Multi-feature fusion for cropland extraction in complex scenes

In the current complex and dynamic agricultural landscape, cropland fragmentation and the phenomena of non-agricultural and non-grain land use are widespread, posing significant challenges to the accurate extraction of cropland. To address issues such as unclear cropland boundaries and inaccurate plot extraction in complex scenarios, this paper proposes a Multi-Feature Progressive Fusion Unet (MPFUnet) for cropland segmentation in remote sensing images. This method fully leverages both spatial information and geometric edge information of cropland, using the Unet as the backbone network for spatial feature extraction and a multi-feature attention module is proposed to make up for the Unet network's lack of local subtle feature perception, which includes a spatial attention module and an edge feature enhancement module. The spatial attention module obtains global features by concatenating average pooling maps and maximum pooling maps across channel dimension firstly, and then obtains the spatial attention map through the activation function, which reflects the spatial importance of each pixel. While the edge feature enhancement module improves the perception of multi-scale spatial information by fusing multiple sets of receptive field features under different dilation rates. Based on them, a hierarchical heterogeneous fusion strategy is implemented by multiplying the attention map and multi-scale feature maps to better learn multi-dimensional feature map representation. Thus, the spatial attention features and multi-scale edge information are aggregated at different resolution scales to obtain edge-enhanced spatial feature maps. In order to dynamically adapt to different feature scales, a progressive feature enhancement (PFE) structure is designed in the spatial decoding part of the network, it embeds the spatial information of adjacent layers in each layer to further integrate the global semantic information of high-level features and detailed edge information of low-level features. Furthermore, A layer-by-layer integration approach is adopted to capture and fuse adjacent scale features, which follows the order from the low layer to the high layer of the network to maintain attention consistency between different feature extraction layers. The experiments used multi-source satellite images such as JL-1, GF-1, GF-2, and GF-7 as data sources. The 2-meter high-resolution domestic cropland dataset of Dongpo District, Meishan City, was randomly divided into training, validation, and test sets in a ratio of 3:1:1. The experimental results show that the MPFUnet achieved an accuracy of 92.54%, a recall rate of 94.08%, an average intersection over union (IoU) of 84.32%, and a Kappa coefficient of 87.47%. Compared to the baseline Unet model, these metrics were improved by 8.23%, 7.01%, 10.02%, and 11.33%, respectively. The results were also superior to other methods such as DeepLabv3+, YOLOv3, and Swin-Unet. All of the quantitative experimental outcomes and visual results manifest that the proposed model exhibits excellent performance on regular patches, fragment patches, and complex scene patches. Moreover, the model is capable of effectively integrating spatial texture characteristics and edge detail features, and thus can actively address crop segmentation tasks within diverse scenarios. Additionally, it possesses a robust capacity to accurately identify plot boundaries and small areas. Furthermore, it also demonstrates a high degree of robustness against interference factors in certain complex scenarios. Therefore, it presents an effective and viable solution for farmland exploration in complex scenarios.