Deep learning techniques are widely used for land cover classification in remote sensing, primarily because they can effectively extract complex features from imagery data, which is essential for accurate land cover classification. However, the heterogeneity of data can pose a challenge to the generalizability of deep models. To address this issue, we propose a novel transfer learning framework for land cover classification using teacher-student structure. The proposed framework utilizes the knowledge acquired by the teacher model from large datasets to facilitate fine-tuning of the student model on small datasets, which prevents problems such as overfitting resulting from training large models on small datasets and inferior performance arising from the limited learning capacity of small models. To achieve this goal, we design a loss function base on central moment discrepancy and high-temperature softmax. In addition, we conducted experiments on three distinct pairs of datasets and found that our proposed framework outperforms both training the student model from scratch on the target domain and simply fine-tuning the teacher and student models from the source domain to the target domain. Specifically, we observed an average increase in mIoU of 9.9%, 2.1%, and 4.3%. These results demonstrate the effectiveness and generalizability of our proposed framework for remote sensing land cover classification.

Stereo matching and semantic segmentation are two significant tasks in remote sensing. Recently, deep learning approaches have been applied to these tasks separately. However, the lack of semantic supervision makes the training of stereo matching model susceptible to data disturbance, resulting in inferior generalization ability; foreground objects are sometimes confused with background pixels in RGB images, limiting the classification accuracy. By exploring the relationship between these two tasks, semantic stereo solves these problems simultaneously with multitask learning. Previous methods took semantic stereo as two parallel processing tasks, so they did not take full advantages of the additional information from both tasks and only obtained slight improvement. In this work, we designed a multitask learning framework semantic stereo network (S2Net). The proposed network generates cost volumes with feature maps supervised by semantic information to estimate disparity maps and fuses RGB-D feature maps to predict classification maps, therefore gathering multitask learning information. To enhance the performance of trained model, we also considered the continuity of disparity values and the duality of stereo image pairs in data augmentation. When applied in datasets without training, S2Net obtained 2.937% D1-Error in the WHU dataset, lower than 4.297% of the previous best method, depicting the generalization ability improvement from semantic supervision. In terms of semantic segmentation, the introduction of disparity maps increases the mean intersection over union (mIoU) from 61.375% to 69.096% in the US3D datasets. The experiments on the KITTI semantics benchmark show that our proposed method obtains 60.76% mIoU, achieving state-of-the-art among multitask learning methods.

Semantic segmentation is one of the fundamental tasks of pixel-level remote sensing image analysis. Currently, most high-performance semantic segmentation methods are trained in a supervised learning manner. These methods require a large number of image labels as support, but manual annotations are difficult to obtain. To address the problem, we propose an efficient unsupervised remote sensing image segmentation method based on superpixel segmentation and fully convolutional networks (FCNs) in this letter. Our method can achieve pixel-level images segmentation of various scales rapidly without any manual labels or prior knowledge. We use the superpixel segmentation results as synthetic ground truth to guide the gradient descent direction during FCN training. In experiments, our method achieved high performance compared to current unsupervised image segmentation methods on three public datasets. Specifically, our method achieves an Adjusted Mutual Information (AMI) score of 0.2955 on the Gaofen Image Dataset (GID) dataset, while processing each image of size 7200 × 6800 pixels in just 30 seconds.

Object-based image classification (OBIC) on very-high-resolution (VHR) remote sensing (RS) images is utilized in a wide range of applications. Nowadays, many existing OBIC methods only focus on features of each object itself, neglecting the contextual information among adjacent objects and resulting in low classification accuracy. Inspired by a spectral graph theory, we construct a graph structure from objects generated from VHR RS images and propose an OBIC framework based on truncated sparse singular value decomposition and graph convolutional network (GCN) model, aiming to make full use of relativities among objects and produce an accurate classification. Through conducting experiments on two annotated RS image data sets, our framework obtained 97.2% and 66.9% overall accuracy, respectively, in automatic and manual object segmentation circumstances, within a processing time of about 1/100 of convolutional neural network (CNN)-based methods’ training time.

Object detection has attracted a lot of attention in the field of image automatic interpretation. Detectors based on convolution neural networks (CNNs) applied in natural scene images encode detection results with horizontal bounding boxes (HBBs), which can not accurately calibrate the position and shape of the arbitrary-orientation objects on remote sensing images (RSIs). To solve these issues, we propose an object detection framework named multi-oriented rotation-equivariant network (MORE-Net) in this letter. The MORE-Net consists of a multi-oriented rotating filter (MORF) and a multi-oriented ground objects detector. The MORF generates rotation-equivariant features by rotating canonical filter according to the predefined discrete directions. Equipped with MORF, the multi-oriented ground objects detector extracts rotation-invariant semantic representations for each decoded rotated region of interest (RRoI). We also propose area smooth (AS) L1 loss to impose tighter shape and position constraints to the RROIs. Extensive experiments and comprehensive evaluations on the large-scale DOTA dataset demonstrate the effectiveness of the proposed framework, which achieves a mean average precision (mAP) value of 81.27 on the DOTA-v1.0 dataset and a mAP value of 78.03 on the DOTA-v1.5 dataset.