Abstract:In order to obtain spatial features distribution from mixed pixels of remote sensing image and further increase accuracy of crop classification and recognition from remote sensing, a double-calculated spatial gravity model (DSGM) based on improvement of spatial attraction model was put forward and applied in research of multispectral images classification and identification in agriculture region at sub-pixel level. Law of gravity was used to describe the spatial correlation and calculate attraction between pixels. Based on the above research, the initialization algorithm of the pixel swapping model (PSM) was improved by spatial attraction model (SAM), and the optimization algorithm of PSM was improved respectively. Finally, all of the models of PSM, SAM and DSGM were applied to the sub-pixel mapping experiments of multispectral images in agricultural region and sub-pixel mapping accuracies of models were compared with each other. The study areas located in typical farming area of Huang-Huai-Hai Plain in North China, and artificial imagery in different degradation scales and GF-1 remote sensing imagery were used as the data sources in the experiment. The final results indicated that (DSGM) model could map effectively at sub-pixel level and its mapping accuracy was superior to those of PSM and SAM. Among them, in artificial image experiment, when sub-pixel degradation scale was 6, overall accuracy and kappa coefficient of DSGM were 93.90% and 0.818, respectively. Compared with K-mean classification, the DSGM model could improve overall accuracy and kappa coefficient by 3.76% and 0.254, respectively. Compared with SAM, DSGM could improve overall accuracy and kappa coefficient by 2.25% and 0.160, respectively. Compared with PSM, DSGM could improve overall accuracy and kappa coefficient by 2.45% and 0.173, respectively. In remote sensing image experiment, when sub-pixel degradation scale was 4, overall accuracy and kappa coefficient of DSGM were 83.13% and 0.742, respectively. Compared with the K-mean classification, DSGM could improve the overall accuracy and kappa coefficient by 9.50% and 0.154, respectively. Compared with SAM, DSGM could improve the overall accuracy and kappa coefficient by 5.44% and 0.088, respectively. Compared with PSM, DSGM could improve the overall accuracy and kappa coefficient by 6.39% and 0.104, respectively. It was seen that DSGM model had feasibility and applicability in sub-pixel mapping, and it could provide a new way to better surpass the limits of remote sensing image spatial resolution. DSGM could further improve accuracy of crop remote sensing classification and recognition and provide strong technical support to obtain accurate information for agricultural remote sensing.
