Résumé
[Display omitted]
•First approach for hydrological connectivity mapping using spectral similarity and machine learning.•Procedures for large training dataset generation for lake connectivity mapping.•L-CONNECT framework achieved 88% overall accuracy in connectivity mapping.•River-lake spectral similarity indices are reliable proxy for connectivity analysis.
Hydrological connectivity based on water surface connectivity controls the water exchange between large rivers and their floodplain lakes, which occurs by channelized flow through floodplain areas and overbank flow, and it is relevant to sustain the ecosystem’s health and biodiversity of floodplain waterbodies. Given the climate change impacts on floodplain aquatic habitats, further studies are needed to understand and quantify the river–lake connectivity and its temporal dynamics. However, few studies are dedicated to objectively estimating hydrological connectivity, and new commercial satellite datasets and machine learning approaches can advance the understanding of this important topic. This paper proposes a new framework for computing the hydrological connectivity of small floodplain lakes (river-Lake CONNECTivity or L-CONNECT) in the Amazon Juruá River. The L-CONNECT framework consists of spectral similarity analysis with machine learning in the river–lake system and three steps were implemented: (i) sampling process based on independent satellite-imagery visual interpretation; (ii) automated similarity features extraction from river-lake system; and (iii) training and validation of machine learning algorithm. A total of 552 3-m PlanetScope SuperDove imagery were acquired in 2020 and 2021 to perform our approach. In general, the L-CONNECT framework achieved 88% overall accuracy. However, not-connected lakes were not easily estimated. We found that the L-CONNECT framework managed to perform accurately over all lakes investigated independent on their distance to Juruá River channel (average accuracy of ∼ 86%), and that there was a low discordance (less than 30%) between the lakes’ hydrological connectivity responses acquired by Sentinel-2 against PlanetScope data. The mapping results showed that lakes were more connected to the Juruá River during the 2021 (not-connected lakes rate of 28%) compared to 2020, which was explained by higher cumulative precipitation during January, February, and March of 2021. Finally, the new L-CONNECT framework proposed here can support hydrological connectivity mapping of small floodplain lakes by considering assumptions relative to the similarity between river and water spectra as a proxy for that.