ATM: Adaptive transformer model for reconstruction of remote sensing multi-temporal non-equispaced sequence imagery

Due to the influence of the imaging characteristics of the solar orbiting satellite and atmospheric conditions, the multi-spectral observation data often have the missing of phase image, which brings difficulties to the efficient and automatic processing of regional remote sensing data. However, the observation sequence contain complex and diverse information, it is difficult for the conventional method to characterize this information. To address this issue, we perform an adaptive transformer model (ATM) for reconstruction of remote sensing multi-temporal non-equispaced sequence imagery. To realize efficient information processing by dynamically adjusting the weight distribution of the internal relations of data, the core workflow of ATM is divided into three steps: feature coding, relationship evaluation and adaptive focusing. Firstly, the input data is encoded as a sequence of vectors, each representing a local feature of the data. Then, the model generates an attention score by calculating the similarity between the query and the key, quantifying the strength of the association between the different features. Finally, a context-aware feature representation is formed by using the normalized attention score weighted aggregation vector. Our experiments, conducted utilizing unmanned aerial vehicle (UAV) datasets and the Landsat-8 datasets, demonstrate that the proposed methodology achieves promising performance in terms of both precision (RMSE decreases by about 2.5 points, and SSIM increases by about 0.05 points) and effectiveness (time cost savings of about 50 %), and the proposed model has wide application prospects in intelligent agriculture, water color inversion and vegetation phenology detection.