Land Cover Monitoring in Landscapes with High Level of Heterogeneity: an Application of Multi-Temporal Feature Analysis and Machine Learning

Spatiotemporal monitoring of Land use and land cover (LULC) in heterogeneous landscapes is challenging while one pixel covers several features, only one land cover type is noted, which is the most dominant feature in the satellite images. This is one of the main sources of uncertainty in land cover maps. In this study, we propose a machine learning approach to improve the accuracy of LULC maps in heterogeneous landscapes based on the multi features time series analysis. This approach receives multi-temporal input to generate one output label map for every input image. Training of random forest machine learning was done based on the five feature groups containing: (i) band reflectance; (ii) satellite-derived indices: NDVI, NDWI, NDBI, (iii) train factors: radar backscatter coefficient, elevation, slope; (iv) climatic features: land surface temperature and evapotranspiration; and (v) texture features: Correlation, Contrast and Homogeneity. Alongside the uncertainty adjustment, the effectiveness of each feature groups are reported in LULC mapping that is benefits for researchers to optimum feature selection and reduce the cost and time of the processing. The results showed that integration of band reflectance, train factors and climatic features outperforms the other feature groups in LULC mapping in landscapes with high level of heterogeneity yielding an overall accuracy of 96% and a Kappa coefficient of 0.94. The findings also revealed that appropriate feature selection can boost the accuracy of land cover area up to 35%, which is significant for a wide range of applications in local to global scale.