关键词: 耕地提取, 对象上下文网络, 空洞空间金字塔池化, 联合损失函数, 山地丘陵地区

Abstract:

Cultivated land serves as the core guarantee of food security, and its accurate extraction is of great significance for optimizing the allocation of land resources and advancing sustainable agricultural development. Although the object context network (OCNet) can capture contextual information from remote sensing images to a certain extent, it suffers from weak extraction capability and single-scale limitations in semantic aggregation when extracting cultivated land in mountainous and hilly regions characterized by fragmented patches, large scale disparities and uneven spatial distribution. To address these drawbacks, this study proposes a multi-scale feature enhancement-OCNet (MSFE-OCNet) algorithm. To strengthen the model's ability to identify cultivated land of diverse sizes, the algorithm constructs an MSFE block integrating the atrous spatial pyramid pooling (ASPP) module for local detail feature extraction and a multi-branch OC module for global contextual modeling. The ASPP module adopts dilation rates of (12, 24, 36), while the multi-branch OC module utilizes receptive field combinations of (1, 2, 4). Meanwhile, a weighted cross entropy-dice combined loss function is introduced during the training phase to mitigate insufficient training of small-sample land cover categories. To validate the performance of the proposed algorithm, this study takes Nanchang, Jingdezhen and Yingtan as representative areas for cultivated land extraction, and conducts a series of comparative and ablation experiments. Across three typical mountainous and hilly study areas (Nanchang, Jingdezhen and Yingtan), the MSFE-OCNet outperforms competing algorithms including U-Net, GTNet, SAM and DeepLabV3+ in all accuracy metrics. For MSFE-OCNet, the overall accuracy (OA), Kappa coefficient and mean class accuracy (MCA) reach 92.87%, 90.50% and 91.33% in Nanchang; 93.69%, 91.56% and 92.15% in Jingdezhen; and 93.71%, 91.59% and 92.22% in Yingtan. For the cultivated land category specifically, the intersection over union (IoU) and F₁-score of the proposed algorithm are 90.03% and 94.76% in Nanchang, 92.73% and 96.23% in Jingdezhen, and 92.05% and 95.86% in Yingtan. Experimental results demonstrate that the proposed algorithm effectively improves the identification integrity and classification accuracy of fragmented cultivated land in mountainous and hilly regions.

Key words: cultivated land extraction, object context network, atrous spatial pyramid pooling, combined loss function, mountainous and hilly regions