Earth observation for monitoring of shifting cultivation

Abstract

Systematic monitoring of shifting cultivation areas is important for assessing forest and cropland conversion in northeast India (NEI). Recent advances in satellite image processing techniques and computational facilities have improved the efficiency of forest cover change mapping. The current study aimed to map shifting cultivation areas by monitoring deforested and burned areas using a random forest (RF) machine learning algorithm-based on satellite data-derived spectral indices and topographic variables (elevation). This study was conducted in three states in NEI: Manipur, Meghalaya, and Nagaland. Cloud-free Landsat satellite images were acquired for the premonsoon season from 2016 to 2018 to map shifting cultivation. Spectral indices related to vegetation (normalized difference vegetation index [NDVI], enhanced vegetation index [EVI], and soil adjusted vegetation index [SAVI]), burned area (normalized burned ratio [NBR], normalized burned ratio-2 [NBR2]), and leaf water (normalized difference moisture index [NDMI]) were employed. The RF classification model was developed and validated with the training data points generated for images collected in 2017 and 2018. The model demonstrated an accuracy of >80% in burned area mapping. The developed model was then applied to the images of 2016 and 2017, and the classification accuracy was assessed with more than 11,000 points, which indicated a high classification accuracy (>90% overall accuracy). Moreover, the comparison with MODIS-burned area products showed good agreement with the identified shifting cultivation areas. A total of 690 km² of forest was deforested and burned during 2016–17, which was reduced to 620 km² during 2017–18. The indices NDMI, NBR, NBR2, and elevation were identified as the dominant contributors in burned area mapping. The current study demonstrates the utility of machine learning approaches for burned-area mapping using multitemporal satellite images. Specifically, the study demonstrates the utility of spatiotemporal deforestation and burned area mapping using machine learning approaches that are useful to forest resource managers, conservationists, and decision-makers.