基于空–谱融合和深度神经网络的机载高光谱影像作物精细识别

Crop Fine Identification of Crops in Airborne Hyperspectral Images Based on Space-Spectral Fusion and Deep Neural Network

摘要: 针对目前农作物精细识别中由于没能充分利用影像特征，导致识别精度不足的局限，本文提出了一种耦合空–谱融合框架和深度神经网络的机载高光谱遥感影像作物精细识别方法。该方法首先提取作物的GLCM纹理、形态学轮廓、端元丰度三种空间信息，构建决策级融合模型将这三种空间信息与作物的光谱信息进行融合计算。然后引入基于深度神经网络和条件随机场的混合模型进行高精度作物识别。本研究选择了河北雄安有人机数据集进行实验，结果表明，本文方法能融合生成具有互补性的作物特征数据，并能有效减少识别过程中的噪声影响，保持地物边缘，获取高精度的农作物识别信息。

Abstract: Aiming at limitations of insufficient accuracy due to the insufficient utilization of image features in crop fine recognition, this paper proposes a crop fine recognition method for airborne hyperspectral remote sensing images by coupling a space-spectral fusion framework and a deep neural network. This method first extracts three spatial information of crops: GLCM texture, morphological contour, and end element abundance, and constructs a decision-level fusion model to fuse and calculate the three spatial information with the spectral information of crops. Then, a hybrid model based on deep neural networks and conditional random fields is introduced for high-precision crop recogni-tion. This study selected a human-machine dataset from Xiong’an, Hebei for experimentation, and the results showed that the proposed method can fuse and generate complementary crop feature data, effectively reduce the impact of noise during the recognition process, maintain the edge of land features, and obtain high-precision crop recognition information.