Efficient Remote Sensing in Agriculture via Active Learning and Opt-HRDNet

As the foundation of human survival, the development of agriculture has always played an important role in social development. The quality of agricultural development determines the speed of social progress. With the development of computer science, using computer technology to solve problems related to agricultural development has become an important research direction in current computer development. In recent years, remote sensing detection has received widespread attention, and the application of remote sensing detection technology in agriculture can provide great convenience for the development of agriculture. Benefit from the development of deep learning, remote sensing detection has made gigantic achievements. However, we have to face some challenges. First, deep learning depends on large scale of data with annotations, which expends inestimable human resources. In addition, as the depth of detection network increases, the amount of parameters explosively extends. In this work, we carry out the research based on the two problems. At the beginning, an active learning method considering classification and localization task is proposed. Our method can choose some few but valuable images. It uses 34.48% amount of training set, up to 97.4% baseline performance, and realize the compression of the scale of dataset, which reduces the trouble of manual labeling. Due to imaging and other factors, there exists many small objects in remote sensing images. So we add the mixed convolution, dilated convolution, and mosaic data augmentation modules into HRDNet network. Experiments on the agriculture dataset indicate that the improved algorithm can obtain about 2% higher than HRDNet. To reduce the number of parameters, we adjust the weighted sum ratio of importance scores dynamically. With the pruning ratio of 80%, the model volume has only 184 MB, degrading 70%. Model compression accelerates the detection speed for seven times on NVIDIA AGX Xavier, with a speed of 6 FPS. Our work will lay a foundation for remote sensing detection.