Investigating the Impact of a Low-Rank Tensor-Based Approach on Deforestation Imagery

In this work we handle deforestation as a semantic segmentation problem using convolution-free methods, specifically tensor-based neural networks. The methodology follows a two-step process: first, we enhance the identification of deforestation areas through the application of low-level filters. Second, we merge the filter outcomes with the original image to create an object-level tensor representation. We introduce a tensor-based model that effectively identifies and analyzes complex patterns even with limited data availability. To evaluate our model, its outcomes are compared with established state-of-the-art models. The experimental results demonstrate the effectiveness and efficiency of the proposed approach in accurately mapping land cover patterns, particularly in deforestation detection and analysis. Based on the results, Squeezed-Net perform better when more data are available, while Tensor neural network model (TNN) outperformed when limited data are available.