Background As an effective engineering measure for soil and water loss control, terraced fields provide crucial data support for evaluating regional soil and water conservation benefits and planning agricultural production layouts if their spatial distribution information is acquired rapidly and accurately. Aiming at the problems of difficult boundary and field surface recognition, as well as morphological fragmentation and dispersion of terraced fields in high-resolution remote sensing images, this study constructed a set of high-resolution terraced field datasets and proposed the AMSL-Net model. This model balances the requirements of segmentation accuracy and lightweight, providing technical support for the refined monitoring of terraced fields.

Methods In the encoder stage of the AMSL-Net model, the lightweight MobileNetV3 is employed as the backbone network. The ASPP-CBAM module is introduced to extract multi-scale features of terraced fields, and long skip connections are utilized to establish long-range dependencies, enabling the fusion of features at different levels of a given image. In the decoder stage, deformable convolutions are applied to adapt to the shape variations of terraced fields, and the DySample dynamic upsampler is used to enhance classification accuracy.

Results Five models, namely Seg-Net, CPF-Net, PSP-Net, U-Net++, and DeepLabV3, were selected for comparative experiments on the same test set, and the results are as follows: 1) Comparative experiment design and core accuracy performance: The proposed model performs excellently in multiple performance metrics, with an overall accuracy (OA) of 96.18%, an intersection over union (IoU) of 90.97%, and an F1-score of 95.24%, which is a significant improvement in accuracy compared with other methods. 2) Model lightweight performance: The proposed model has only 6.05 M parameters (Params) and 6.84 G floating-point operations (FLOPs), which significantly reduces the computational cost compared with traditional models. 3) Ablation experiment verification: The results of ablation experiments verify that each module has an obvious promoting effect on terraced field recognition.

Conclusions This model exhibits high accuracy in segmenting terraced fields from high-resolution remote sensing images, offering a valuable reference for the refined monitoring and management of terraced fields.