Using the optical camera in remote sensing is limited in various environmental conditions. This paper presents a system of combining deep learning and image transform algorithms to detect landslide location in satellite images. In the deep learning part, a convolution neural network is used to classify satellite images contain landslides. From landslide images classified, in order to accurately identify landslides under different lighting conditions, this paper proposes a transformation algorithm Hue – Bi-dimensional empirical mode decomposition (H-BEMD) to determine the landslide region and size. After the location of landslide is detected, we discover the size change of the landslide based on different time points. In this study, we record an accuracy of up to 96% in the classification process, and the accuracy of landslide location almost absolute.

Edge computing enables rapid data processing and decision-making on satellite payloads. Deploying deep learning-based techniques for low-light image enhancement improves early detection and tracking accuracy on satellite platforms, but it faces challenges due to limited computational resources. This paper proposes an edge-computing-enabled inference model specifically designed onboard satellites. The proposed model follows an encoder-decoder architecture to generate the illumination map with low multiplication matrix complexity, 25.52 GMac of 1920 X 1200 image size. To reduce nanosatellite hardware consumption with a single-precision floating-point format, the edge-computing-enabled inference model proposes a quantized convolution that computes signed values. The proposed inference model is deployed on Arm Cortex-M3 microcontrollers onboard satellite payload (86.74 times faster than normal convolution model) but also has a similar quality with the low-light enhanced in full-precision computing of lightweight training model by using the peak signal-to-noise ratio (PSNR, average of 28.94) and Structural Similarity Index (SSIM, average of 0.85) metrics.

Detecting tiny objects in remote sensing images presents a significant challenge, especially moving vehicles or ships smaller than $32 \times 32$ pixels. This paper uses Space-to-depth Convolution (SPD-Conv), an architecture design to increase feature extraction of tiny objects. Additionally, the proposed network adopts four detection heads to enhance the tiny object bounding boxes. Compared to the baseline model the proposed network increases by 14.3% of Average Precision at 50% intersection over union (AP50).

Bi-dimensional empirical mode decomposition (BEMD) transforms 2D data into some intrinsic mode functions (IMFs). In this paper, BEMD is used in image preprocessing to determine features of landslides. BEMD omits noisy areas of a landslide image, correctly identifying features of the landslide to conduct training through the Convolutional Neural Networks (CNN) architecture. In satellite image recognition technology, optimization of data transferring between the ground station and the satellite is important. Hence, using BEMD at the pre-training stage of CNN (identifying landslide identification area) reduces the size of the trained model up to 3 times compared to applying the normal CNN architecture with the original images.

This paper proposes an edge computing-based video coding implementation on an Earth observation satellite (SAT-video coding), which can encode video using limited resources and the power of mini/microsatellites. SAT-video coding proposes a hardware-related quantization (Q) function, hardware reduction of the motion estimation (ME) method, and simplified entropy coding (EC), which reduces the computation complexity. The hardware-related Q reduces hardware resource and power consumption by 72% and 55%, respectively, compared with traditional Q implementation. The hardware reduction of ME reduces resource use compared with regular ME implementation (59% of lookup tables [LUTs] and 79% of Registers). The total number of LUTs used for the simplified EC function is also much lower than other EC hardware implementations. The SAT-video encoder IP uses fewer hardware resources, and the power consumption is estimated at 0.0894 W at a high working frequency (125 MHz). The SAT-video encoding speed is 18.95 frames per second for $2560\times 2560$ video. Therefore, the proposed SAT-video coding is an edge computation suitable for micro/minisatellites. The coding efficiency records the highest compression ratio at 33.8, with a peak signal-to-noise ratio of 34.46 dB. With the important task of designing edge computing based on satellite video encoding, these are adequate values for remote sensing video.

This paper proposes a deep learning approach for small surface water recognition (SWRNet) using multispectral satellite imaging, which reduces the computational complexity by 18.66 times and increases the accuracy of surface water recognition by up to 14.1%. The proposed model uses near infrared (NIR) combined with RGB spectral imagery to increase the accuracy of surface water recognition. In addition, since surface water only accounts for a small percentage of the remote sensing dataset, thus creating an imbalance problem, a proposed loss function is introduced to combine region-based and distribution-based loss. This paper introduces an adaptive factor that automatically adjusts the weighting between distribution- and region-based loss functions. The proposed adaptive factor is determined based on the loss value of the previous training step. The mean intersection over union (mIoU) of surface water between predicted and ground truth regions is recorded as 0.80.