Design and implementation of CNN-based Custom Net Architecture with improved inference time for Realtime Remote Sensing Application

Authors:

Abstract
Convolutional Neural Network(CNN) is popular methodology for image classification in remote sensing. However, high-resolution images from nano/microsatellites face bandwidth limitations and cloud-cover issues. Real-time cloud detection is essential due to its impact on bandwidth and playback time. Despite CNNs’ success in ground-based detection, their complexity hampers real-time use on resource-limited satellites. This work introduces SICNet, a lightweight CNN for automatic cloud detection, using the Xilinx-Vitis-AI TensorFlow framework. Trained on 20,000 images, it achieves over 85% accuracy with ≤2 ms inference time on a Xilinx ZCU102 Evaluation Kit. This platform extends to applications like fire and object detection in remote sensing.