Remote Sensing for Agriculture using UAV and Satellite data with Machine Learning

农业在发达国家和发展中国家的经济中都占主导地位，是加拿大的一个重要部门。精准农业旨在优化田间规模的化肥、水和农药等投入，以提高作物的质量和产量，降低农民的成本，并将对环境的负面影响降至最低。现代遥感传感器技术使精准农业成为可能，例如无人机上的传感器用于极高空间分辨率的场内变量检索和测绘，卫星用于相对高空间分辨率的大面积测绘。然而，在将遥感用于农业时，还有一些悬而未决的问题和挑战。例如，光学卫星图像受到云层覆盖的影响。在世界一些地区，作物生长期云层不断，难以或不可能获得多时相光学卫星图像。在作物类型分类中，据我们所知，间作(同时种植一种以上作物)类型尚不能利用卫星遥感进行分类。利用合成孔径雷达(SAR)卫星数据进行生物物理参数的反演仍不能满足单模SAR对作物进行时间序列和实时监测的要求。现有的无人机传感器大多是多光谱传感器，光谱带宽较宽，只覆盖部分光谱范围，对作物生化变量的提取能力有限。为了解决这些问题和挑战，在本建议中，(1)我们将利用深度学习算法，从多时相和多源(光学和雷达)卫星数据融合中开发出准确和详细的作物/间作类型分类的新方法。(2)我们将开发新的方法，利用机器/深度学习的多时相、多模式合成孔径雷达卫星数据，准确地反演时间序列的作物高度和土壤水分。(3)我们将使用基于无人机的高光谱和具有深度学习的光探测和测距(LiDAR)数据来提取田间植物体内的生化和生物物理变量，并用于生物量和产量估计。在这一资助期间，将培训6名博士、5名硕士、5名本科生和其他博士后。新的和改进的作物分类和生物物理和生化变量的提取算法不仅将有助于遥感图像分析、植物和土壤遥感信息提取的研究领域，而且将为作物监测和精准农业做出贡献。上述方法将用于作物模型、作物监测和产量预测，并将为可持续和智能农业提供急需的信息。这项研究将在粮食安全、可持续农业、改善作物管理、适应全球气候变化和提高加拿大人的生活质量方面惠及加拿大。

Agriculture plays a dominant role in the economies of both developed and developing countries and is an important sector in Canada. Precision agriculture aims to optimize inputs at field-scale such as fertilizer, water, and pesticides to improve quality and yield of crops, reduce costs to farmers, and minimize negative environmental impacts. Precision agriculture is enabled by the modern remote sensing sensor technology, such as sensors on Unmanned Aerial Vehicles (UAVs) for very high-spatial-resolution intra-field variable retrieval and mapping, and satellites for relatively high-spatial-resolution large area mapping. However, when using remote sensing for agriculture, there are unresolved issues and challenges. For example, optical satellite images are affected by cloud cover. In some parts of the world, there is constant cloud cover during the crop growing period, making it difficult or impossible to obtain multi-temporal optical satellite images. In crop type classification, to our knowledge, the intercrop (planting more than one crop together) types could not be classified yet using satellite remote sensing. The retrieval of biophysical parameters using Synthetic Aperture Radar (SAR) satellite data still could not meet the requirements of time series and real time crop monitoring using single mode SAR. Most available UAV sensors are multispectral sensors with a few broad spectral bands, and only cover a portion of spectral range, and their ability to extract crop biochemical variables is limited. To solve these issues and challenges, in this proposal, (1) we will develop new methods for accurate and detailed crop/intercrop type classification from the fusion of multi-temporal and multi-source (optical and radar) satellite data using deep learning algorithms. (2) We will develop new methods for accurate retrieval of time series crop height and soil moisture using multi-temporal and multi-mode SAR satellite data with machine/deep learning. (3) We will use the UAV-based hyperspectral and Light Detection and Ranging (LiDAR) data with deep learning to retrieve intra-field plant biochemical and biophysical variables, and for biomass and yield estimation. During this grant funding period, six PhDs, five Master's, five undergraduate students and additional post-doctors will be trained. The new and improved algorithms for crop classification and retrieval of biophysical and biochemical variables will contribute not only to the research fields of remote sensing image analysis, plant and soil information retrieval from remote sensing, but also to crop monitoring and precision agriculture. The methods proposed above will be used for crop models, crop monitoring and yield prediction and will provide much needed information for sustainable and smart agriculture. The research will benefit Canada in terms of food security, sustainable agriculture, improving crop management, adapting to global climate change and improving the quality of life of Canadians.