Time-of-planting mapping method for apple orchards based on standard spectral endmembers spaces

Abstract: A highly efficient and convenient mapping can greatly contribute to access the plantation year in a large-scale orchard using remote sensing. Limited studies on the mapping orchard plantation age can be divided into two categories, namely: 1) using the spectral feature differences, and 2) using the vegetation phenology represented by remote sensing images. However, current studies cannot avoid the influence of mixed image elements on the spectral information of features. Alternatively, the linear hybrid decomposition model can be expected to effectively estimate the orchard plantation age in a large scale. The complex hybrid image can further be decomposed into different pure end elements for the physical information. This study aims to integrate the surface standard end element space using Landsat series images, in order to mapping the orchard plantation information. The following parts were included: 1) The area of apple orchard was firstly mapped to incorporate four standard end elements of substrate (SL, rock and soil), vegetation (GV, photosynthetic foliage), dark matter (DA, shadows), and water (WA, water bodies) into the original image, particularly with the random forest for the land use classification. 2) The Landsat8-OLI, Landsat7-ETM+, and Landsat5-TM sensor images were used to conduct the linear spectral mixture decomposition. Then, the time series curves of vegetation end element were constructed to determine the slow growth interval of apple orchard. The four-point method was applied to explore the maximum environmental carrying capacity of apple orchard in the study area. 3) The starting point of apple orchard plantation was found to fit the logistic growth model for the subsequent mapping of the orchard plantation information. The main findings were as follows. 1) Four end elements from the linear spectral mixture decomposition were used to better represent the surface component information in the orchard. The accuracy of feature extraction was also effectively improved after the fusion of the standard end element space and the random forest. Specifically, the overall accuracy of classification mapping reached up to 88.80% than before, with the Kappa coefficient of 0.86. Besides, there was a better interpretation of the orchard, with the accuracy of 92% than before. 2) An excellent stability was obtained to relatively present the vegetation end element time series curves. Among them, three Landsat series sensor images were used to extract the feature information during operation. The variation of land cover/use was easily used to capture the vegetation end member time series curves. Thus, the Logistic growth model better performed on the biological processes of vegetation growth. The fruit tree growth model was also fitted for the higher accuracy and stability, particularly with the overall fit of 0.751, and the mean error of 1.86 years. The finding can provide a strong reference to determine the plantation information and plantation year of fruit trees with the higher accuracy than before.