Innovation in land use and land cover classification for landslide analysis

Land use and land cover (LULC) analysis has become increasingly significant in environmental studies due to its direct impact on the environment. Changes in LULC affect the ecological and climatic balance, in addition to increase the terrain’s susceptibility to hazardous phenomena. However, one of the key challenges in analyzing LULC time series is the presence of classification errors, which can result in invalid transitions. Invalid transitions, which represent unlikely or impossible land cover changes in a given period, can lead to misinterpretations of causes or consequences of hazard events, especially in highly susceptible areas, such as mountainous regions. 

To enhance the accuracy of identifying these changes, a team of researchers from countries such as Brazil, Ecuador, and China presented a method that integrates the Random Forest (RF) algorithm with the temporal approach of the Compound Maximum a Posteriori (CMAP) algorithm, referred as RF-CMAP. Unlike traditional methods that treat each year independently, the CMAP algorithm considers the temporal dynamics of LULC changes. It evaluates the probability of transitions over time, ensuring that reported changes are consistent with observed natural processes. By integrating the advantages of RF with CMAP, the new RF-CMAP method reduces invalid transitions and improves LULC classification. 

Improving the accuracy of land cover change analysis 

The classification process integrated the probabilities of each LULC class to be classified in each image pixel, as determined by the RF algorithm, with the temporal approach provided by the CMAP algorithm. For this purpose, Landsat images from three years (2000, 2008 and 2016) were used for the analysis, and compared results to those obtained using the traditional RF method. Although both methods presented high performance, with overall accuracy values higher than 0.87, the RF-CMAP method outperformed RF in all analyzed years, correcting 99.92 km2 (12% of the total area) of invalid transitions identified in RF classifications. 

The study also highlights the validations and performance analyses of the classifications generated by each model. For example, the overall accuracy of LULC change areas between 2000 and 2008 was 0.622 for RF and 0.703 for RF-CMAP, with RF-CMAP correcting 78% of errors related to invalid transitions during this period. The error correction rate by RF-CMAP increased to 81% for the period 2008-2016. In addition, RF-CMAP significantly reduced the salt-and-pepper effect, enhanced the homogeneity of classified regions, and eliminated errors observed in RF classifications. This included a notable improvement in the classification of areas with no LULC change, such as forest, bare soil, and water. Between 2000 and 2008, RF-CMAP corrected 50% more errors in these regions than traditional RF method. 

The key role of technology in preventing disasters caused by natural hazards. 

The study area, the Rolante River Basin, experienced heavy rainfall in 2017 that triggered more than 300 landslides. Extreme rainfall events, when combined with LULC changes, can increase soil instability. In this context, accurately identifying LULC changes is essential to understanding the factors contributing to hazardous phenomena occurrence and preventing future disasters. 

In this research, 35% of the landslides could be related to invalid transitions between 2000 and 2008, and 16% between 2008 and 2016. These invalid transitions can misrepresent the environmental conditions leading to landslides. For example, 35% of these landslides could be associated with reforestation, even though there is no evidence of reforestation in this area over an eight-year period. The RF-CMAP method effectively avoided these invalid transitions, correctly classifying 66% of landslide affected areas, compared to only 21% with the traditional RF model. 

In conclusion, integrating advanced technologies, such as RF and CMAP, represents an important advance in the temporal analysis of LULC changes, offering valuable insights to improve disaster risk management. By addressing invalid transitions in landslide-prone regions, this model has the potential to substantially enhance disaster prevention and safeguard of vulnerable communities. As remote sensing technologies and predictive algorithms continue to improve, their widespread adoption could revolutionize the sustainable management of natural resources and support disaster risk management.