ContexNestedU-Net: Efficient Context-Aware Semantic Segmentation Architecture for Precision Agriculture Applications Based on Multispectral Remote Sensing Imagery

Precision agriculture relies on semantic segmentation models to optimize crop yield and minimize environmental impact.ContexNestedU-Net is proposed to improve the capture of contextual information for efficient utilization of multispectral remote sensing images in precision agriculture applications.For this purpose, it includes a novel redesign of the convolutional blocks in the Nested U-Net model.Through the application of depthwise separable convolution in the convolution blocks, the ContexNestedU-Net efficiently preserves unique spectral information.Subsequently, dilated convolution is applied to capture rich contextual information.Three image sets are utilized in the experiments, one from the WorldView-3 satellite and the others from aerial vehicles.Extensive experiments demonstrate that the ContexNestedU-Net outperforms other U-Net-based models for various precision agriculture tasks.When using NDVI images, the proposed architecture improves the Jaccard index by 13% for tree objects, 0.9% for crop objects, and 4.5% for wheat yellow-rust objects compared to Nested U-Net.In addition, the ContexNestedU-Net model reduces the number of trainable parameters from 36.63 to 19 compared to Nested U-Net, and the computational complexity (GLOPs) decreases from 849.3 to 302.4.