基于深度学习AlexNet的遥感影像地表覆盖分类评价研究

发布时间：2018-04-06 00:13

本文选题：深度学习　切入点：地理国情普查　出处：《地球信息科学学报》2017年11期

【摘要】：地表覆盖分类信息是反映自然、人工地表覆盖要素的综合体,包含植被、土壤、冰川、河流、湖泊、沼泽湿地及各类人工构筑物等元素,侧重描述地球表面的自然属性,具有明确的时间及空间特性。地表覆盖分类信息数据量大、现势性强、人工评价费时,其自动化评价长期以来存在许多技术难点。本文基于面向对象的图斑分类体系,引入深度卷积神经网络对现有地理国情普查-地表覆盖分类数据进行分类评价,并通过试验利用AlexNet模型实现地表覆盖分类评价验证。试验结果表明,该方法可有效判读耕地、房屋2类图斑,正确分类隶属度优于99%,而由于数据较少、训练不充分,林地、水体图斑正确分类隶属度不高,分别为62.73%和43.59%。使用本文方法,经过大量数据充分微调的深度学习AlexNet可有效地计算图斑的地类隶属度,并实现自动地表覆盖分类图斑量化评价。
[Abstract]:The classification information of surface cover is a complex that reflects the elements of nature and artificial surface cover, including vegetation, soil, glacier, river, lake, swamp wetland and all kinds of artificial structures, etc., which mainly describes the natural properties of the earth's surface.Has definite time and space characteristic.There are many technical difficulties in automatic evaluation of land cover classification, such as large amount of data, strong present situation and time-consuming manual evaluation.Based on the object oriented classification system of map spot, this paper introduces the deep convolution neural network to classify and evaluate the existing geographic situation census and land cover classification data, and realizes the ground cover classification evaluation and verification by using AlexNet model.The experimental results show that this method can effectively distinguish cultivated land, house two types of spots, the correct classification and membership degree is better than 99m, but due to less data, insufficient training, forest land and water spots, the correct classification and membership degree are 62.73% and 43.59%, respectively.By using the method of this paper, the subclass membership degree of map spot can be calculated effectively by depth learning AlexNet, which is fully fine-tuned by a large amount of data, and the quantification evaluation of surface cover classification map can be realized.
【作者单位】：中国矿业大学;国家测绘产品质量检验测试中心;中国测绘科学研究院;
【基金】：国家自然科学基金项目(41671440)
【分类号】：P237

【相似文献】