Assessing the accuracy of detected breaks in Landsat time series as predictors of small scale deforestation in tropical dry forests of Mexico and Costa Rica

Abstract

Tracking the occurrence of deforestation events is an essential task in tropical dry forest (TDF) conservation efforts. Ideally, deforestation monitoring systems would identify a TDF clearing with near real time precision and high spatial detail, and alert park managers and environmental practitioners of illegal forest clearings occurring anywhere in a region of interest. Over the past several years there have been significant advances in the design and application of continuous land cover change mapping algorithms with these capabilities, but no studies have implemented such methods over human dominated TDF environments where small–scale deforestation (<5 ha) is widespread and hard to detect with moderate resolution sensors. The general objective for this research was to evaluate the overall accuracy of the BFASTSpatial R Package for detecting and monitoring small-scale deforestation in four sites located in tropical dry forest landscapes of Mexico and Costa Rica using greenness and moisture spectral indices derived from Landsat time series. Results show a high degree of spatial agreement (90%–94%) between the distribution of TDF clearings occurred during the 2013–2016 period (as indicated by VHR imagery interpretation) and BFASTSpatial outputs. NDMI and NBR2 had the best performance than other indices and this is evidenced by the combined overall, user's and producer's accuracies. In particular, NBR2 were the most accurate predictor of deforestation with an overall accuracy of 94.5%. Our results also imply that monitoring sites at an annual basis is feasible using BFASTSpatial and LTS, but that lower confidence should be given to sub-annual products given significant systematic temporal differences between the BFASTSpatial monthly product and reference data. The possibility of including more clear observations at the spatial resolution of Landsat (30-m) or higher will greatly increase the spatial and temporal accuracies of the method. Given its performance, BFASTSpatial can help monitor hotspots of small-scale TDF loss across Central and North America at little or no cost. Users of the method should have a strong knowledge of the local land use and land cover dynamics and the ecophysiology of vegetation types present in the landscape. This local expertise is necessary for interpreting and validating results as well as communicating its output to decision-makers and stakeholders.