Evolutionary Architecture Search for Chlorophyll-$a$ Prediction in Lakes using Sentinel-2
Architecture search yields a 409-parameter Sentinel-2 bloom model, lifting AUC to 0.820.
The paper searches an extended multilayer-perceptron space with regularized evolution for a published Sentinel-2 algal-bloom classifier, keeping the original features and lake-level split. Selection uses only inner cross-validation AUC. The best network raises 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. At about 1.6 kB, it is intended as an onboard screening trigger.
- Held-out AUC rises from 0.790 to 0.820 and accuracy from 0.733 to 0.748.
- The network has 409 parameters, 26 times fewer than the hand-designed reference.
- At about 1.6 kB it is sized for onboard bloom screening.
- Search converges on one narrow tanh layer with RMSprop and weight averaging.
From the original article
1 min read at arxiv.orgSmall 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.
By Kursat Komurcu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.