photogramm. Remote Sens. 221, 1 –19.
    doi: 10.1016/j.isprsjprs.2025.01.034
    Zheng, S., Lu, J., Zhao, H., Zhu, X., Luo, Z., Wang, Y., et al. (2021). “Rethinking
    semantic segmentation from a sequence-to-sequence perspective with transformers,” in
    2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , (Los
    Alamitos, California: IEEE Computer Society (CVPR 2021 proceedings)), 6877 –6886.
    doi: 10.1109/CVPR46437.2021.00681
    Zhu, X. X., Tuia, D., Mou, L., Xia, G.-S., Zhang, L., Xu, F., et al. (2017). Deep learning in
    remote sensing: A comprehensive review and list of resources. IEEE Geosci. Remote
    Sens. Mag. 5, 8 –36. doi: 10.1109/MGRS.2017.2762307
    Tian et al. 10.3389/fevo.2026.2011654
    Frontiers in Ecology and Evolution frontiersin.org17

论文原文与出版信息

Frontiers in Ecology and Evolution
查看论文原文

继续探索