Terraced field extraction in UAV imagery using improved DeepLabv3+ network

Accurately and timely extracting the area and geographic distribution of terraced fields is crucial for achieving dynamic monitoring of soil erosion and precise agricultural management. In this study, we present an improved DeepLabv3+ model for extracting terraced fields from ultra-high-resolution UAV images using a typical terraced area in Dingnan County, Jiangxi Province, China. The model employs the EfficientNetv2 backbone for feature extraction, integrates the CBAM attention mechanism, and optimizes the loss function. This approach effectively balances the accuracy and speed of terraced field extraction from ultra-high-resolution UAV images. By achieving high extraction accuracy while reducing model parameters, the model meets the requirements for timeliness. The experimental results indicate that the proposed terraced field extraction model is feasible for UAV images, with an overall extraction accuracy of 90.57% for terraced fields, demonstrating high extraction precision. In terms of extraction speed, the average frame rate reaches 121.03, which suggests that the method meets the requirements for timeliness. This approach offers a new and improved strategy for enhancing the accuracy and speed of terraced field extraction.