2D Target/Anomaly Detection in Time Series Drone Images Using Deep Few-Shot Learning in Small Training Dataset

Abstract

Optimization problem is an important challenge in two-dimensional (2D) target/anomaly detection as a real-world application. As manual time series drone image interpretation is time-consuming and expensive, deep learning methods are of high interest for 2D target/anomaly detection. Despite 2D target and anomaly detection from time series drone images based on deep learning models is an active field in remote sensing engineering, but annotating remote sensing time series data is costly for training step. To build robust machine learning methods in remote sensing, deep few-shot learning approaches have been developed from real-world and real-time datasets based on drone images in small training data as an optimized solution. In this chapter, we focus on two real-world applications of 2D target/anomaly detection based on a new deep few-shot learning method, which can be widely used in urban management and precision farming. The experiments are based on two time series multispectral datasets, including traffic monitoring (as a target) and weed detection (as an anomaly). Compared with the few-shot learning with different backbones, the proposed method, called SA-Net, demonstrates better performance and good generalization ability for 2D target/anomaly detection.