Learning-Based Methods for Detection and Monitoring of Shallow Flood-Affected Areas: Impact of Shallow-Flood Spreading on Vegetation Density
This study aims to investigate the impacts of shallow flood spreading on vegetation density using a time-series collection of Landsat images spanning 2012–2020. To do this, Support Vector Machine (SVM), Random Forest (RF), Classification and Regression tree (CART) and Deep Learning Convolutional Neural Network (DL-CNN) algorithms were employed for flood-affected areas mapping and monitoring. The models were trained by using 214, 235, 230, and 219 ground truth data for years 2012, 2014, 2017 and 2020 respectively. Our accuracy assessment via the area under curve (AUC) method reveals that the DL-CNN outperforms the SVM, the RF and the CART models for detecting and mapping shallow-flood-affected areas. The findings of this study further revealed significant changes in the NDVI values within a period before and after flood occurrence. While the mean values of the NDVI were estimated 0.232, 0.221, 0.213, and 0.232 for years 2012, 2014, 2017, and 2020, respectively, prior to flood spreading, these values increased up to 0.464, 0.476, 0.355 and 0.444, respectively following flood occurrence. Furthermore, physical-chemical soil properties (e.g., clay, EC, Na, and MgHCO3), have grown considerably in the study region following the flood spreading.
Kazemi Garajeh, M., Weng, Q., Hossein Haghi, V., Li, Z., Kazemi Garajeh, A., & Salmani, B. (2022). Learning-Based Methods for Detection and Monitoring of Shallow Flood-Affected Areas: Impact of Shallow-Flood Spreading on Vegetation Density. Canadian Journal of Remote Sensing, 1-23.
A Comparison between Sentinel-2 and Landsat 8 OLI Satellite Images for Soil Salinity Distribution Mapping Using a Deep Learning Convolutional Neural Network
In this paper, we aim to compare the suitability of Sentinel-2 and Landsat 8 OLI images for detecting and mapping soil salinity distribution (SSD) using a deep learning convolutional neural network (DL-CNN) approach. We first identified and selected six SSD predisposing variables to train the models. These variables are the normalized difference vegetation index (NDVI), land use, soil types, geomorphology, land surface temperature, and evaporation rate. Next, we collected 219 ground control points from the top 20 cm of the soil surface and randomly divided them into training (70%) and validation (30%) datasets. We then evaluated the different activation, loss/cost, and optimization functions and, finally, employed ReLu, Cross-Entropy, and Adam as the most effective activation function, loss/cost function, and optimizer, respectively. The results showed that the Sentinel-2 image (94.78% overall accuracy and a Kappa of 93.14%) is more suitable for detecting and mapping SSD than the Landsat 8 OLI image (91.45% overall accuracy and a Kappa of 90.45%). Our findings also demonstrated that the DL-CNN approach can support fast and reliable image analysis and classification. As such, this research is a promising step toward understanding, controlling, and managing the complex mechanisms of soil salinization.
