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

size (equivalent to 0.25 ha), reducing the false detections arising from noise. The algorithm removes and replaces pixels below the selected threshold with the pixel value of the largest neighbour polygon. Finally, a morphological binary closing was applied to fill holes smaller than the chosen disk radius (r = 5) [39]. Land Cover (LC) change. A LC classification scheme was developed to study the impact of the natural disasters like drought and flood on agricultural food productivity and human activity. This motivated the inclusion of the cropland class, containing food producing arable land as well as food producing orchards. Meanwhile the Settlement class was created to include any apparent permanent human structure. The two principles “background” classes in this product where the Grassland/Shrubland class and the Soil/Sparse vegetation class. The pri- mary differentiator between these two classes is the quantity of vegetation present, with Grass- land/Shrubland dominated by low bushes, and occasional thick grass coverage, whilst the Soil class contains minimal vegetation. The study area is covered mostly with grasslands and shrub- land whilst the majority of the croplands, forest, and settlement regions are distributed along the Shebelle River. The classification scheme can be seen below in Table 3: The entire of four spectral bands of Sentinel-2 with 10m resolution (blue, green, red, near infrared) and two 20m resolution (Short Wave Infrared 1 and 2) were used at the Level-2A in

the training and the evaluation of the model. Additionally, the NDVI was used to enhance the distinction between the cropland and forest classes. A U-Net based model was chosen due to its superior performance in preliminary studies, both in terms of validation metrics and visual results. U-Net architecture is a type of Convolu- tional Neural Network (CNN) with layers that “slide” multiple trained filters to the image, cre- ating several output channels. In U-Net, all convolutions have a 3x3 kernel size, meaning that a pixel and its immediate neighbours are all selected. The CNN implemented using Python 3 programming language. The encoding part of the network used skip connections to form so called ResBlocks, which advantageously allow the input image to propagate deeper into this section of the network, thereby preventing issues such as vanishing gradients from being prominent. The loss function used in the network training was a weighted average of the cross-entropy and dice loss between the two images, with the dice cross comprising 90% of the loss function and the cross-entropy the remaining 10%. The network was trained using the Adam Optimiser combined with a learning rate sched- uler, which gradually reduces the learning rate throughout training by a constant factor of 0.92 per epoch. The network was trained for 100 epochs in total. The U-Net model had various hyperparameters which affected the details of the network architecture and training Table 3. Classification scheme of AOI land cover with six members. ID Class Description 0 Agricultural Fields (green)

Arable land used for crop growing. Food producing orchards are also including in this class. 1 Forest (dark green) Area covered by dense tree coverage, such that the underlying ground coverage cannot be seen. 2 Grassland / Shrubland (brown) Open areas covered by dense non-arable vegetation including predominately from grass and bushes. Class may also include sparse tree coverage. 3 Settlement (grey) Region forming a permanent human settlement. 4 Soil / Sparse vegetation (yellow) Open area covered by bare soil, sand and low-density vegetation. 5 Water (blue) Water bodies including lakes, rivers, ponds, and canals. Also includes regions of flooding covered by ground water. https://doi.org/10.1371/journal.pone.0304202.t003 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 12 / 27