Evolutionary Architecture Search Finds a 1.6 kB MLP That Boosts Algal Bloom Detection on Sentinel-2 Data

Forum topic · 小凯 · 2026-10-09

Summary

A new arXiv paper (2610.10496) by Kursat Komurcu and Linas Petkevicius explores whether neural architecture search (NAS) can improve small tabular models used in operational Earth observation. The authors revisit a published Sentinel-2 algal bloom classifier, keeping the task, expert-designed spectral features, and lake-level train/test split fixed. Using regularized evolution over an extended multilayer-perceptron search space, with selection based solely on inner cross-validation AUC, they discover networks that raise held-out AUC from 0.790 to 0.820 and accuracy from 0.733 to 0.748 with only 409 trainable parameters — 26x fewer than the strongest hand-designed reference. The search consistently converges on a recipe practitioners might not choose by default: a single narrow layer, RMS normalization, tanh activation, step-decay RMSprop, and weight averaging. The resulting 1.6 kB model is small enough to run as an onboard screening trigger on satellites. Code is available at https://github.com/VU-AIML/automl4eo-bloom-nas.

Paper Overview

Field: Machine Learning Authors: Kursat Komurcu, Linas Petkevicius Published: 2026-10-07 arXiv: 2610.10496

Abstract (translated)

Small tabular datasets with expert-designed spectral features are the norm in operational Earth observation, and the networks applied to them are typically hand-designed. We revisit one such published model — a Sentinel-2 algal bloom classifier — and ask what architecture search adds, holding the task, the features and the lake-level train/test split of the original study fixed. Searching an extended multilayer-perceptron space with regularized evolution, and selecting on inner-cross-validation AUC only, we find networks that improve held-out AUC from 0.790 to 0.820 and accuracy from 0.733 to 0.748 while using 409 trainable parameters, 26 times fewer than the strongest hand-designed reference. The search converges on a consistent recipe — a single narrow layer, RMS normalisation, $\tanh$ activation, step-decay RMSprop and weight averaging — that practitioners are unlikely to arrive at by default. The model is 1.6 kB, small enough to serve as an onboard screening trigger, which is precisely the application scenario motivating this work.

Key Findings

Regularized evolution over an extended MLP search space outperforms hand-designed architectures on a fixed chlorophyll-a / algal bloom classification task.
Held-out AUC improves from 0.790 to 0.820; accuracy from 0.733 to 0.748.
Winning network uses only 409 trainable parameters (26x fewer than the strongest hand-designed reference) and weighs just 1.6 kB.
The consistent winning recipe: a single narrow layer, RMS normalization, tanh activation, step-decay RMSprop, and weight averaging.
Tiny model size enables onboard satellite screening triggers — the practical motivation of the study.
Experimental protocol: task, features, and lake-level train/test split held fixed from the original study; model selection using inner-cross-validation AUC only.

Resources

Paper: https://arxiv.org/abs/2610.10496
Code: https://github.com/VU-AIML/automl4eo-bloom-nas