Radar for rainforest: A monitoring system for land cover change in the Colombian Amazon

An internationally recognized problem is the uncontrolled conversion of forest into other land cover types. Most attention is focussed on the rapid disappearance of tropical rainforests, although deforestation is also a problem on higher latitudes. Efforts to curb and regulate the clearing of tropical rainforests have often been hampered, not only by the complexity of the social, economical and political factors driving deforestation, but also by the scarcity of up-to-date information on the location, extent and speed of the process. As extensive field data collection is not feasible because of poor accessibility, remote sensing, especially radar, can provide a viable solution for mapping and monitoring such areas. The monitoring system presented here, is based on images of the ERS-1&2 satellites. It has a generic structure, making it suitable for application in a wide range of areas by replacing only a few specific components. The system was applied in the pilot site of San Jose del Guaviare, in the Amazon rainforest of Colombia. People from other parts of the country migrate to this area in search of land and livelihood, clearing parts of the forest, cultivating crops and ultimately planting pastures for cattle breeding. Field data were collected between 1990 and 1994 in the pilot area. During the period 1992-1994, seven ERS-1 images were available. Descriptions of type and structure of the land cover were analyzed together with remote sensing data, to formulate land cover classes that are both meaningful to the user of the system as well as interpretable by means of time series of ERS-1&2 images. These land cover classes are characterized by image characteristics, such as the average level of radar backscatter per segment and the variation in the level of backscatter over three subsequent images. Simulation of level and variation of backscatter of different land cover classes with the theoretical model UTACAN confirmed the experimental findings. Prior to classification, images were segmented with the RSEG algorithm, based on edge detection and segment growing. The process of land cover change occurring in the pilot area, was modeled by a land cover change model BOSTOS. This model is used to calculate the possible changes in land cover classes between two moments in time. By using this model, unlikely changes were excluded and monitoring becomes more accurate. Two sets of formal rules were developed for the classification of the time series of ERS-1 images of the pilot site: The first set is based on the relation between land cover class and backscatter signature, as described above. The second set describes the possible changes in land cover as calculated by BOSTOS. These sets were applied to the seven images for image classification and land cover change detection. Results show that, while it is difficult to map land cover with one single ERS-1 image, three images over a time span of one year can separate forests from pastures, provided the images were taken in different seasons. Secondary vegetation is a very heterogeneous class, confused both with forest and with pastures. Time series of ERS-1 images over several years can also separate secondary vegetation from forests and pastures, with accuracy increasing with the time the monitoring system is functioning and the number of images.