Enhancing the Accuracy and Temporal Transferability of Irrigated Cropping Field Classification Using Optical Remote Sensing Imagery

编号 030033301

推送时间 20220307

研究领域 森林经理 

年份 2022 

类型 期刊 

语种 英语

标题 Enhancing the Accuracy and Temporal Transferability of Irrigated Cropping Field Classification Using Optical Remote Sensing Imagery

来源期刊 REMOTE SENSING

期 第333期

发表时间 20220219

关键词 irrigated field classification;  Landsat;  NDVI time series;  Gaussian mixture model;  interannual climate variability;  regional analysis; 

摘要 Mapping irrigated areas using remotely sensed imagery has been widely applied to support agricultural water management; however, accuracy is often compromised by the in-field heterogeneity of and interannual variability in crop conditions. This paper addresses these key issues. Two classification methods were employed to map irrigated fields using normalized difference vegetation index (NDVI) values derived from Landsat 7 and Landsat 8: a dynamic thresholding method (method one) and a random forest method (method two). To improve the representativeness of field-level NDVI aggregates, which are the key inputs in our methods, a Gaussian mixture model (GMM)-based filtering approach was adopted to remove noncrop pixels (e.g., trees and bare soils) and mixed pixels along the field boundary. To improve the temporal transferability of?method one?we dynamically determined the threshold value to account for the impact of interannual weather variability based on the dynamic range of NDVI values. In?method two?an innovative training sample pool was designed for the random forest modeling to enable automatic calibration for each season, which contributes to consistent performance across years. The irrigated field mapping was applied to a major irrigation district in Australia from 2011 to 2018, for summer and winter cropping seasons separately. The results showed that using GMM-based filtering can markedly improve field-level data quality and avoid up to 1/3 of omission errors for irrigated fields.?Method two?showed superior performance, exhibiting consistent and good accuracy (kappa > 0.9) for both seasons. The classified maps in wet winter seasons should be used with caution, because rainfall alone can largely meet plant water requirements, leaving the contribution of irrigation to the surface spectral signature weak. The approaches introduced are transferable to other areas, can support multiyear irrigated area mapping with high accuracy, and significantly reduced model development effort

服务人员 付贺龙

服务院士 唐守正

PDF文件 浏览全文