Incorporating optical and polarimetric synthetic-aperture radar (SAR) data to estimate impervious surface is useful but challenging due to their different geometric imaging mechanism and the high diversity of urban land covers. The recent development of deep convolutional networks (DCN) opens a promising opportunity by automatically extracting the deep features from both data sets. In this study, a small-patched DCN (SDCN) was designed to estimate the impervious surface from optical and SAR data. Benchmark methods, e.g., GoogLeNet, VGG16, ResNet50, and the support vector machine were employed for comparison. Two study sites in the most complex metropolitan of China, the Guangdong-Hong Kong-Macau Greater Bay Area, were selected to assess the proposed method. Experimental results indicated the effectiveness of proposed SDCN with a better accuracy outperforming other benchmark methods. Furthermore, we found that 60%–80% of training samples performed comparably with the whole training set, indicating that a large number of training samples may not be necessary in all cases, depending on the settings of some factors (e.g., number of epochs). Generally, SDCN appears more suitable than other methods in terms of combining the optical and SAR data and improved the accuracy of estimating impervious surface.

In recent years, the number of geological disasters in Sichuan Province has significantly increased due to the influence of earthquakes and extreme climate, as well as the disturbance to the geological environment by human activities. Thus, geological disaster monitoring is particularly important, which can provide some scientific basis for disaster prevention and reduction. In this paper, the interferometric synthetic aperture radar (InSAR) technology was introduced to monitor potential geological hazards, taking parts of Dujiangyan City, Wenchuan County, and Mao County in Sichuan Province, China as examples. Firstly, the data such as Sentinel-1A Terrain Observation with Progressive Scans (TOPS) Synthetic Aperture Radar (SAR) images and Precision Orbit Determination (POD) precise orbit ephemerides from 2018 to 2020, high-resolution optical satellite images and Digital Elevation Model (DEM) were collected. Secondly, the Differential InSAR (D-InSAR), Persistent Scatterer InSAR (PS-InSAR), Small Baseline Subset InSAR (SBAS-InSAR), Offset-Tracking, and Distributed Scatterer InSAR (DS-InSAR) algorithms were used to invert the surface deformation of the study area. Thirdly, according to the deformation results obtained by experiments, we used GF-1 and GF-2 optical images as a reference and combine the experimental results of InSAR algorithms to delineate the areas affected by geological disasters. A total of 49 geological disaster areas were obtained, mainly including landslides, collapses, and debris flow. Through field verification, the overall accuracy rate of InSAR deformation monitoring achieved 69.23%, and the accuracy rate of new potential hazards monitoring reached 63.64%. Among all InSAR methods, the DS-InSAR method outperformed and monitored the geological disaster areas well. Finally, the study area was divided into three elevation intervals and the applicability of different InSAR algorithms in different elevation intervals was discussed.

In recent years, the recognition accuracy of semantic segmentation model on natural images can yield a very high level. Thus, it is of great significance to utilize semantic segmentation algorithm to obtain land use classification with remote sensing images. However, due to the large differences between natural images and remote sensing images, the standard semantic segmentation algorithm is not effective for land use classification of remote sensing images. In this paper, the structure of high-resolution network (HR-Net) algorithm is improved according to the difference between the two kinds of images to make it more suitable for remote sensing images. Furthermore, in order to overcome the dependence of the semantic segmentation algorithm on a large number of high-quality prior data sets, some research experiments are conducted with the improved HR-Net domain adaptation model, and both of the adversarial domain adaptation model and the fusion domain adaptation model based on improved HR-Net and CycleGAN are designed to reduce the workload of manually labeling data. The extensive experimental results show that the classification of our improved HR-Net algorithm and the two domain adaptation models outperform other algorithms, that demonstrates the effectiveness and superiority of our algorithms.

Wildfire is a severe natural disaster that poses a significant threat to the natural environment, as well as the safety of human life and property. The timely detection of wildfires plays a critical role in minimizing their detrimental impact. Himawari-8, a geostationary satellite equipped with an advanced Himawari imager (AHI) sensor, can provide full disk data every ten minutes, thus enabling near real-time and large-scale monitoring of wildfires. In this paper, a wildfire detection method based on the spatiotemporal features of Himawari-8 data is proposed. First, a temporal convolutional network (TCN) is employed to predict the brightness temperature of the bands related to wildfire detection, achieving prediction results with MAE of 0.28 K, MSE of 0.30 K ² , and MAPE of 0.10 %. Then, various feature strategies are devised from spectral, spatial, and temporal aspects, and machine learning models are utilized for wildfire detection research. Among the considered strategies, strategy 4, which integrates spectral, spatial and temporal features with the random forest algorithm, exhibits the most effective wildfire detection performance. It achieves a precision of 0.62, an omission of 0.34, and an F1-score of 0.64. Compared with the threshold method, precision increased by 0.05, omission decreased by 0.31, and F1-score increased by 0.21. To further evaluate practical applicability, the combination of strategy 4 and the random forest is employed for wildfire detection near power grid transmission lines. In this scenario, out of the 295 real wildfires, 253 are successfully detected, resulting in a recall of 0.86. These experimental results affirm the effectiveness of the proposed method for wildfire detection.

Remote sensing integrated with deep learning (DL) improves wildfire assessment. The research has been done to scrutinize the areas affected by disastrous wildfires in Yunnan using DL. Wildfire identification and demarcation of the affected area have been limited to primitive thresholding and outdated machine learning classification techniques. Therefore, the research work incorporated DL in the wildfire scrutiny, and several of the most important considerations are investigated. The proposed research objective is to exploit the recent advancement of capsule-based DL together with the wildfire domain. The proposed dense structure provides highly efficient detection and segmentation of the burned area (BA). The BA dense capsule network (BA_EnCaps) is employed to extract and localize the burned zone with an overall accuracy of 98%. The model is evaluated quantitatively using accuracy, binary_cross-entropy, dice_loss, and mean square error (mse). The research aims to utilize the segmentation model to estimate the BA with great results. BA-EnCaps shows excellent accuracy in discriminating the spectral indices for the burned zone. The proposed method surpasses other segmentation benchmark techniques (U-Net, U-Net3p, SegCaps, Deep U-Net, and U-Net+) by substantially lessening the computing power. Finally, BA_EnCaps is compared with standard segmentation techniques and shows that DL-based models can assess wildfire better than conventional algorithms.

Due to an increasing demand of animal husbandry in arid (semiarid) area in China, grassland monitoring based on big data has become a proliferating research focus in recent years. However, the spectral of grass is interfered by topographic relief or that of forest in remote sensing image classification, leading to confusing pixels. In this letter, Tangbula grassland in the middle section of the Tianshan Mountains in Xinjiang (China) was selected as the research area. Upon analysis, we developed a novel composite multifeature deep learning method of capsule network to realize rapid and high-precision remote sensing recognition of the mountainous grassland, through combining the spectral bands with all the extracted features [normalized difference vegetation index (NDVI), topographic, and texture]. Using the new method, the accuracy of the grassland and overall classifications reached the highest values of 91.60% and 96.13%, respectively, greater than those of 85.58% and 92.18%, respectively, of normal classifications without input from texture and topographic features. Compared with the other methods, the method we applied is better than support vector machine (SVM), random forest, and artificial neural network in terms of grassland extraction and classification accuracy.