基于卷积神经网络的高分遥感影像耕地提取研究

高效精准地提取遥感影像中的耕地对农业资源监测以及可持续发展具有重要意义，针对目前多数传统全卷积神经网络（FCN）模型在提取耕地时存在重精度而轻效率的缺陷，本文建立基于FCN的轻量级耕地图斑提取模型（LWIBNet模型），并结合数学形态学算法进行后处理，开展耕地图斑信息的自动化提取研究。该LWIBNet模型汲取了轻量级卷积神经网络和U-Net模型的优点，以Inv-Bottleneck模块（由深度可分离卷积、压缩-激励块和反残差块组成）为核心，采用高效的编码-解码结构为骨架，将LWIBNet模型分别与传统模型的耕地提取效果、经典FCN模型的轻量性和精确度进行对比，结果表明，LWIBNet模型比表现最优的传统模型Kappa系数提高12.0%，比U-Net模型的参数量、计算量、训练耗时、分割耗时分别降低96.5%、87.1%、78.2%和75%，且LWIBNet的分割精度与经典FCN模型相似。

Cultivated Land Extraction from High Resolution Remote Sensing Image Based on Convolutional Neural Network

It is of great significance for agricultural resources monitoring to accurately extract cultivated land map information from remote sensing images.To improve the defects of traditional models for extracting cultivated land and solve the problem that most FCN model pays more attention to accuracy but ignores the consumption of time and computing resources, a lightweight model for extracting cultivated land map spots was established based on FCN (LWIBNet), and post-processing combined with mathematical morphology algorithm were used to carry out automatic extraction of cultivated land information. LWIBNet drew on the advantages of lightweight convolutional neural network and U-Net model, and it was built with the core of Inv-Bottleneck (composed of deep separable convolution, compression-excitation block and inverse residual block) and the skeleton of efficient coding-decoding structure. Compared LWIBNet with the cultivated land extraction effect of traditional model, and the computational resources and time consumption of classical FCN model.The results showed that LWIBNet was 12.0% higher than the Kappa coefficient of the best traditional model, and compared with U-Net, LWIBNet had 96.5%, 87.1%,78.2% and 75% less parameters, calculation, training time and split time-consuming, respectively. Moreover, the segmentation accuracy of LWIBNet was similar to that of the classical FCN model.