Characterising RADD Forest Disturbance Alerts using High Resolution Planet Data

By Etse Lossou

Abstract

Forest disturbances around the world continue to take place at alarming rates. Until recently, there was reliance on near real-time forest disturbance monitoring systems that were based on optical satellite data. These platforms are not always efficient in dense forests such as the Congo Basin because of persisting cloud cover. Today, Radar for Detecting Deforestation (RADD) alerts which is a novel cloud-penetrating forest monitoring system based on Sentinel-1, is continuously providing gap-free observations for the tropics every 6 to 12 days. Nevertheless, RADD alert is unable to provide useful information on the type of forest disturbances that are detected. To provide missing information, this study relied on freely available high-resolution optical PlanetScope imagery to characterise the RADD alerts from September 2020 to August 2021. This period was divided into four equal parts. One of the challenges that was solved was the spatial offset between the PlanetScope imagery and the RADD alerts. Using NDVI distributions, the 100th, 90th, 75th, 25th, 10th, percentile datasets were created out of the PlanetScope imagery. From the percentile datasets with four spectral bands (Blue, Green, Red, and NIR), five vegetation indices were computed. In addition, Gray Level Co-occurrence Matrix was generated from the spectral bands and vegetation indices. In total, forty-five features were derived. Using a combination of three feature selection methods namely the Recursive Feature Elimination, Recursive Feature Elimination with Cross Validation, and Random Forest, twenty-five (25) features were retained and used as input for the pixel-based classification. Four types of disturbances were identified by using Random Forest algorithm in Google Earth Engine. These disturbances were: Agricultural disturbances, Logging disturbances, Mining disturbances, and Road disturbances. The Random Forest algorithm performed well in classifying the disturbances correctly. This is based on the relatively high Macro Average F1-scores ranging from 39% to 81%. The map of the pixel-based disturbances were then pooled to the level of the RADD alert. To effectively characterise the RADD alerts with the PlanetScope imagery, a suitable percentile range should be chosen. A percentile range between 25 and 75 will be a good match for Agricultural disturbances, logging disturbances and road disturbances. For Mining disturbances, a percentile range between 50 and 90 should be considered. This study did not only explore the use of the new RADD alert, but also it successfully characterised the alerts using the newly released high-resolution optical PlanetScope imagery. It therefore provides the evidence that the PlanetScope imagery could provide the needed information for the effective monitoring of forests within the tropics. This study can also be used as a reference to develop other methods that can automatically remove the spatial offset or automatically determine the suitable percentile to use for each disturbance event.

Keywords: RADD alerts; PlanetScope; Vegetation Indices; Gray Level Co-occurrence Matrix; Feature selection; Random Forest