Exploiting Sentinel-2 dataset to assess flowing status in non-perennial rivers. A case study: Sangone river.

Non-perennial rivers are the most common type of rivers on Earth today. Due to anthropogenic pressures, such as changes in land uses, water withdrawals and climate change, the shifting from a perennial to a non-perennial condition is becoming faster and the preservation of temporary rivers and streams is in jeopardy. Their ubiquity and crucial role in biodiversity and rivers’ ecosystems defense are widely recognized. However, the lack of social perception of their importance, together with the complexity of evaluating their ephemerality, makes it difficult for policymakers to draw up appropriate legislation to establish their degraded condition and promote their preservation. In this thesis, it was used hydrogeologic data, Sentinel-2’s images and a Random Forest algorithm to detect and foresee the daily flowing status of three segments of the Sangone river in Piemonte, with the aim of assessing the duration and frequency of each flowing status. By evaluating the reflectance signature of sediments, vegetation, and water in the riverbed, and with the help of ground truth data and high-resolution images, it was found that the false color image with SWIR, NIR, and RED’s Sentinel-2 bands allows the optimum discrimination of river water compared with other classes. Furthermore, this composition allows performing the supervised classification of the segments to determine the alternation of three flowing statuses during the years: “Flowing” (F), “Ponding” (P) and “Dry” (D). Completing the dataset with significant meteorological and hydrogeological data, a Random Forest algorithm was implemented to predict the flowing status for days with a cloudy image or no image at all in the period 2015-2021, and the Boruta package to determine the most significant explanatory variables. The outcome of the supervised classification shows an unbalanced dataset, where the image with flowing status was always more than 70% of the total. Thus, the RF model with the best prediction capacity is the oversampling double Boolean model that has firstly distinguished between F/NF (not flowing, where both D and P statuses are included) statuses, and then between D/P. The accuracy obtained for each model is in the range of 0.89-0.99. The outputs highlight an important annual variability of NF status, which goes from 0 to 166 days per year. The cumulative 30-days rainfall [R30] and 90-days rainfall [R90], in some cases also at 10 days [R10], with the average of previous 90-days-maximum air temperature [TMAX90] proved to have good predictive capabilities. Specifically, it was determined the water table level was the most significant explanatory variable to distinguish both F/NF and D/P flowing statuses. The flowrate’s measurements have a significant impact in the first F/NF model, whereas the average of the previous 30-days average relative humidity has a powerful prediction capacity in the D/P model. Furthermore, the models let to determine threshold values for some explanatory variables, which makes it possible for the body in charge of on-time monitoring and prediction of the flowing statuses of Sangone. All the results can be valuable data on fighting the deterioration of the Sangone river and the extinction of non-perennial rivers in general.