Multispectral Remote Sensing Classification using Differentiable Weightless Neural Networks

Abstract

Remote sensing image classification plays a critical role in environmental monitoring, land-use analysis, and urban planning. Although deep convolutional neural networks achieve high classification accuracy, they often require substantial computational and memory resources, limiting deployment in edge and resource-constrained systems. This paper presents a computationally efficient framework for multispectral satellite image classification using Differentiable Weightless Neural Networks (DWNs) and handcrafted feature representations. The proposed approach combines spectral indices, band statistics, HSV histograms, and texture descriptors within a multi-scale spatial pyramid framework. The extracted features are binarized using distributive thermometer encoding and classified through lookup-table-based DWN inference. Ablation studies and hyperparameter analysis evaluate the contribution of individual feature groups and LUT configurations. Experimental results on the EuroSAT benchmark show that the selected configuration (L=8 thermometer levels and n=4 LUT inputs) achieves 97.00% ± 0.05% (Cohen's κ = 0.9666, macro-F1 = 0.9686) on the full 13-band multispectral variant and 93.24% ± 0.12% on the RGB variant, with a total deployable footprint of 0.091 MB — including LUT contents, the learnable input mapping, and the thermometer encoder — over 492× smaller than ResNet-18 and roughly 100× smaller than MobileNetV2. The improvement from RGB to 13-band input is consistent with the added discriminative value of red-edge, NIR, and shortwave infrared channels, although the two pipelines also differ in feature composition. The proposed method provides an interpretable and hardware-efficient alternative for remote sensing classification in resource-constrained environments.