Cartography of the Invisible: Revealing Small-scale Agriculture by RapidEye Imagery in the Region of Lower Tocantins, PA.

Abstract

Remote sensing-derived data combined with the use of digital classification techniques provides a synoptic view and information about the temporal dimension of the studied phenomena on the territory, making possible to generate information about the landscape dynamic and spatial pattern in regions of large territorial extensions such as Amazon. In the mapping of large areas, such as those performed by INPE for the Legal Amazon with satellite images, land use and land cover classes are in general defined as a function of the spatial and radiometric resolution of the sensors used, so that small areas with diversified uses are generally aggregated into single and mixed class. This is the example of the so-called Mosaic of Uses class of the Land Use and Land Cover Monitoring System of the Amazon-TerraClass. This class represents in part, household agriculture, however, as this mapping is carried out based on TM or OLI images from Landsat series satellites, due to its spatial resolution of 30m and the minimum mapping area of 6.25 hectares (ha) defined by TerraClass, the identification of small production scale is compromised, since they are aggregated with other land use and land cover areas no longer distinguishable. For studies that seek to give visibility to small-scale economy spatial patterns, it is necessary to refine these classes with resolution data of better definition. In this context, the main objective of this work is to test three semi-automatic classification algorithms based on pixel and regions were tested, such as MAXVER, Bhattacharya and k-nearest neighbor (KNN) to evaluate theirs refinement capacity of the Mosaic of Uses class of the data produced by TerraClass-2014. The study area comprises part of the municipalities of Cametá, Mocajuba, and Baião, located in the northeastern region of Pará, where the production of cassava, black pepper, cacao, and açaí are economically important for the local population. For the mapping of the categories contained in the Occupancy Mosaic class, images of RapidEye, REIS sensors, orthoimage with 5m spatial resolution were used. The accuracy of the tested algorithms were estimated in 26%, 38% and 78% for MAXVER, Bhattacharya and k-nearest neighbor algorithms, respectively. In addition to the greater Global accuracy (78%), the k-nearest neighbor algorithm presented better results in relation to secondary vegetation, hydrography, and dirty pasture classes, with more than 90% of accuracy. The small-scale agriculture class presented 62% of accuracy, while the other two algorithms tested did not exceed 8%. The methodological approach developed demonstrated the feasibility of using spatial high-resolution images and semi-automatic methods to distinguish land use and land cover classes present into TerraClass's Mosaic of Uses class. The methodology can be used to complement the existing databases for the Amazon (TerraClass, MapBiomas, and IBGE), emphasizing small-scale agricultural categories, giving visibility to their production systems, frequently neglected in large extent mappings.