University of Twente Student Theses


Detecting Artisanal Small-Scale Gold mines with LandTrendr multispectral and textural features at the Tapajós river basin, Brazil.

Fonseca Gomez, Alejandro (2021) Detecting Artisanal Small-Scale Gold mines with LandTrendr multispectral and textural features at the Tapajós river basin, Brazil.

[img] PDF
Link to full-text:
(only accessible for UT students and staff)
Abstract:Artisanal Small Scale Gold Mines (hereafter ASGM) is one the major disturbances in the Amazon Forest, which lately has received scientific attention owing to the environmental impact in water sources because of the use of mercury in the amalgamation process, the deforestation trend in the mine settlements, the association with insurgent groups, and direct linkage with 8 out 17 United Nations SDGs. UN Environment Program under the Minamata Convention on Mercury and the planetGOLD program has provided some guidelines to foster the National Action Plans (NAP) to incorporate formalization, mitigation, and adaptation processes of the ASGM activities, which leads to the reduction of mercury usage, decreasing the environmental impact and improving the socio-economic conditions of the population-related with the ASGM. The NAP encourages the local government – among others- the implementation and development of a monitoring system program dedicated to delimitating and identify the ASGM and assess the environmental impact that allows the land title and formalization of areas dedicated to. Current methods of monitoring ASGM activities are time-consuming, resource-intensive, and unable to cover extended areas such as military campaigns and aerial reconnaissance by aircraft. However, some different approaches in remote sensing have been used to face the NAP requirements and improve the monitoring of the ASGM using spectral land characteristics, concluding that mixed label classes are not separable by just the use of spectral information. Therefore, this research proposes an innovative remote sensing approach that combines multifeatured spectral and textural analysis with Landsat time series to map and detect the ASGM in the Tapajós River Basin (Brazil) from 2000 to 2019. LandTrendr is the time series algorithm employed that uses Landsat images to perform temporal segmentation of pixel value trajectories over time to detect forest disturbances. Every pixel is fitted with a linear regression model represented by a set of vertex and segments rendering the land cover changes in a 2D profile, allowing to incorporate the spectral and temporal attributes to classify forest loss because of ASGM activities. The entire set of predictors (86) were prone to assess feature relevance/importance through Variable Selection Using Random Forest (VSURF), which uses a stepwise regression in an ensemble model to efficacy remove noise considering the OOB error and the mean decrease error to select the most important features in ASGM detection. After performing feature importance and feature reduction, a final set of 33 features (26 textures, 5 spectral indices, and 2 topographic data) were used for binary (Mine – No Mine) using Random Forest Classifier. For the tuning parameter, 500 ntrees (number of trees) and 6 mtry (number of drawn candidate variables in each split) with 50 iterations were used for classification purposes, achieving an overall accuracy of 90.8%. Additionally, a probabilistic classification (probability of ASGM presence) was also performed considering that the threshold between classes is not clearly defined in the landscape and the transition areas from one class to another are fuzzy. An average vote counting of the threes in the Random Forest model was used to calculate the probability of ASGM occurrence at pixels level. The results show that the proposed method is reliable and robust to detect the forest disturbances produced by ASGM activities and arise as an alternative to achieve the NAP.
Item Type:Essay (Master)
Faculty:ITC: Faculty of Geo-information Science and Earth Observation
Programme:Geoinformation Science and Earth Observation MSc (75014)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page