CROPUP: Historical products are all you need? An end-to-end cross-year crop map updating framework without the need for in situ samples

In situ samples are essential for crop mapping, but the collection of samples is time-consuming and labor-intensive, and the samples are usually only valid for the current year, due to the crop rotation across years. In this paper, we discuss an alternative solution, i.e., whether using transfer learning to mine useful information from historical products can achieve cross-year crop mapping without the need for in situ samples. However, there are two main challenges that limit the application of historical products: 1) the label mismatch problem, which is caused by the limited accuracy of the historical products and the cross-year changes in crop planting; and 2) the cross-year phenological mismatch problem, where the number and the date of the satellite imagery time series (SITS) are inconsistent across years, hindering the transferability of deep learning models. To address these issues, we propose an end-to-end CRoss-year crOp maP UPdating (CROPUP) framework for crop mapping without the need for any in situ samples. Specifically, to solve the cross-year phenological mismatch problem, an UNequal tIme-series feaTure Extraction (UNITE) network is first introduced to unify the feature dimensions of the SITS of different years, which is then followed by a feature alignment module to align the key cross-year phenological features. In addition, to solve the label mismatch problem, the CROPUP framework introduces a noise-free label estimation loss to reduce the noisy labels in the historical products dynamically during training, which promotes the accuracy of cross-year crop mapping in an iterative manner. The CROPUP framework was verified in the Corn Belt in the U.S. for a long-term and multi-scene analysis and in Jianghan Plain in China for a large-area analysis, using Landsat 8 and Sentinel-2 SITS. The CROPUP framework is highly efficient, and still robust in the case of historical products with a high noisy label ratio. It also shows strength in early-season crop mapping. In addition, the experiments undertaken in this study indicated that the validity period for historical products is within about 5 years, and the accuracy decreases with an increase in time interval. We believe that the CROPUP framework will be a promising and efficient tool to support large-scale crop map updating without the need for in situ samples.