ABSTRACT

Afghanistan’s annual opium survey relies upon time-consuming human interpretation of satellite images to map the area of potential poppy cultivation for statistical sample design. Deep Convolutional Neural Networks (CNNs) have shown ground-breaking performance for image classification tasks by encoding local contextual information, in some cases outperforming trained analysts. In this study, we investigate the development of a CNN to automate the classification of agriculture from medium-resolution satellite imagery as an alternative to manual interpretation. The residual network (ResNet50) CNN architecture was trained and validated for delineating the agricultural area using labelled multi-seasonal Disaster Monitoring Constellation (DMC) satellite imagery (32 m) of Helmand and Kandahar provinces. The effect of input image chip size, training sampling strategy, elevation data, and multi-seasonal imagery were investigated. The best-performing single-year classification used an input chip size of 33 × 33 pixels, a targeted sampling strategy and transfer learning, resulting in high overall accuracy (94%). The inclusion of elevation data marginally lowered performance (93%). Multi-seasonal classification achieved an overall accuracy of 89% using the previous two years’ data. Only 25% of the target year’s training samples were necessary to update the model to achieve >94% overall accuracy. A data-driven approach to automate agricultural mask production using CNNs is proposed to reduce the burden of human interpretation. The ability to continually update CNN models with new data has the potential to significantly improve automatic classification of vegetation across years.

摘要

阿富汗的年度鸦片调查依靠费时的人工对卫星图像的解释来绘制潜在罂粟种植面积，以进行统计样本设计。深度卷积神经网络（CNN）通过对本地上下文信息进行编码，显示了在图像分类任务中的突破性性能，在某些情况下，性能优于训练有素的分析师。在这项研究中，我们调查了CNN的发展，以从中分辨率卫星图像中自动进行农业分类，以替代人工解释。残余网络（ResNet50）的CNN架构经过培训和验证，可使用赫尔曼德省和坎大哈省的带标签的多季节灾害监测星座（DMC）卫星图像（32 m）来描绘农业区域。输入图像芯片尺寸的影响，研究了训练采样策略，海拔数据和多季节影像。表现最好的一年级分类使用的输入芯片尺寸为33×33像素，采用了针对性的采样策略并进行了转移学习，因此总体准确性很高（94％）。包含高程数据会稍微降低性能（93％）。根据前两年的数据，多季节分类的总体准确度达到89％。更新模型以达到> 94％的总体准确度仅需要目标年份的培训样本的25％。为了减少人工解释的负担，提出了一种数据驱动的使用CNN自动化生产农业口罩的方法。用新数据不断更新CNN模型的能力有可能显着改善多年来对植被的自动分类。