A satellite image often contains spatial gaps—missing data (e.g., NoData pixels) that result from issues like cloud masking, sensor errors, or striping. Standard deep learning models cannot inherently process images with these gaps. Therefore, these missing areas must be effectively addressed or "filled" through a process called imputation or gap filling. Several methods are employed for this, ranging from simple techniques that rely on surrounding pixel values to more advanced, sophisticated models. The choice of method is critical, as it directly impacts the quality of the final training data and, consequently, the accuracy of the deep learning model. For the case, the simple gap filling relying on the neighbouring valid pixels will be applied. The exact algorithm is called ‘Navier-stokes’, a prominent inpainting algorithm in computer vision, and applied through OpenCV python library (Bertalmio et al., 2001).
For training the model, only VV and VH bands were selected from Sentinel 1. Although it is possible to train a model with only this much information. It might not give the optimal output. Therefore, we consider adding some features based on the existing radar bands. In the literature Radar vegetation index (RVI) and Radar Backscatter Cross Ratio(BCR) have been proven effective in vegetation monitoring specially in the agricultural use cases(Hu et al., 2024; Vreugdenhil et al., 2018). These features have been calculated and fused with the radar bands using the following formulas.
\[ RVI = \frac{8VH_{linear}}{VV_{linear} + VH_{linear} + e}\]
\[ BCR = \frac{VV_{linear}}{VH_{linear} + e}\]
Where \(VV_{linear}\) and \(VH_{linear}\) stand for the VV and VH backscatters in the linear scale, and \(e\) stands for the error margin.
Along with these remote sensing features, we also add an auxiliary data ‘Day of the year (DOY)’, which means the number of the day in the year when the image was collected, and is also important in crop phenology.