Mapping of plastic greenhouses and mulching films from very high resolution remote sensing imagery based on a dilated and non-local convolutional neural network

As the important components of modern facility agriculture, both plastic greenhouses and mulching films have been widely utilized in agriculture production. Due to the similarity of spectral signatures, it remains a challenging task to separate plastic greenhouses and mulching films from each other. Meanwhile, deep learning has achieved great performance in many computer vison tasks, and has become a research hotspot in remote sensing image analysis. However, deep learning has been rarely studied for the accurate mapping of agricultural plastic covers, especially for the long-neglected issue of the separation between plastic greenhouses and mulching films. Therefore, this study aims to propose a deep learning model to detect and separate plastic greenhouses and mulching films from very high resolution (VHR) remotely sensed data, providing the agricultural plastic covered maps for relevant decision-makers. In specific, the proposed model is a dilated and non-local convolutional neural network (DNCNN), which consists of several multi-scale dilated convolution blocks and a non-local feature extraction module. The former contains a series of dilated convolutions with various dilated rates, which is to aggregate multi-level spatial features hence to account for the scale variations of land objects. While the latter utilizes a non-local module to extract the global and contextual features to further enhance the inter-class separability. Experimental results from Shenxian, China and Al-Kharj, Saudi Arabia show that the DNCNN in this study obtains a high accuracy with an overall accuracy of 89.6% and 92.6%, respectively. Compared to standard convolution, the inclusion of dilated convolution could raise the classification accuracy by 2.7%. In addition, ablation analysis shows that the non-local feature extraction module could also improve the classification accuracy by about 2%. This study demonstrates that the proposed DNCNN yields an effective approach for the accurate agricultural plastic cover mapping from VHR remotely sensed imagery.