Search

Crop classification is a crucial task in agricultural remote sensing, with the accuracy of such classification heavily relies on field sampling. Reusing historical samples can minimize the reliance on annual field sampling for crop classification. Although previous research primarily focused on the classification accuracy based on the reused historical samples, the underlying factors that influence the accuracy have not been adequately investigated. In this study, we employed a two-dimensional convolutional neural network (2D-CNN) model for crop classification reusing historical samples and investigated the factors influencing the classification accuracy. First, we calculated a normalized difference vegetation index (NDVI) time series from historical data to characterize crop growth patterns. Secondly, we assessed three different time-series construction methods. Subsequently, we used the 2D-CNN model to automatically extract abstract features of crop growth patterns. Finally, we designed various strategies for reusing historical samples to explore the influencing factors. Experiments were conducted in Kuitun, Xinjiang Uygur Autonomous Region, China, from 2016 to 2020, employing a long time series of Sentinel-2 images as remote sensing data. Our results indicated that optimal 2D-CNN models using an irregular satellite image time series (irSITS) outperformed random forest models in inter-annual classifications (the overall accuracies for 2016–2020 were 0.78, 0.61, 0.89, 0.89, and 0.70, respectively). In addition, we identified that the primary factors affecting classification accuracy were 1) the time-series construction method used; 2) the crop growth patterns; and 3) the sample diversity. By reusing historical samples and considering the factors influencing classification accuracy, this study provides valuable insights into high-quality crop classification mapping under limited field sample conditions. [ABSTRACT FROM AUTHOR]