Comparing 2D and 3D convolutional neural networks for land cover fraction and change mapping

By Lennart Murk

Abstract
Mapping land cover fractions and their changes is essential for policy-making, resource management, disaster response, deforestation and climate change monitoring. Land cover fractions are a closer representation of reality as they can capture multiple classes in a single pixel, and allow for customizable maps. Convolutional Neural Networks have been widely used for land cover classification studies; however, their capability for land cover fraction and change mapping remains underexplored. A 2D CNN, 2D CNN-RNN and a 3D CNN were compared for land cover fraction and change mapping. Sentinel-2 imagery with 10 bands, and fractional land cover data at 20 meter resolution were used as input data for the models. The 2D CNN was trained on yearly time series, while the 3D CNN was trained on yearly and monthly time series. All CNN architectures incorporate a ResNet-50-like encoder and a U-Net-like decoder. The RMSE and MAE for land cover mapping indicate that the monthly 3D CNN (28.6, 17.7) achieves the lowest error compared to the other models. The overall metrics for changed locations show that the monthly 3D CNN has the lowest RMSE (27.9) for change. Additionally, the per-bin metrics prove that the monthly 3D CNN again performs best for mapping change in land cover fractions when "no change" is not considered. The results suggest that including temporal information in predictions has a positive effect on mapping change. The monthly 3D CNN performed best for land cover fraction and change mapping. Future research should explore different CNN encoders, decoders, and loss functions. Additionally, future research should explore the use of a hybrid CNN-transformer, and pure transformer architectures.