Agricultural Field Extraction with Deep Learning Algorithm and Satellite Imagery

Automatic detection of borders using remote sensing images will minimize the dependency on time-consuming manual input. The lack of field border data sets indicates that current methods are ineffective. This article seeks to promote the detection of field borders from satellite images with general process based on a multi-task segmentation model. ResUNet-a is a convolutional neural network with a completely linked UNet backbone that supports sprawling and conditional inference. The algorithm will significantly increase model efficiency and its generalization by re-constructing connected outputs. Then individual field segmentation can be accomplished by post-processing model outputs. The model was extremely exact in field mapping, field borders, and thus individual fields using the Sentinel-2 and Landsat-8 images as inputs. The multitemporal images replacement with a single image similar to the composition time decreased slightly. The proposed model is able to reliably identify field borders and remove irrelevant limits from the image to acquire complex hierarchical contextual properties, thus outstriking classical edge filters. Our method is supposed to promote individual crop field extraction on a scale, by minimizing overfitting.

Automated summary

This article introduces a method for detecting field borders in remote sensing images using a multi-task segmentation model. The method utilizes the ResUNet-a convolutional neural network to achieve fast and accurate border detection, and performs field segmentation using post-processing of model outputs. Experimental results demonstrate that the method can effectively identify field borders and produce precise field border and segmentation results.