Efficient Flood Segmentation from Sentinel-1 SAR Imagery Using SegFormer and CNN-Based Deep Learning Models

Abstract

Having the good and timely data of flood extent is crucial for disaster management, disaster risk assessment and environmental management. Synthetic Aperture Radar (SAR) imagery is another source of data that can be used to map flooding since it can be acquired under cloudy conditions and does not require daylight. However, the backscatter characteristics of SAR imagery is heterogeneous that makes accurate delineation of flood at pixel level difficult.  A light Weight Vision Transformer - based semantic segmentation architecture, SegFo.rmer, is compara.tively evalu.ated in this stu.dy with four CNN-bas.ed mod.els: U-Net, Weig.hted U-Net, .""Deep.LabV3+ and Atten.tion U-Net for pix.el-lev.elflo.oddeline.ation usi.ng Sent.inel-1 SAR data bas.ed on the SEN1Flo.ods11 bench.mark. A unif.iedexperi.mental prot.ocol was used for data preproc.essing & mod.eltrai.ningtest.ing and perfor.manceevalu.ation". Accu.racy, preci.sion, rec.all, F1-sco.re/Dice coeffi.cient, Inters.ection over Uni.on (IoU), rece.iveropera.tingcharact.eristic area und.er the cur.ve (ROC-AUC), qualit.ative segmen.tation anal.ysis, train.able param.eters, and infer.ence time were used to comp.are the evalu.atedmod.els. The results of SegFormer showed the highest overall accuracy, F1/Dice score, IoU, and ROC-AUC. Attention U-Net had maximum recall (70.17%) and UNET had the lowest measured inference time (6.23 ms/image). The number of trainable parameters in SegFormer was as low as 3.71 million, significantly lower than the CNN architectures evaluated. The results show that SegFormer achieves a good compromise between segmentation quality and parameter efficiency for Sentinel-1 SAR flood segmentation, and the comparative study reveals significant trade-offs between flood sensitivity, segmentation overlap, and computational efficiency.