AI-Based Crop Detection using Spectral Remote Sensing Images

中/高特殊分辨率免费卫星图像的可用性使具有挑战性的动态土地利用和土地覆盖制图问题有可能得到解决。这个项目是关于开发用于作物检测和绘图的新人工智能技术。作物检测是基于人工智能的时间序列分析的第一步，旨在为许多社会经济应用提供基础信息。例如作物控制和产量估计、变化监测、供应链和粮食安全、轮作、保险和施肥服务等气候变化政策。缺乏地面真实数据是作物检测的一个主要问题。大多数历史数据的时间序列分析都是如此。另一方面，光谱卫星图像可以感知一年中特定作物在视觉和化学特性上的变化。因此，本项目的重点是开发有效利用光谱波段进行作物检测的人工智能技术。这是基于(1)一个无监督框架来识别有效的光谱波段，用于时间序列分析和作物检测。为此，利用作物指纹和其他植被指数来识别重要的光谱波长。(2)利用视觉变形预测策略开发新的频谱注意模型的监督框架。

The availability of free satellite images in medium/high special resolution has enabled possible solutions for the challenging dynamic land use and land cover mapping problems. This project is about developing new AI techniques for crop detection and mapping. Crop detection is the first step in AI-based time series analyses, aiming to provide fundamental information for many socio-economic applications. Examples are crop control and yield estimation, change monitoring, supply chain and food security, climate change policies such as crop rotation, insurance, and fertilization services.The lack of ground truth data is a major problem for crop detection. That is the case for most time-series analyses of historical data. On the other hand, the crop-specific variations in visual and chemical characteristics during a year are sensed by spectral satellite images. Therefore, this project is focused on developing AI techniques effectively utilizing the spectral bands for crop detection. This is achieved based on (1) an unsupervised framework to identify effective spectral bands for time-series analysis and crop detection. For this aim, the crops fingerprints and other vegetation indexes are used to identify the important spectral wavelengths. (2) a supervised framework for developing a novel spectral attention model using visual transformers prediction strategies.