A Comparison of Remote Sensing Change Detection Techniques in East African Montane Forests




Cook, Chelsea

Journal Title

Journal ISSN

Volume Title



Deforestation is a major concern for many countries throughout the world, including those in East Africa. It is essential to accurately keep track of changes in forests over time both for historical reference, as well as for government decision-making purposes. A large variety of change detection algorithms have been developed for remote sensing to detect fluctuations in forests. Yet, it is hard to determine which one of the multitude of techniques available is the best method for a particular study. Therefore, this study focused on using remote sensing to determine which one method out of the five different change detection techniques chosen to perform this project provides the best results for detecting conversions in East African forests. The five change detection techniques compared: 1) image differencing (ID), 2) post-classification comparison (PCC), 3) normalized difference vegetation index differencing (NDVID), 4) principal component analysis (PCA) and 5) visual interpretation (VI). These five methods are compared with the use of two Landsat Thematic Mapper (TM) Surface Reflectance Climate Data Record (CDR) images of subset areas at three separate sites. The sites are Mount Kenya and the Mau Forest Complex in Kenya and Mount Kilimanjaro in Tanzania. These sites were selected due to the large amount of recent deforestation. The results for all five of the change detection methods were compared for each of the three sites and found that the PCC technique yielded the highest overall accuracy at 90%, followed closely by PCA at 89%. The VI technique came in third at 88%, with ID and NDVID having the lowest accuracies at 87% and 83% respectively. The forest loss (FL) user’s and producer’s accuracies were also compared with PCC, PCA and VI receiving the highest accuracies at Mount Kenya. PCA had the highest user’s accuracy of 81.1% and PCC had the highest producer’s accuracy of 98.5%.VI ranked second for both user’s (56.6%) and producer’s (77.4%) accuracies. At the Mau Forest Complex, PCC had the highest user’s accuracy of 96.0% followed by NDVID at 94.7%. ID was the third highest at 92.4%. PCA received the best producer’s accuracy at 99.1%, with PCC in second at 97.8% and VI in third at 90.6%. Finally, at Mount Kilimanjaro, both ID and VI had the same highest user’s accuracies of 85.6%, followed by PCA at 82.1%. However, both ID and VI received low producer’s accuracies of 46.8% and 42.8% respectively. Both PCC and PCA did the best overall with user’s accuracies of 80.1% (PCC) and 82.1% (PCA) and producer’s accuracies of 77.6% (PCC) and 76.0% (PCA). When all three sites were averaged, PCC had the highest overall average for FL accuracy at 83.9%, followed by PCA at 81.4%, then VI at 76.2%, next ID at 68.9% and then finally NDVID at 55.0%. The findings from averaging the FL accuracies support the results of the overall accuracies, which conclude that the most beneficial techniques from this study are as follows: 1) PCC, 2) PCA, 3) VI, 4) ID and 5) NDVID.



Remote sensing, Change detection, Comparison of techniques, East Africa, Montane Forests, Accuracy assessment