Model morphing supported large scale crop type mapping: A case stuy of cotton mapping in Xinjiang, China

• Self-attention model was used to learn unique feature knowledge of target crop. • Constructing time-continuous model series using model morphing technique. • The approach shows strong transfer generalization ability without training samples. Long-term, large-scale crop distribution mapping is crucial for agricultural policy and resource management. While high-resolution multispectral remote sensing has been widely used for crop type mapping, three major challenges remain: 1) spatiotemporal heterogeneity in cloud-free and shadow-free observations, 2) the lack of sufficient ground truth samples, and 3) limited generalization of identification models over extended periods. To address these challenges, this paper constructs a time-continuous sequence model that captures the unique feature pattern between the target-crop and non-target crops (referred to as the knowledge model). Specifically, a morphing approach was first employed to interpolate intermediate models between two pre-trained non-adjacent knowledge models. Then, a date-continuous sequence model that estimate the probabilistic of growth patterns of target crop was generated. This date-continuous sequence model mitigates spatiotemporal heterogeneity issues at the pixel level across large regions. Additionally, crop-specific knowledge model addresses sample scarcity and enhances generalization during long-term applications. The method was test using a long-term cotton mapping task (2000, 2005–2023) in Xinjiang, China. The results demonstrate that: 1) The sequence of knowledge model can effectively capture feature differences between cotton and non-cotton throughout the growing period, resulting in knowledge feature has a higher separability compared to original spectral and vegetation index features; 2) Segmenting knowledge features with Unet enables effective mapping cotton and non-cotton without ground samples. The estimated planting area from our mapping results shows excellent consistency with official statistics (R 2 = 0.97). The correlation between our 2018–2021 results and previously published data reached 0.8, 0.88, 0.88, and 0.89. 3). The stable and excellent mapping accuracy proves that resonation of connectivity and reachability in parameter space between two networks with identical architecture, and model morphing is a feasible way to overcome the spatial–temporal heterogeneity in valid observations in large regions.