A large number of publications have incorporated deep learning in the process of remote sensing change detection. In these Deep Learning Change Detection (DLCD) publications, deep learning methods have demonstrated their superiority over conventional change detection methods. However, the theoretical underpinnings of why deep learning improves the performance of change detection remain unresolved. As of today, few in-depth reviews have investigated the mechanisms of DLCD. Without such a review, five critical questions remain unclear. Does DLCD provide improved information representation for change detection? If so, how? How to select an appropriate DLCD method and why? How much does each type of change benefits from DLCD in terms of its performance? What are the major limitations of existing DLCD methods and what are the prospects for DLCD? To address these five questions, we reviewed according to the following strategies. We grouped the DLCD information assemblages into the four unique dimensions of remote sensing: spectral, spatial, temporal, and multi-sensor. For the extraction of information in each dimension, the difference between DLCD and conventional change detection methods was compared. We proposed a taxonomy of existing DLCD methods by dividing them into two distinctive pools: separate and coupled models. Their advantages, limitations, applicability, and performance were thoroughly investigated and explicitly presented. We examined the variations in performance between DLCD and conventional change detection. We depicted two limitations of DLCD, i.e. training sample and hardware and software dilemmas. Based on these analyses, we identified directions for future developments. As a result of our review, we found that DLCD’s advantages over conventional change detection can be attributed to three factors: improved information representation; improved change detection methods; and performance enhancements. DLCD has to surpass the limitations with regard to training samples and computing infrastructure. We envision this review can boost developments of deep learning in change detection applications.

The accurate location of clouds in images is prerequisite for many high-resolution satellite imagery applications such as atmospheric correction, land cover classifications, and target recognition. Thus, we propose a novel approach for cloud detection using machine learning and multi-feature fusion based on a comparative analysis of typical spectral, textural, and other feature differences between clouds and backgrounds. To validate this method, we tested it on 102 Gao Fen-1(GF-1) and Gao Fen-2(GF-2) satellite images. The overall accuracy of our multi-feature fusion method for cloud detection was more than 91.45%, and the Kappa coefficient for all the tested images was greater than 80%. The producer and user accuracy were also higher at 93.67% and 95.67%, respectively; both of these values were higher than the values for the other tested feature fusion methods. Our results show that this novel multi-feature approach yields better accuracy than other feature fusion methods. In post-processing, we applied an object-oriented method to remove the influence of highly reflective ground objects and further improved the accuracy. Compared to traditional methods, our new method for cloud detection is accurate, exhibits good scalability, and produces consistent results when mapping clouds of different types and sizes over various land surfaces that contain natural vegetation, agriculture land, built-up areas, and water bodies.

Earth observation data obtained from remote sensors must undergo radiometric calibration before use in quantitative applications. However, the large view angles of the panchromatic multispectral sensor (PMS) aboard the GF-4 satellite pose challenges for cross-calibration due to the effects of atmospheric radiation transfer and the bidirectional reflectance distribution function (BRDF). To address this problem, this paper introduces a novel cross-calibration method based on data assimilation considering cross-calibration as an optimal approximation problem. The GF-4 PMS was cross-calibrated with the well-calibrated Landsat-8 Operational Land Imager (OLI) as the reference sensor. In order to correct unequal bidirectional reflection effects, an adjustment factor for the BRDF was established, making complex models unnecessary. The proposed method employed the Shuffled Complex Evolution-University of Arizona (SCE-UA) algorithm to find the optimal calibration coefficients and BRDF adjustment factor through an iterative process. The validation results revealed a surface reflectance error of <5% for the new cross-calibration coefficients. The accuracy of calibration coefficients were significantly improved when compared to the officially published coefficients as well as those derived using conventional methods. The uncertainty produced by the proposed method was less than 7%, meeting the demands for future quantitative applications and research. This method is also applicable to other sensors with large view angles.

Leaf area index (LAI) and fraction of photosynthetically active radiation (FPAR) absorbed by vegetation are two of the essential biophysical variables used in most global models of climate, hydrology, biogeochemistry, and ecology. Most LAI/FPAR products are retrieved from non-geostationary satellite observations. Long revisit times and cloud/cloud shadow contamination lead to temporal and spatial gaps in such LAI/FPAR products. For more effective use in monitoring of vegetation phenology, climate change impacts, disaster trend etc., in a timely manner, it is critical to generate LAI/FPAR with less cloud/cloud shadow contamination and at higher temporal resolution—something that is feasible with geostationary satellite data. In this paper, we estimate the geostationary Himawari-8 Advanced Himawari Imager (AHI) LAI/FPAR fields by training artificial neural networks (ANNs) with Himawari-8 normalized difference vegetation index (NDVI) and moderate resolution imaging spectroradiometer (MODIS) LAI/FPAR products for each biome type. Daily cycles of the estimated AHI LAI/FPAR products indicate that these are stable at 10-min frequency during the day. Comprehensive evaluations were carried out for the different biome types at different spatial and temporal scales by utilizing the MODIS LAI/FPAR products and the available field measurements. These suggest that the generated Himawari-8 AHI LAI/FPAR fields were spatially and temporally consistent with the benchmark MODIS LAI/FPAR products. We also evaluated the AHI LAI/FPAR products for their potential to accurately monitor the vegetation phenology—the results show that AHI LAI/FPAR products closely match the phenological development captured by the MODIS products.

The detection and monitoring of changes in urban buildings, as a major place for human activities, have been considered profoundly in the field of remote sensing. In recent years, comparing with other traditional methods, the deep learning-based methods have become the mainstream methods for urban building change detection due to their strong learning ability and robustness. Unfortunately, often, it is difficult and costly to obtain sufficient samples for the change detection method development. As a result, the application of the deep learning-based building change detection methods is limited in practice. In our work, we proposed a novel multi-task network based on the idea of transfer learning, which is less dependent on change detection samples by appropriately selecting high-dimensional features for sharing and a unique decoding module. Different from other multi-task change detection networks, with the help of a high-accuracy building mask, our network can fully utilize the prior information from building detection branches and further improve the change detection result through the proposed object-level refinement algorithm. To evaluate the proposed method, experiments on the publicly available WHU Building Change Dataset were conducted. The experimental results show that the proposed method achieves F1 values of 0.8939, 0.9037, and 0.9212, respectively, when 10%, 25%, and 50% of change detection training samples are used for network training under the same conditions, thus, outperforming other methods.