The first fine-resolution mapping of contour-levee irrigation using deep Bi-Stream convolutional neural networks

Abstract

Agricultural irrigation accounts for nearly 70% of global freshwater withdrawal. Among irrigation practices, contour-levee cascade irrigation is of particular interest as it is water-intensive and widely used in many rice production regions. Despite its significant environmental implications, no study has quantified the distribution of contour-levee irrigation. One major challenge of remote sensing-based contour-levee field detection is how to accurately identify the thin and curved levee lines whose appearance varies dramatically in different fields. This paper presents a new deep network-based method that jointly optimizes semantically meaningful features to quantify the contour-levee fields. This new method uses a bi-stream encoder-decoder architecture to capture spectral information and gradient features. To maintain image gradient sharpness, a skip connection approach is employed to facilitate gradient propagation across long-range connections. Moreover, the new method uses deep supervision to generate more informative features from the earlier hidden layers and superpixel segmentation to reduce classification noise as a post-processing step. By testing against 41 images across 10 Arkansas counties, the average accuracy was 86.23% and the method achieved 15%-17% improvement over benchmark methods. The results show that IrrNet-Bi-Seg maintains good transferability and is thus promising for larger-scale applications.

Keywords

• Remote sensing
• Deep learning
• Irrigation practice
• Agriculture water conservation