Deciphering climate-induced displacement in Somalia: A remote sensing perspective

parameters that needed to be tuned in order to get an optimal result. There were two main hyperparameters that defined the structure of the U-Net, its depth, and its starting features. The depth of the U-Net is defined as the number of skip connections between convolutional encoding/decoding blocks. The depth of the model is 5, as there are 5 total skip connections between encoder/decoder blocks. The other key hyperparameter in the structure of the net- work was the number of starting channels that were used. A depth of 5 was shown to give the highest accuracy and F1 score on the validation set and was able to fit in the GPU memory. For the number of starting channels, as value of 64 was chosen by the same process as before of experimenting with values from 16 to 128, with 64 giving the highest accuracy on the validation. For training and testing the U-Net model, a series of polygons were created using Sentinel- 2 imagery. The polygons, selected during the two annual growing seasons, were devoid of any distortions. The U-Net model was trained on 128x128 images containing the seven input channels. The first approach of gathering training samples, inspired by an approach [40], syn- thetic training images were created from the training polygons. The algorithm to generate the synthetic training images via tessellation worked by randomly selecting a polygon from the training polygon set and then rasterizing it which involved calculating all input features. Each rasterised polygon was sampled randomly weighted by its area, meaning that larger polygons

were more likely to be selected than smaller ones. In total, 1024 128x128 training images and 128 testing images of 128x128 were created. The second approach to generating training images for the U-Net was to select a square ground region and attempt to label each pixel in the image using the Sentinel-2 data as a refer- ence. Each pixel in an image was hand labelled into one of the six classes defined in the classifi- cation schema. After augmentation this led to a total of 985 ground truth images of 128x128 in size, generation in total 2009 training images from both methods available for training. These polygons were then divided with 70% of the polygons being used for training and 30% for testing. The LC Change was calculated based on the detection of changes between two LC maps: a before and after map. The LC change between the maps can then be calculated by performing a simple mapping of the form: LCChange ¼ LCafter þ LCbefore ∗ #classes ð8Þ The number of classes was 6 resulting to a total of 36 possible LC Change classes represent- ing each possible transition of any class to any other class (including conservation). In the case where one or both LC values were unavailable for a given pixel due to the cloud mask, then the LC Change at this pixel was also masked. All possible transitions are seen in Table 4. Analysis and results Drought and floods In total, 48 ADI mosaic were produced monthly with 20 m spatial resolution in GeoTIFF for- mat. Fig 7 shows an example of drought map from August 2017, where the AOI was predomi-

nantly affected by moderate and severe drought, especially in the south near the lower Shebelle River. River floods, caused by water overflowing riverbanks due to prolonged rainfall, are a main cause of environmental migration in Somalia. A flood algorithm was developed using a change detection technique with Sentinel-1 data from before and after a flood. The Digital Elevation Model was used to calculate the slope of the study area. Unlike the monthly drought maps, flood maps were produced based on flood events. Fig 8 displays the overall flood mosaic map. PLOS ONE Deciphering Climate-induced Displacement in Somalia: A Remote Sensing Perspective PLOS ONE | https://doi.org/10.1371/journal.pone.0304202 August 9, 2024 13 / 27