Spatial Cross Attention Based Lightweight Deep Learning Model for Enhanced Tree Classification in Satellite Images

Résumé fourni par la source

Satellite images contain rich information that is widely used in forest monitoring, landslide detection, and precision agriculture. However, processing of these highdimensional data requires a robust deep learning model. In this study, we design a lightweight deep learning model with spatial cross attention (SCA) to improve the tree and no-tree classification from the Sentinel-2A dataset. Our model has three convolution layers, followed by the ReLU activation and max pooling. The classical deep learning model extracts shallow, high-dimensional features and may miss the edge and boundary region features. We applied the SCA module after the third convolution block to overcome the issue. The SCA provided attention to the feature map obtained from the convolution block to enhance the focus on the complex region of the satellite image. Further, experimental results are compared with the VGG19, Inception V3 and MobileViT to prove the model's effectiveness. Our model showed improvement in precision of$\mathbf{6. 4 6} \boldsymbol{\%}, \mathbf{4. 2 9 \%}$, and$\mathbf{2. 2 1 \%}$compared to its counterparts.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.