Oral Session 12 | Friday, October 7, 11:40–12:00 | Abstract 569
Deep Learning-Based Retrieval of Cyanobacteria-Sensitive Orange band for Landsat-8/9 and Sentinel-2 Imagery
The sensors aboard Landsat-8/9 and Sentinel-2 do not capture an orange band (~ 620 nm), which is central for estimating phycocyanin concentration as an indicator of cyanobacteria. Recent studies estimate the orange band for Landsat imagery leveraging the panchromatic band. However, the current approach is not applicable to sensors like Sentinel-2 without a panchromatic band. Here, we propose a novel method based upon deep learning that allows for estimating the orange band based on multispectral bands only. We build upon a convolutional neural network (CNN) architecture and train the model by leveraging a massive number (100 k) of radiative transfer simulations. The simulations are performed for a broad range of bio-optical conditions with low (Landsat-9 and Sentinel-2) and high (Sentinel-3) spectral resolutions. Then, the low spectral resolution information of either Landsat-9 or Sentinel-2 is fed to the network as inputs, considering the simulated orange band of Sentinel-3 as output. The hyperparameters of the deep network are automatically tuned by performing a Bayesian optimization. We applied the proposed method to in-situ (2381 samples from several lakes in Germany) and satellite spectral data (San Francisco Bay). The orange band predictions are validated against the reference orange band available from either in-situ hyperspectral measurements or near-simultaneous Sentinel-3 overpasses. Furthermore, we compared the results with the panchromatic band-based approach for the Landsat-8/9 analyses. The results indicate the high potential of the proposed deep learning model in predicting cyanobacteria-sensitive orange band for both Landsat-8/9 and Sentinel-2 (R 2 = 0.95, relative error less than 15%).
Milad Niroumand-Jadidi, Fondazione Bruno Kessler, 0000-0002-9432-3032
Francesca Bovolo, Fondazione Bruno Kessler