Abstract:Land-use/land-cover (LULC) classification of high spatial resolution (HSR) remote sensing imagery has been successfully improved using deep learning techniques. However, the current deep learning-based classification methods necessitate the division of remote sensing imagery into smaller and fixed image patches, primarily due to computational constraints arising from the extensive size of these images. This approach limits the receptive field of the classification network and hinders the handling of different-scale LULC objects. A key problem is how to automatically select the appropriate scale of patch for different objects with a deep learning network. To address this challenge, a scale-aware classification network (SAN) based on deep reinforcement learning (DRL) is proposed. In SAN, the state of each image patch is represented by a reduced-resolution version of the high-spatial-resolution (HSR) remote sensing image, referred to as a 'thumbnail', and a positional encoding. The scale selection actions are performed by a scale control agent. A feature indexing module is also proposed to enhance the ability of the agent to distinguish the location of the current image patch. The action switches the patch scale and the viewing area of context branch of a two-branch classification network, which extracts and fuses the features of the multi-scale images. The SAN framework adjusts the network parameters to perform the appropriate scale selection action based on the mapping reward received for the selected scale. In this way, the SAN framework is able to introduce more appropriate contexts by adjusting the scale of the network input based on RL, without the need for labeled scale selection samples. The experimental results obtained using two publicly available datasets and a newly built dataset demonstrate that SAN outperforms the previous LULC deep learning methods with fixed patches, particularly for large-scale mapping applications. When compared to state-of-the-art approaches such as GLNet and WiCoNet, which combine global and local information for segmentation, as well as CascadePSP and MagNet, renowned for their progressive segmentation capabilities, SAN consistently demonstrates approximately 10% higher accuracy. The codes for this research are openly available at http://rsidea.whu.edu.cn/resource_sharing.htm.

Scale-aware Deep Reinforcement Learning for High Resolution Remote Sensing Imagery Classification

A lightweight and stochastic depth residual attention network for remote sensing scene classification

Superpixel-Based Long-Range Dependent Network for High-Resolution Remote-Sensing Image Classification

SAN: Scale-Aware Network for Semantic Segmentation of High-Resolution Aerial Images

Single Remote Sensing Image Super-Resolution Via a Generative Adversarial Network with Stratified Dense Sampling and Chain Training

Unbalanced Class Learning Network With Scale-Adaptive Perception for Complicated Scene in Remote Sensing Images Segmentation

Global to Local: A Scale-Aware Network for Remote Sensing Object Detection.

A Stage-Adaptive Selective Network with Position Awareness for Semantic Segmentation of LULC Remote Sensing Images

Seeing Beyond the Patch: Scale-Adaptive Semantic Segmentation of High-resolution Remote Sensing Imagery Based on Reinforcement Learning

High-Resolution Remote Sensing Image Semantic Segmentation via Multiscale Context and Linear Self-Attention

Scale-Aware Neural Network for Semantic Segmentation of Multi-Resolution Remote Sensing Images

SSN: Scale Selection Network for Multi-Scale Object Detection in Remote Sensing Images

Cascaded Deep Learning Model for Accurate Land Use and Land Cover Classification

Adaptive Spatial-Scale-Aware Deep Convolutional Neural Network for High-Resolution Remote Sensing Imagery Scene Classification

Classification of Very-High-Spatial-Resolution Aerial Images Based on Multiscale Features with Limited Semantic Information

A Multiscale Attention Network for Remote Sensing Scene Images Classification

SFSANet: Multiscale Object Detection in Remote Sensing Image Based on Semantic Fusion and Scale Adaptability

Scale Aware Adaptation for Land-Cover Classification in Remote Sensing Imagery

CSANet: A Channel-Spatial Attention Network for Remote Sensing Image Change Detection

Super Resolution Guided Deep Network for Land Cover Classification From Remote Sensing Images

Remote Sensing Image Land Cover Classification with Label Noise Based on Deep Reinforcement Learning