Field Boundary Change Detection

Hummingbird Technologies与STFC的Hartree中心合作，提议建立一个全自动的人工智能工具，用于农业领域的边界检测和变化检测，该工具在全球范围内具有商业可行性，并且足够强大，可以解决农业中遇到的不同地质、土地利用和农业差异。该项目将以蜂鸟现有的研发项目为基础，哈特里中心将在深度神经网络和高性能计算领域贡献关键知识。挑战是双重的：首先是一个静态的挑战——定义边界。这将不使用任何先验信息，但其目的是训练一个模型，该模型可以识别跨定义区域的字段边界，或者在字段内单击后局部识别字段边界。为了确保我们能够在全球范围内做到这一点，需要考虑在不同地区实施的土地管理实践，以及仅在特定地区常见的人为特征。为了确保我们生成的东西能够准确地做到这一点，我们需要识别和定义模型使用的数据集中的每个特征。其次是动态的——使用先前的边界信息，我们将开发一种深度学习算法，该算法会随着时间的推移定期运行，以确定受土地利用变化影响的区域。该算法需要利用大型卫星图像数据集，这些数据集包含许多由不同特征定义的字段样本，旨在鲁棒地预测边界的几何变化和土地利用的变化。随着时间的推移，在捕获不同日期的图像时，检测也会包含明显的噪声和校准问题，并且需要将这些数据采集问题与实际变化区分开来。工具的彻底验证将需要通过一个连续的过程来执行，这样我们就可以度量模型的性能，并确保我们向客户提供正确的结果。

Hummingbird Technologies, in collaboration with the STFC's Hartree Centre, are proposing to build a fully automated AI tool for agricultural field boundary detection and change detection that is both commercially viable on a global scale and robust enough to tackle the differing geological, land use and farming variances encountered in agriculture.The project will build on Hummingbird's existing R&D projects, with the Hartree Centre contributing to critical knowledge gaps in deep neural networks and high performance computing.The challenge is twofold:First is a static challenge - defining the boundary. This will not use any prior information but will aim to train a model that recognises field boundaries across a defined region or, alternatively, locally after a single-click inside a field. To ensure we are able to do this on a global scale, the land management practices that are implemented in different geographies need to be taken into account as well as man made features that are only common in a specific locality. To ensure we are producing something that can do this accurately, we need to identify and define each feature within the dataset used by the model.Second is dynamic - using prior boundary information we will develop a deep learning algorithm that runs periodically over time to identify areas which are subject to land use change. This algorithm will need to make use of large satellite image datasets that contain many samples of fields which are each defined by different features, aiming to robustly predict both geometric change of the boundaries and land use changes. Detection over time will also contain significant noise and calibration issues when capturing imagery on different dates, and these data acquisition issues will need to be discriminated from actual changes. A thorough validation of the tool will need to be performed through a continuous process so that we can measure model performance and ensure we are providing the correct results to customers.