基于条纹池化和迁移学习的耕地分割模型

耕地的数量和质量是保持农业可持续发展的关键，是政府部门的决策依据。目前这些信息的获取主要依靠人力，不仅浪费大量人力、财力，而且效率较低，不准确，因此利用卫星遥感影像识别分割耕地图像具有重要价值。本文提出了一种耕地图像分割神经网络SP-Vnet，其包含了条纹池化模块和空洞卷积的V型分割模型，并与迁移学习、图像形态学方法等结合，实现了卫星遥感图像中耕地的精确分割和提取。与目前六个主流的语义分割网络模型相比，本文提出的SP-Vnet在最近MathorCup遥感图像耕地分割的竞赛数据集上，取得了更高的整体准确率OA、F1值和平均交并比（mIoU）。实验表明，SP-Vnet能够加强网络的全局特征表征能力，显著提高了耕地识别的准确率，同时结合图像形态学方法的后处理操作，提升了耕地分割边缘的平滑性和准确性。

Cultivated Land Segmentation Model Based on Strip Pooling and Transfer learning

The quantity and quality of cultivated land is the key to maintain the sustainable development of agriculture. Currently these information are obtained manually, which not only costs a lot of manpower and financial resources, but also has low efficiency and accuracy. Thus, it is much valuable to use satellite remote sensing images to identify and extract cultivated land. In this paper, an image segmentation network named SP-VNET is proposed, which contains Strip Pooling module and V-type segmentation model of void convolution while the transfer learning and image morphology processing are also applied. In this way, the cultivated land of satellite remote sensing image can be recognized and segmented more accurately. Compared with the six major semantic segmentation networks, the proposed SP-Vnet obtains the higher segmentation accuracy of OA, F1, and mean crossover ratio (mIoU) on the Mathorcup remote sensing image competition dataset for block recognition and segmentation. Experimental results show that SP-Vnet can enhance the global ability of the network to extract features, improve the segmentation ability, and significantly improve the accuracy of the cultivated land recognition. By combining with the post-processing operation of the traditional image processing method, our network obtains the smoother and the more recognition accuracy of the cultivated land edges.