Object-based, multi-sensor habitat mapping of successional age classes for effective management of a 70-year secondary forest succession

Abstract

Multi-temporal change detection over decades including the pre-satellite era is challenging due to the different image types available over time, and this explains the scarcity of long-term studies of vegetation succession which can play a pivotal role in the restoration of biodiversity in regenerating forests. This study describes a semi-automated, object-based habitat classification method for change detection of tropical forest succession since 1945. The study uses a set of black and white aerial photographs and high-resolution satellite images which differ in quality and resolution, to investigate forest successional patterns and their implications for informed ecosystem and land rehabilitation management. For optimized habitat boundary delineation from black and white aerial photographs and panchromatic satellite images, three levels of hierarchical image object primitives were created. The minimum object sizes of 50 m2, 500 m2, and 1000 m2 maximized inter-object and minimized intra-object variability according to the scale of habitat patches and imagery used. Object-Based Image Analysis (OBIA) provided additional Grey-Level Co-occurrence Matrix (GLCM) textural features of segmented objects which helped to incorporate knowledge-based rule-sets into the final habitat classification which was done manually. Results show accuracies for grassland greater than 94%, monoculture plantations were distinguished from natural forest with 95% accuracy, and isolated mature stands of natural forest achieved 75% accuracy. Consideration of multi-date images increased the accuracy of distinguishing between mixed plantations and natural forest as well as between shrubland and young secondary forest. The resulting maps of vegetation structure at five time periods from 1945 to present gave new insights into the ecological processes of secondary forest succession. These include the surprising rapid rate of natural forest regeneration, at an annual rate of 7.7% from 1945 to 2014, and an even faster rate of 11% during a period when hill fires were controlled. The last areas to succeed to forest are those which are still, or at some time have been under exotic mono-cultural plantations. This suggests that long term protection from hill fire would be a better option for assisting natural succession in the landscape than plantations, which are both costly, and act as barriers to natural succession. Overall, with more than 92% mapping accuracy, the method can be adapted for other multi-temporal, multi-sensor studies as it enables inclusion of spatial theories by dividing the satellite image into time-consistent geographic entities according to the scale of target objects and image resolution. The accurate maps of forest cover patches at different successional stages can also help in site specific management of the recovering forest, such as introduction of shrub seedlings to bridge bottlenecks in seed dispersal according to shrub density and dispersal distances for forest birds. Late successional tree species can also be introduced in areas where only early successional species are present after 50 years of succession.