Abstract
Pará is a Brazilian state leader in deforestation and deserves special attention due to the intense artisanal mining activity that has caused severe environmental damage to the Amazon Rainforest. Remote sensing is an important tool for identifying areas degraded by mining activities. However, the large territorial extension of the Amazon Rainforest and the equally large corresponding database make the mapping by photointerpretation a costly and slow process. This study attempts to overcome this obstacle by employing the Geographic Object-Based Image Analysis (GEOBIA) approach together with Data Mining techniques in the automatic identification of areas degraded by artisanal mining in the Crepori National Forest (CNF). A NDVI image and a multiband image derived from Sentinel-2 data were segmented and the former proved to be more appropriate to the development of this research. The use of the Correlation-based Feature Selection (CFS) algorithm in attribute selection led to a 55% database dimensionality reduction. Additionally, the results obtained in the decision tree construction by the J48 algorithm showed that the spectral attributes were the most relevant in the classification of artisanal mining areas, especially the attributes related to the near infrared (NIR) band. The attributes of textural and spatial origin also contributed to the model, whereas the contextual attribute was not relevant to our classification problem. The results from classification demonstrated that the ‘Vegetation’ class is the largest in the Crepori National Forest, representing 99.50% of the total area, followed by ‘Areas Degraded by Artisanal Mining’ and ‘Other Anthropized Areas’, representing 0.17% of the total area, and, lastly, the ‘Hydrography’ class totaling 0.16%. Total anthropization in the CNF decreased between 2014 and 2017, from 2,955 ha to 2,506 ha. It is worth noting that, when compared with the Brazilian Forest Service's (Serviço Florestal Brasileiro) data, our results reveal that more than 50% (679.46 ha) of artisanal mining areas mapped in 2017 were installed after 2014, majorly in the CNF southern region. The performance of our classification model is good, reaching a global accuracy of 88.18% and a Kappa coefficient of 0.84. In class-by-class indexes, the method presented a minimum precision of 0.79 and a minimum recall of 0.75, both referring to the ‘Other Anthropized Areas’ class.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have