STFMCNN: A Novel Convolutional Neural Network for Spatiotemporal Fusion of Remote Sensing Images

STFMCNN: A Novel Convolutional Neural Network for Spatiotemporal Fusion of Remote Sensing Images

Spatiotemporal fusion (STF) is a crucial technique for generating high-resolution images from multiple low-resolution images acquired at different times. This paper presents STFMCNN, a novel convolutional neural network (CNN) specifically designed for STF in remote sensing applications.

Comparison with Existing Methods

STFMCNN is compared with three popular STF methods:

• STARFM: A well-established weight function-based STF method widely used as a benchmark.* FSDAF: A hybrid STF method known for its excellent performance in reconstructing abrupt changes.* StfNet: A state-of-the-art learning-based STF method utilizing CNNs.

Experimental Setup

To ensure fair comparison, the default parameters of STARFM, FSDAF, and StfNet were used. StfNet used the same training and validation datasets as STFMCNN, and the predicted FR images were linearly fused using a global function.

Evaluation Metrics

The accuracy of predicted FR images was assessed using four metrics:

• RMSE (Root Mean Squared Error): Measures the difference between the predicted and ground truth images. A lower RMSE indicates higher accuracy.* SSIM (Structural Similarity Index): Measures the similarity in structural information between the predicted and ground truth images. A higher SSIM indicates higher accuracy.* CC (Correlation Coefficient): Measures the linear correlation between the predicted and ground truth images. A higher CC indicates higher accuracy.* AD (Average Difference): Measures the average difference between the predicted and ground truth images. A positive AD indicates overestimation, while a negative AD indicates underestimation.

Results and Discussion

STFMCNN outperformed all other methods in terms of RMSE, SSIM, CC, and AD. This indicates that STFMCNN is particularly effective in reconstructing visible abrupt changes and achieving high accuracy in predicted FR images.

Conclusion

STFMCNN offers a significant improvement over existing STF methods, demonstrating the potential of CNNs for accurate and efficient spatiotemporal fusion of remote sensing images. The proposed method holds great promise for various applications, including environmental monitoring, urban planning, and disaster management.