Lightweight Convolutional Neural Network for Land Use Image Classification

Abstract

Convolutional Neural Networks (CNN) have proven to be pivotal in advancements in the domain of computer vision. One of the most important and highly researched applications of CNN is in land use classification using aerial imagery. While there is a lot of research conducted using advanced techniques like transfer learning, data augmentation, CNN cascading and many more to elevate the performance of classification models, the power of combination of simpler and computationally efficient approaches of CNN configuration like relevant filtering, better normalization, and accurate placement of dropouts is often underestimated. Although highly deep and the complex architectures provides better accuracy, they exhibit adverse performance in terms of computational cost, time and efforts required to develop and deploy the models. Thus, there is always a tradeoff between improved accuracy and ease of development and implementation. This paper demonstrates a lightweight CNN configuration that results in significantly high validation accuracy of 88.29% on well-known UC Merced land use image classification dataset without underfitting or overfitting. This accuracy is competitive with that achieved by many advanced states of the art architectures on the same dataset. This shows that putting time and effort in correct utilization and configuration of simple features is always worth a consideration before pursuing complex and computationally expensive approaches. Development and implementation of  such  simple  architectures would be particularly useful in land use classification in developing countries and/or municipalities with limited budgets and less powerful systems with memory constrains