Abstract

In recent years, research has been conducted in scientific institutions of the Ministry of Agriculture of the Russian Federation and the Russian Academy of Sciences on introducing new technologies for the use of aerospace information in agriculture. This article, using the example of Stavropol krai, considers the possibility of using cloud services such as Google Earth Engine (GEE) and Kaggle machine learning systems for mapping agricultural fields using deep learning methods based on remote sensing data. Median images of the Sentinel 2 space system for the 2022 growing season are used as data for the selection of training and validation samples. The total volume of the prepared training samples is 3998 images. One problem for researchers and manufacturers in the field of agriculture is a lack of centralized and verified sources of geospatial data. Deep learning methods are able to solve this problem by automating the task of digitizing the geometries of agricultural fields based on remote sensing data. One of the limitations in the widespread use of deep learning is its high demand for computing resources, which are not always available to a researcher or manufacturer in the field of agriculture. This paper describes the process of preparing the necessary data for working with a neural network, including correcting and obtaining satellite images using GEE, their standardization for training a neural network in Kaggle, and further use locally. A neural network of the U-net architecture is used as part of the study. The final classification quality is 97%. The threshold of division into classes according to the classification results is established empirically and amounts to 0.62. The proposed approach makes it possible to significantly reduce the requirements for the local use of PC computing power. All the most resource-intensive processes related to the processing of satellite images are performed in the GEE system, and the learning process is transferred to the resources of the Kaggle system. The proposed combination of cloud services and deep learning methods can contribute to a wider spread of the use of modern technologies in agricultural production and scientific research.

中文翻译：

---

利用深度学习和云服务根据地球遥感数据绘制农田地图

摘要

近年来，俄罗斯联邦农业部和俄罗斯科学院科研机构就引进航天信息在农业中的应用新技术进行了研究。本文以斯塔夫罗波尔边疆区为例，考虑使用 Google Earth Engine (GEE) 和 Kaggle 机器学习系统等云服务，利用基于遥感数据的深度学习方法绘制农田地图的可能性。 Sentinel 2 空间系统 2022 年生长季的中值图像被用作选择训练和验证样本的数据。准备的训练样本总数为3998张图像。农业领域的研究人员和制造商面临的一个问题是缺乏集中且经过验证的地理空间数据源。深度学习方法能够通过自动执行基于遥感数据的农田几何形状数字化任务来解决这个问题。深度学习广泛使用的限制之一是其对计算资源的高需求，而农业领域的研究人员或制造商并不总是能够获得这些资源。本文描述了准备使用神经网络所需数据的过程，包括使用 GEE 校正和获取卫星图像、在 Kaggle 中训练神经网络的标准化以及在本地进一步使用。 U-net 架构的神经网络被用作研究的一部分。最终分类质量为97%。根据分类结果进行分类的阈值是根据经验确定的，为0.62。所提出的方法可以显着降低对本地使用 PC 计算能力的要求。所有与卫星图像处理相关的资源最密集的过程都在GEE系统中进行，学习过程转移到Kaggle系统的资源中。云服务和深度学习方法的结合可以有助于现代技术在农业生产和科学研究中的更广泛应用。