Research on self-collected UAV forage hyperspectral imagery based on three-dimension auto-encoding multi-scale feature extraction

Unmanned Aerial Vehicle (UAV) hyperspectral technology is capable of achieving large-scale, high-efficiency image acquisition, providing more granular information in both spectral and spatial dimensions. This technology overcomes the limitations of traditional methods, which often suffer from low efficiency and limited coverage. The application of UAV hyperspectral technology to agricultural remote sensing significantly enhances the efficiency of crop and plant identification and monitoring, thereby offering robust support for the management and conservation of grassland resources. To this end, we collected low-altitude UAV hyperspectral images of forage fields on site and established a comprehensive hyperspectral dataset for forage. Furthermore, we propose a three-dimensional auto-encoding multi-scale multi-layer convolution network (3DAMCN) for the research on the identification and classification of hyperspectral images. The 3DAMCN calculates weighted features across three dimensions and refines the representation of key features through compression and reconstruction processes. Subsequently, a series of continuous multi-dimensional and multi-scale 2D and 3D convolution structures are utilized to extract image features, fostering a stronger integration of global and local features and augmenting data mining capabilities. We conducted experiments on three datasets and evaluated model performance using six evaluation metrics, including overall accuracy (OA) and computational time. By comparing our model with state-of-the-art deep learning approaches, we verified its effectiveness and feasibility.