High space- and time-resolution drought monitoring using harmonized Landsat and Sentinel data with Drone imagery

Drought is gaining global attention due to its irrefutable and irreparable damages. Aiming at exploiting the great potential of remote sensing platforms to facilitate drought monitoring and characterization, even through multi-sensor-based approaches, this contribution underscores the efficacy of harmonizing Landsat and Sentinel data, driven by high-resolution drone imagery, to monitor drought conditions on a local scale over a large farm located in the Calabria Region, southern Italy.

To accomplish the monitoring, the Normalized Difference Vegetation Index (NDVI) has been exploited, and the cloud coverage has been evaluated at a local level so as to discard the images that are locally cloudy and shadowy and retain instead those locally cloud-free for further process. Machine learning techniques, including Support Vector Machines (SVM), Random Forest (RF), Feedforward Neural Networks (FFNN), and Convolutional Neural Networks (CNN), were employed to develop accurate cloud and shadow masks. The approach was enhanced with special spatial filtering considering seven bands for the cloud masking and the SWIR1 band for shadow masking, leading to remarkable accuracies of 96.9% for Sentinel and 89.4% for Landsat imagery.

Remote sensing data harmonization from different sources was driven by high-resolution drone imagery. Specifically, on July 12, 2024, a drone survey was carried out, and the reflectance in its Red and NIR bands (needed for NDVI calculation) was compared with that provided by satellite data for the same date, highlighting that Sentinel’s reflectance is radiometrically closer to that provided by the drone.

Subsequently, Landsat and Sentinel data were harmonized, and Landsat data were modified to converge to the Sentinel data. In order to do this, over the six months ranging from April 15 to October 15, 2024, a linear relationship between the Landsat and Sentinel Red and NIR spectral bands was determined in the dates when both images were available at most one day of distance. Then, the linear equation coefficients were also estimated for Landsat images acquired at more than one day of distance from Sentinel ones, applying a linear interpolation over time between the closest dates with simultaneous or near-simultaneous (i.e., one-day difference) acquisition between the two platforms.

The procedure was tested by comparing the extracted NDVI values (namely, Sentinel NDVI and harmonized Landsat NDVI) with the local information about agricultural activities and with other four high-resolution drone surveys, implying the effectiveness of the proposed methodology. The proposed integrated approach not only improves the monitoring of drought conditions but can also help agricultural management and disaster response in vulnerable regions.

Acknowledgments: This study was funded by The Next Generation EU - Italian NRRP, Mission 4, Component 2, Investment 1.5, call for the creation and strengthening of ‘Innovation Ecosystems’, building ‘Territorial R&D Leaders’, Project Tech4You - Technologies for climate change adaptation and quality of life improvement, n. ECS0000009.