Land Classification in Satellite Images by Injecting Traditional Features to CNN Models

TL;DR

This paper introduces a method to enhance small CNN models for satellite land classification by injecting traditional features, significantly improving their accuracy while maintaining low memory requirements.

Contribution

The study proposes a novel approach of injecting traditional features into small CNN models to boost their land classification accuracy on satellite images.

Findings

Traditional feature injection improves small CNN accuracy.

Significant accuracy gains on EuroSAT dataset.

Effective for models with sizes from 0.5 MB to 528 MB.

Abstract

Deep learning methods have been successfully applied to remote sensing problems for several years. Among these methods, CNN based models have high accuracy in solving the land classification problem using satellite or aerial images. Although these models have high accuracy, this generally comes with large memory size requirements. On the other hand, it is desirable to have small-sized models for applications, such as the ones implemented on unmanned aerial vehicles, with low memory space. Unfortunately, small-sized CNN models do not provide high accuracy as with their large-sized versions. In this study, we propose a novel method to improve the accuracy of CNN models, especially the ones with small size, by injecting traditional features to them. To test the effectiveness of the proposed method, we applied it to the CNN models SqueezeNet, MobileNetV2, ShuffleNetV2, VGG16, and ResNet50V2 having size 0.5 MB to 528 MB. We used the sample mean, gray level co-occurrence matrix features, Hu moments, local binary patterns, histogram of oriented gradients, and color invariants as traditional features for injection. We tested the proposed method on the EuroSAT dataset to perform land classification. Our experimental results show that the proposed method significantly improves the land classification accuracy especially when applied to small-sized CNN models.