Land Use Land Cover Classification with U-Net: Advantages of Combining Sentinel-1 and Sentinel-2 Imagery

Abstract

The U-net is nowadays among the most popular deep learning algorithms for land use/land cover (LULC) mapping; nevertheless, it has rarely been used with synthetic aperture radar (SAR) and multispectral (MS) imagery. On the other hand, the discrimination between plantations and forests in LULC maps has been emphasized, especially for tropical areas, due to their differences in biodiversity and ecosystem services provision. In this study, we trained a U-net using different imagery inputs from Sentinel-1 and Sentinel-2 satellites, MS, SAR and a combination of both (MS + SAR); while a random forests algorithm (RF) with the MS + SAR input was also trained to evaluate the difference in algorithm selection. The classification system included ten classes, including old-growth and secondary forests, as well as old-growth and young plantations. The most accurate results were obtained with the MS + SAR U-net, where the highest overall accuracy (0.76) and average F1-score (0.58) were achieved. Although MS + SAR and MS U-nets gave similar results for almost all of the classes, for old-growth plantations and secondary forest, the addition of the SAR band caused an F1-score increment of 0.08–0.11 (0.62 vs. 0.54 and 0.45 vs. 0.34, respectively). Consecutively, in comparison with the MS + SAR RF, the MS + SAR U-net obtained higher F1-scores for almost all the classes. Our results show that using the U-net with a combined input of SAR and MS images enabled a higher F1-score and accuracy for a detailed LULC map, in comparison with other evaluated methods.