On the Neuron Response Features of Convolutional Neural Networks for Remote Sensing Image

Jie Chen, Min Deng, Haifeng LiarXiv 2014 · DOI · arXiv

Abstract: In this paper, some patterns of the Neuron Response of deep Convolutional Neural Networks were observed.

Introduction

Deep learning has revolutionized remote sensing, turning massive streams of satellite data into actionable maps. However, for many researchers, these models remain "black boxes." This study, conducted by researchers at Central South University, attempts to peel back the layers of Convolutional Neural Networks (CNNs) to understand their internal logic when processing complex geographical data. Specifically, it addresses the "semantic gap"—the disconnect between raw pixel values and high-level human concepts like land-use categories.

Methodology

The researchers utilized the Caffe framework to experiment with two classic architectures: AlexNet and CaffeRefNet. To study the impact of depth, they created versions of these networks with 3, 4, and 5 convolutional layers (e.g., AlexNet-3Conv). The core analysis relied on the Deep Visualization Toolbox (DVT) to identify "stable response filters"—filters in the highest convolutional layer that consistently fire when shown images of a specific class. The study utilized the UC Merced (UCM) Dataset, consisting of 21 land-use classes with 256×256 pixel RGB images at 0.33m resolution.

Results and Discussion

Category-Specific Depth Requirements

A primary finding of this research is that "deeper is better" is not a universal rule in remote sensing. As illustrated in Figure 1, the optimal network depth varies significantly by terrain type. For instance, complex categories like "Buildings" and "Baseball Diamonds" achieved peak accuracy with 5 convolutional layers. Conversely, classes such as "Forests" and "Harbors" reached peak performance with only 3 layers, suggesting that these land types are defined by simpler, low-level textures rather than complex hierarchical features.

Classification accuracy of AlexNet with varying convolutional depths across 21 land-use categories.

Semantic Selectivity and Stable Signatures

Through deconvolutional visualization, the authors discovered that each land-use type activates a unique, stable combination of filters. These high-level filters act as selective semantic detectors for human-understandable attributes. For example, specific filters were found to focus exclusively on individual houses in residential areas or the water-dock interface in harbors. Figure 2 demonstrates these stable activation patterns across multiple samples, proving that the network develops consistent "signatures" for distinct geographical features.

Visualization of stable neuron response features across different remote sensing image samples.

Robustness and Occlusion Thresholds

Robustness tests using systematic image occlusion revealed a distinct "tipping point" for model performance. Most architectures maintained reasonable accuracy until approximately 20% of the input image was obscured. Beyond this 0.2 ratio, accuracy declined linearly. Notably, deeper networks (5 layers) demonstrated significantly higher resilience to occlusion than shallower 3-layer models, maintaining over 50% accuracy even with substantial portions of the image removed.

Conclusion

This study provides empirical evidence that CNNs develop stable, semantically meaningful internal representations for remote sensing imagery. The findings suggest that network architecture should be tailored to the complexity of the target land-use categories. Future work will explore the mathematical relationships governing these neuron responses and evaluate the consistency of these signatures across different satellite sensors and geographical regions.

Summary updated

Abstract

In this paper, some patterns of the Neuron Response of deep Convolutional Neural Networks were observed.
BibTeX
@misc{chen2014neuron,
  author = {Jie Chen and Min Deng and Haifeng Li},
  title = {{On the Neuron Response Features of Convolutional Neural Networks for Remote Sensing Image}},
  year = {2014},
  eprint = {1408.5549},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1408.5549},
}
Figure