LOCATING THE FRONTIERS OF DEFORESTATION

Kakamega Forest - Kenya; Evidence from multi-date satellite imagery and tree-structured classifiers.

by: Alex Awiti and Markus Walsh

In view of the manifold benefits that are attributable to forests, their progressive depletion must rank high in the global research and development agenda. Here, we present a summary of findings related to deforestation and soil degradation from the Kakamega Forest in Kenya and the Eastern Rainforests of Madagascar.

A significant proportion of the original forest in the Kakamega, South Nandi and North Nandi forest blocks is being lost primarily due to cultivation encroachment, settlement and to a small extent, government sanctioned deforestation for commercial tea production. This loss exerts an enormous influence on basin hydrology, climate and global biogeochemical cycles.

There is urgent need to focus on the principal loci of deforestation, i.e the main deforestation fronts. An alert strategy through remote sensing (the most comprehensive and objective source of data now available) would greatly enhance our capacity to understand the character and scale of deforestation in the principal localities. This strategy would highlight those areas where safeguarding measures are needed and where there will be the biggest payoff from timely intervention.

The overall objective of the study was to provide an alerting strategy, based on remote sensing and statistical tree structured analysis tools, to identify deforestation fronts, both present and prospective.

Image processing

An important step in monitoring vegetation change is the ability to compare changes in light reflectance among or between images from different dates and sites in different scenes. These comparisons require that digital numbers from each scene be transformed to common reference values through calibration. A calibration process was applied to raw images to remove time and scene dependent effects caused by atmospheric scattering, solar illumination angle and sensor decay through time.

A total of 42 pseudo invariant features (PIFs), which are objects spatially well defined, spectrally and radiometrically stable were identified. The digital numbers of these invariant features were extracted from Landsat TM 1986 and Landsat ETM 2000 using bands 1, 2, 3, 4, 5 and 7.  A linear regression was applied to the resulting values file using Landsat ETM 2000 as the reference image (independent variable) and Landsat TM 1986 as the overpass image (dependent variable). The slope and intercept of the regression line, referred to as gain and offset respectively were used to calibrate the image dates.

Selection of training sites

The following forest cover categories were considered for selection of training sites: (1) Forest that remained forest (f). (2) Non- forest that remained non- forest (nf). (3) Non- forest that changed to forest or regeneration/re-growth (r). (4) Forest that changed to non- forest (d). Then 72 training sites were randomly selected for each category by visual assignment based on bands 4, 3, 2 composite of the two image dates. The training sites were used to extract reflectance from bands 1, 2, 3, 4, 5, and band7 using Landsat TM 86 and Landsat ETM 2000.

Back to top

Classification Tree

Classification and Regression Tree (CART® Version. 4.0) statistical software was used to analyze reflectance data extracted from each band for the two image dates using the respective forest cover category training sites (f, nf,d and r). CART’s major role was to produce an accurate set of data classifiers by uncovering the predictive structure of reflectance data. The initial node on a tree is called the root. From the root, the model was fit using a binary recursive partitioning.

Findings

Classification rules and tree topology

From the initial node on the tree, the data was successively broken into left and right branches with the splitting rules defined by the predictor variable, reflectance.

Forest Change Dynamics

The classification rule defining each splitting node was used to process four binary images through image band reclassification. These were subsequently combined using image cross tabulation module in IDRISI 32. Reclassification was applied again to make the respective class assignments.

Back to top

Classification accuracy

100% of the non-forest and deforested pixels were classified correctly. 98.64% and 96.33% of forest and forest re-growth pixels respectively were classified correctly. Cross validation using 50 randomly held segment reveals prediction success probabilities of 0.99 for forest and deforested pixels.

.

Non-forest and forest re-growth pixels had prediction success of 0.96 and 0.93 respectively.

Conclusion

This study provides one of the most objective and comprehensive approaches for locating the frontiers of tropical deforestation. It provides a rational spatial sampling domain for addressing the main causes of deforestation; historical, present and prospective. These results will also guide the implementation of replicated spatial analogue surveys designed to assess the impacts of forest conversion on soil productivity (greenhouse bioassay), carbon, macronutrient stocks (N, P, K, S, CEC) and soil physical properties (bulk density, water stable aggregates, texture).

Back to top