Spatiotemporal cube for identifying Cd-dominated heavy metal stress in rice using signal decomposition and hotspot analysis from GF-6 time series images

Crops often experience heavy metal stress alongside other agricultural stresses. However, distinguishing heavy metal stress from other types of stress is difficult because of the similar canopy reflectance responses. The study area is located in the Zhuzhou section of the Xiangjiang River Basin, Hunan Province, China. A total of 39 GaoFen-6 Wide Field View (WFV) images with less than 20% cloud cover were acquired during the rice-growing period (1 April to 31 October) from 2019 to 2022. The GF-6 WFV sensor provides multispectral imagery with 16 m spatial resolution and red-edge spectral bands, which are suitable for constructing the red-edge chlorophyll index. To capture the persistent spectral response of rice to heavy metal stress, this study constructed a CIre time series during the rice growth period and decomposed it using complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) to isolate the relatively stable low-frequency fluctuation component associated with heavy metal stress. The extracted component was further analysed using emerging hot spot analysis (EHSA) to identify spatiotemporal stress patterns, from which a heavy metal stress index (HMSIEC-n) was developed to quantify rice stress intensity. Field-sampled and laboratory-measured soil Cd concentration was used to validate whether the detected rice stress pattern was consistent with the actual contamination level. The HMSIEC-n showed a significantly and positively correlated with measured soil Cd concentration (r = 0.844), and severe heavy metal stress was mainly concentrated along the Xiangjiang River and in the nearby northwestern and central parts of the study area. These findings indicate that the combination of signal decomposition and EHSA can effectively identify and characterize heavy metal stress in rice by capturing the spatiotemporal characteristics of the CIre response. This study will use denser field sampling, more complete time series, and optimized spatiotemporal analysis to improve model generalizability.