基于深度学习和数学形态学的经济欠发达地区农村住房智能识别研究

Intelligent Identification of Rural Housing in Economically Underdeveloped Areas Based on Deep Learning and Mathematical Morphology

中国大部分农村住房处于一种无规划的自发性建设状态。这种无序扩张是对土地资源的严重浪费,尤其是耕地资源,因此需要通过对农村住房监测来规范农房建设,同时保护耕地资源。而中国目前的农村住房监测手段主要以全国土地调查人工巡查为主,该方法收集到的信息缺乏实时性和可靠性。针对该问题,文章提出一个基于深度学习和数学形态学的经济欠发达地区农村住房智能识别模型,该模型基于高分辨率遥感影像数据和PaddlePaddle框架,创新性地引入数学形态学的膨胀腐蚀方法,并与MobileNetV2进行松散耦合,其在农村住房识别监测中的平均精度达到84.5%,且具有一定的泛化性。与ResNet34模型进行对比,该模型平均识别精度比ResNet34模型高10.6%,对地类边界的提取效果更精细,对农村住房轮廓的识别效果更好。

With the rapid growth of China’s economy and the supportive government policies and funds, the demand for and the expansion of rural housing in China are on the rise, which tightens requirements for rural housing. However, most of the present rural housing suffers from unplanned growth. The sprawl of rural housing adversely affects the quantity and quality of land resources, in particular, productive agricultural lands;therefore,it is necessary to regulate the growth of rural housing and protect farmlands through spatiotemporally continuous monitoring. Currently, the monitoring of rural housing in China is mainly conducted via in-situ inspection of the national land survey, which is restricted by unfavorable conditions(e.g., weather, outbreaks, and traffic) as well as impairing real-time and reliable control over information collected. To resolve this issue, this study proposed an intelligent model to recognize rural housing in underdeveloped areas based on deep learning and mathematical morphology(MobileNet-MM). The model was based on high-resolution remote sensing data, MobileNetV2(a convolutional neural network architecture well performing on mobile devices), and mathematical morphology.First, the obtained data were segmented and manually screened and tagged to construct a training dataset. Second,the training dataset was used to train MobileNet-MM, with the expansion operation being used to compensate for identification errors of deep learning. Finally, the accuracy of MobileNet-MM to identify and monitor rural housing was tested, resulting in 84.5% accuracy. The comparison of the accuracies of MobileNet-MM and ResNet34(a state-of-the-art image classification model) indicated that ResNet34 misclassified a large area of rural housing that was mainly distributed on the edge of the region as well as cropland and vegetation as rural housing, with its weak ability to recognize actual rural housing. The MobileNet-MM model predicted rural housing accurately and land boundary precisely, with the misclassified area being scattered, and its average accuracy, is 10.6% higher than that of ResNet34. The novelties of this study were two-fold:(1) a high-resolution training dataset of rural housing in underdeveloped areas was generated, which provides data support for the development of subsequent models;and(2) an intelligent model to recognize rural housing in underdeveloped areas(MobileNet-MM) based on deep learning and mathematical morphology was proposed.