Scene classification, aiming to identify the land-cover categories of remotely sensed image patches, is now a fundamental task in the remote sensing image analysis field. Deep-learning-model-based algorithms are widely applied in scene classification and achieve remarkable performance, but these high-level methods are computationally expensive and time-consuming. Consequently in this paper, we introduce a knowledge distillation framework, currently a mainstream model compression method, into remote sensing scene classification to improve the performance of smaller and shallower network models. Our knowledge distillation training method makes the high-temperature softmax output of a small and shallow student model match the large and deep teacher model. In our experiments, we evaluate knowledge distillation training method for remote sensing scene classification on four public datasets: AID dataset, UCMerced dataset, NWPU-RESISC dataset, and EuroSAT dataset. Results show that our proposed training method was effective and increased overall accuracy (3% in AID experiments, 5% in UCMerced experiments, 1% in NWPU-RESISC and EuroSAT experiments) for small and shallow models. We further explored the performance of the student model on small and unbalanced datasets. Our findings indicate that knowledge distillation can improve the performance of small network models on datasets with lower spatial resolution images, numerous categories, as well as fewer training samples.

With the successful launch of GaoFen4 (GF4), on-orbit high accuracy geometric calibration for the high-resolution geostationary optical satellite will be a new research topic. With the improvement in the geometric resolution from geostationary orbit, it will become more and more difficult to meet the requirements of both high geometric resolution and large coverage for the available reference data. The purpose of this paper is to explore a new self-calibration mode for GF4 and future high-resolution geostationary optical area array cameras based on the fewest ground control points (GCPs). To overcome the problems of overparameterization, strong correlation and lower significance of the traditional rigorous imaging model, the simplified physical internal model is proposed, and its effectiveness in describing and compensating for the camera internal distortion is verified. Based on the simplified physical internal model, the self-calibration method based on two GCPs and evenly distributed tie points of two images is proposed for the high accuracy estimation of the calibration parameters. The GCPs can be used to provide the absolute geographical constraints for scale information, and the tie points can be used to provide the global constraints for optimum estimation. After calibration, the internal distortion is well compensated, and the positioning accuracy with relatively few GCPs is shown to be better than 1.0 pixels for both the panchromatic and near-infrared sensor and the intermediate infrared sensor. This paper will provide a new usable concept and approach for the future higher resolution geostationary area array optical camera to overcome the stringent requirements of both high resolution and a large area of reference data for the traditional calibration method.

Using deep learning to improve the capabilities of high-resolution satellite images has emerged recently as an important topic in automatic classification. Deep networks track hierarchical high-level features to identify objects; however, enhancing the classification accuracy from low-level features is often disregarded. We therefore proposed a two-stream deep-learning neural network strategy, with a main stream utilizing fine spatial-resolution panchromatic images to retain low-level information under a supervised residual network structure. An auxiliary line employed an unsupervised net to extract high-level abstract and discriminative features from multispectral images to supplement the spectral information in the main stream. Various feature extraction types from the neural network were selected and jointed in the novel net, as the combined high- and low-level features could provide a superior solution to image classification. In traditional convolutional neural networks, increased network depth might not influence the network performance perceptibly; however, we introduced a residual neural network to develop the expressive ability of the deeper net, increasing the role of net depth in feature extraction. To enhance feature robustness, we proposed a novel consolidation part in feature extraction. An adversarial net improved the feature extraction capabilities and aided digging the inherent and discriminative features from data, with increased extraction efficacy. Tests on satellite images indicated the high overall accuracy of our novel net, verifying that net depth or number of convolution kernels affected the classification capability. Various comparative tests proved the structural rationality for our two-stream structure.

A spacecraft-borne optical navigation camera is one of the key instruments for optical autonomous navigation, and the line of sight (LOS) of the camera directly affects the accuracy of navigation. We developed an on-orbit calibration approach for a navigation camera to ensure the accuracy, in which a stepwise calibration is performed, external parameters estimated, and internal parameters estimated in a generalized camera frame determined by external parameters. In addition, we proposed a batch and sequential on-orbit estimation method to save on-orbit computing power, and established a strategy to reject misidentified reference stars while keeping consistency in on-orbit parameters estimation. After the calibration, the accuracy of LOS in inertial frame satisfied the need in optical autonomous navigation. The results have been proven precise and robust in three experiments.

Gaofen-1 02/03/04 satellites, the first civilian high resolution optical operational constellation in China, have Earth observation capabilities with panchromatic/multispectral imaging at 2/8 m resolution. Satellite jitter, the fluctuation of satellite points, has a negative influence on the geometric quality of high-resolution optical satellite imagery. This paper presents an improved jitter detection method based on parallax observation of multispectral sensors for Gaofen-1 02/03/04 satellites, which can eliminate the effect of the relative internal error induced by lens distortion, and accurately estimate the parameters of satellite jitter. The relative internal error is estimated by polynomial modelling and removed from the original parallax image generated by pixel-to-pixel image matching between two bands of images. The accurate relative time-varying error and absolute distortion caused by satellite jitter could be estimated by using the sine function. Three datasets of multispectral images captured by Gaofen-1 02/03/04 satellites were used to conduct the experiments. The results show that the relative system errors in both the across- and along-track directions can be modelled with a quadratic polynomial, and satellite jitter with a frequency of 1.1–1.2 Hz in the across-track direction was detected for the first time. The amplitude of the jitter differed in the three datasets. The largest amplitude, from satellite 04, is 1.3 pixels. The smallest amplitude, from satellite 02, is 0.077 pixels. The reliability and accuracy of the detection results were verified by using two groups of band combinations and ortho-images with a 1 m resolution. The comparison results show that the detection accuracy is improved by approximately 30% using the proposed method.

The linear array push broom imaging mode is widely used for high resolution optical satellites (HROS). Using double-cameras attached by a high-rigidity support along with push broom imaging is one method to enlarge the field of view while ensuring high resolution. High accuracy image mosaicking is the key factor of the geometrical quality of complete stitched satellite imagery. This paper proposes a high accuracy image mosaicking approach based on the big virtual camera (BVC) in the double-camera system on the GaoFen2 optical remote sensing satellite (GF2). A big virtual camera can be built according to the rigorous imaging model of a single camera; then, each single image strip obtained by each TDI-CCD detector can be re-projected to the virtual detector of the big virtual camera coordinate system using forward-projection and backward-projection to obtain the corresponding single virtual image. After an on-orbit calibration and relative orientation, the complete final virtual image can be obtained by stitching the single virtual images together based on their coordinate information on the big virtual detector image plane. The paper subtly uses the concept of the big virtual camera to obtain a stitched image and the corresponding high accuracy rational function model (RFM) for concurrent post processing. Experiments verified that the proposed method can achieve seamless mosaicking while maintaining the geometric accuracy.