Impacts of temporal smoothing methods on crop type identification

Abstract: Accurate crop type mapping depends mainly on the reliable acquisition of multi-temporal remote sensing data with high quality. Traditionally, some satellite observations can be easily affected by cloudy and rainy weather, thereby generating undesired images. These low-quality images can normally be excluded or reconstructed using the temporal filters for the remote sensing classification in practice. However, temporal filtering can also eliminate some useful information in the temporal trajectory of original images, leading to the uncertain identification of crop types. Meanwhile, the land cover mixtures within a pixel grid can pose a great influence on the performance of temporal filters, with the decrease in the spatial resolution of images. This study aims to comprehensively investigate the impacts of Savitzky-Golay (S-G) and Harmonic Analysis of Time Series (HANTS) filter on the crop type mapping over different spatial resolutions using the random forest classifier, Harmonized Landsat Sentinel-2 (HLS, 30 m), and Moderate-resolution Imaging Spectroradiometer (MODIS, 500 m) data. Specifically, six indices of time-series vegetation were selected to identify three crop types (soybeans, corn, and rice) in Heilongjiang Province in China. According to the derived maps of crop type, the temporal filtering posed a greater impact on the spatial distribution of crop classification in the decametric spatial resolution image (HLS) than that in MODIS. The evaluation results showed that the overall accuracy of the crop type maps derived by S-G and HANTS was reduced by 1.73 and 5.17 percentage points, compared with the original observations for decametric resolution images. By contrast, the overall accuracies were 84.73%, 85.51%, and 83.05% using the original, S-G, and HANTS observations over the low-resolution images, respectively, indicating no significant change in the crop type mapping before and after temporal filtering. Compared with the time-series vegetation indices over different crop types, both the intra- and inter-class differences of crop types changed significantly after temporal filtering in the decametric resolution images, indicating some great impacts on the accuracy of crop type mapping. Specifically, the inter-class difference between two similar crop types was incorrectly identified as the inherent noise by the temporal filtering. In this case, the temporal filtering had reduced the difference for the low accuracy of crop type mapping. For instance, the user's accuracies of soybeans and corn were reduced by 4.60 and 8.77 percentage points, respectively, before and after the HANTS filter. The temporal filtering had improved the accuracy of crop type identification (e.g., the user's accuracy of rice was improved after HANTS filter), because the significant intra-class difference of each crop type was further reduced in the time-series vegetation indices, compared with the observations without filtering. In terms of crop type mapping on the low-resolution images (i.e., MODIS), the temporal filtering posed no impacts on the accuracy of crop type identification, particularly for the larger effects of land cover mixtures in the pixel grid. Consequently, there were some impacts of temporal filtering on the performance of crop type mapping over different spatial scales. The finding can provide theoretical references and technical support for the spatial distribution of crop types.