Workflow from Scientific Research

Schematic representation of ForestViT model for deforestation detection. The ForestViT model is inspired by the vision transformer idea from [9] and the encoder part of the NLP Transformer. The standard Transformer receives a 1-D token embedding sequence as input. Here, the images are split into fixed-sized patches and fed into the model. A learnable positional embedding vector is assigned to every patch to utilize the order of the input sequence. The ForestViT model assigns the existing classes for each output. We highlight that each image patch could be assigned in more than one class.

DOI