Discriminating native and plantation forests in a Landsat time-series for land use policy design

Abstract

The upper Warren catchment of southwest Australia is an area of high biodiversity threatened by the loss of native vegetation and dryland salinity. Over the last 20 years it has been the target of a series of policies that incentivise land conversion to plantation forestry. Remote sensing has a key role in measuring the effectiveness of policy initiatives and trends in the plantation forest area across the landscape as this allows a prediction of the future effects on dryland salinity and aquatic biodiversity. Despite its importance to land-use policy, accurate data on land cover dynamics are not available. Regular, low cost monitoring of long-term change in the spatial distribution of plantation forests through remote sensing is a critical input into environmental policy for the catchment. To this end, a 35 year time-series of Landsat imagery was acquired, and three different classifiers were tested (Support Vector Machines-SVM; Random Forests-RF; and Classification and Regression Trees-CART) on the four bands held in common between the four Landsat Sensors (MSS 1979; TM 1992; ETM+ 2003; OLI 2014). The six major land-use and land-cover classes considered were agriculture, water, native forest, sand dunes, plantation forest and harvested native forest. In classifying the imagery the SVM and RF outperformed the CART across all classes. SVM presented an improved fit over RF, with highest overall accuracy for the 1979 image (92%) and lowest for the 2003 imagery (80%). Eucalypt dominated plantation forests reaching full canopy cover were subject to the highest rates of misclassification inasmuch as they share spectral properties with the Eucalypt dominant native forest. Despite this spectral similarity between both forest types the SVM classifier gave a slightly higher F-score for most individual classes than the RF or CART. Further, the RF class predictions lead to a spurious salt-and-pepper effect, in more cases than CART or SVM. The accuracy of the SVM classifications provide a basis for demonstrating the effects of past policy initiatives and designing efficient environmental and conservation policy in the future.

Comprehensive review on land use/land cover change classification in remote sensing

“…SVM Classification 4,20,40,41,44,47,51,56,62,63,67 This method helps to find the hyper-plane that separates two or more LU/LC classes in the satellite image.…”

“…Many researchers around the world have carried out LU/LC change analysis research over many years. Some of the study areas for LU/LC analysis have been North-eastern Latvia, 1 forests, 7 Northern Ethiopia, 9 Mexico, 14 Germany, 18 China, 23 Iran, 25 Egypt, 27 United Arab Emirates, 31 India, 32 Iraq, 37 Malaysia, 41 south-western Australia, 44 Eastern Region of Ghana, 50 Dubai, 52 south-western Nigeria, 54 Indonesia, 55 Pakistan 56 and Sri Lanka. 57 The main objective of researchers in choosing their study area depends on the causes of LU/LC change ( Figure 2), and their work assists government, forest departments, land resource and urban planners in taking the necessary actions to protect the Earth's environment.…”

“…1,7,13,16,[22][23][24][25][26]29,30,[32][33][34][35][36][37]40,[47][48][49]53,55,69,70,79 The Image De-striping, Rescaling, Point Spread Convolution and Lookup Table (LUT) Stretch methods have been used during radiometric correction. 9,13,22,23,26,29,40,41,44,47,49,70,71,79 Topographic corrections include normalise, level slicing, route intervisibility, surface difference and terrain elevation modelling of explanatory variables like slope, elevation and aspect etc. 12,23,25,29,35,37,40,74,77,…”

“…Non-linear problems were difficult to solve with the logistic regression model. 4,8,11,19,20,27,28,44 LU/LC classification is performed based on the voting results of each decision tree.…”

“…CART 10,11,19,44 The model is created by predicting the value of dependent variables based on the values of many independent variables.…”