Recognition of grape growing areas in multispectral images based on band enhanced DeepLabv3+

Accurate and rapid identification has been of great importance to obtain the spatial distribution of grape growing areas in recent years. The spatial information can be used to guide the fine management of planting areas and high-quality base construction in an orchard. Recognition of large-scale crop growing areas has been usually implemented using remote sensing images. However, the low accuracy of planting area recognition can be induced by the dispersed patches and the complex background. Fortunately, the convolution operation in deep learning can effectively extract the texture features of images. Among them, semantic segmentation has been one of the most important processing for remote sensing images. In this study, an improved band enhancement DeepLabv3+ (BIE-DeepLabv3+) was proposed for the multispectral image recognition of grape planting areas. An encoder-decoder structure was employed in the DeepLabv3+. Atrous convolution was then applied to encode the multi-scale contextual information in the encoder module. The decoder module was used to effectively capture the sharp object boundaries for the gradual recovery of spatial information. As such, the DeepLabv3+ model was used to require the key features suitable for the high recognition accuracy of grape growing areas with different area sizes and scattered spatial locations. Since the combination of various bands reflected the differences between features, the DeepLabv3+ model was modified to concurrently handle four bands of remote sensing images. In addition, the band enhancement module was also built to determine the interdependencies between the band channel maps. All spectral bands of features were weighted to clarify the semantic dependency relationship among the spectral band feature maps. The ground features were distinguished to fully utilize the spectral information in each band. The dataset was generated through labeling the grape growing areas in the GaoFen-2 remote sensing images taken in 2016 and 2019. Then the model is trained on this dataset. The testing was also performed to verify the improved model using the dataset from the remote sensing images taken in 2020. Experimental results show that the improved model achieved the best classification accuracy, where the mean pixel accuracy and mean intersection over union were 98.58% and 90.27%, respectively. The recognition performance of the improved model was much better than that of the SVM algorithm. Specifically, the mean pixel accuracy and mean intersection over union were improved by 0.38 and 2.01 percentage points on the basis of DeepLabv3+, and the improvement over SegNet were 0.71 and 4.65 percentage points, respectively. More complete grape growing regions were predicted for better edge recognition with the reduced model parameters. Therefore, the BIE-Deeplabv3+ network model can be used to achieve the high accurate segmentation of grape growing areas. The detailed information of the image can also be collected for the spatial correlation of pixels in a large range, particularly for the various planting area sizes and regional dispersion in the grape growing areas. The multi-bands input and BIE module can be used to fully utilize the band information for highlighting differences between the objects. The recognition accuracy of images was significantly improved with similar texture features. Anyway, the effective recognition can be widely expected for the grape planting areas in the remote sensing image with the complex background in a large area.