A comparative study on semantic segmentation-orientated deep convolutional networks for remote sensing image-based farmland classification:A case study of the Hetao irrigation district
In the management of modern agriculture production,the spatial distribution of different crop types is identified as important information about agricultural conditions.Identifying crop types from satellite remote sensing imagery serves as a fundamental method for acquiring such information.Although there exist various algorithms for identifying surface features from remote sensing imagery,reliable farmland classification remains challenging.This study selected three representative semantic segmentation-orientated deep convolutional models,i.e.,UNet,ResUNet,and SegNext,and compared their performance in crop classification using remote sensing images of the Hetao irrigation district from the Gaofen-2 satellite.Using the three algorithms,nine network models with varying complexities were developed to analyze the differences in the performance of various network structures in classifying crops in farmland based on remote sensing imagery,thus providing optimization insights and an experimental basis for future research on relevant models.Experimental results indicate that the six-layer UNet achieved the highest identification accuracy(88.74％),while the six-layer SegNext yielded the lowest accuracy(84.33％).The ResUNet displayed the highest complexity but serious over-fitting with the dataset used in this study.Regarding computational efficiency,ResUNet was significantly less efficient than the other two model types.
deep convolutionsemantic segmentationcrop filed classificationHetao irrigation district
万方数据
