Fanjinshan National Nature Reserve (FNNR) is a biodiversity hotspot in China that is part of a larger, multi-use landscape where farming, grazing, tourism, and other human activities occur. The steep terrain and persistent cloud cover pose challenges to robust vegetation and land use mapping. Our objective is to develop satellite image classification techniques that can reliably map forest cover and land use while minimizing the cloud and terrain issues, and provide the basis for long-term monitoring. Multi-seasonal Landsat image composites and elevation ancillary layers effectively minimize the persistent cloud cover and terrain issues. Spectral vegetation index (SVI) products and shade/illumination normalization approaches yield significantly higher mapping accuracies, compared to non-normalized spectral bands. Advanced machine learning image classification routines are implemented through the cloud-based Google Earth Engine platform. Optimal classifier parameters (e.g., number of trees and number of features for random forest classifiers) were achieved by using tuning techniques. Accuracy assessment results indicate consistent and effective overall classification (i.e., above 70% mapping accuracies) can be achieved using multi-temporal SVI composites with simple illumination normalization and elevation ancillary data, despite the fact limited training and reference data are available. This efficient and open-access image analysis workflow provides a reliable methodology to remotely monitor forest cover and land use in FNNR and other mountainous forested, cloud prevalent areas.

Remote sensing scientists are increasingly adopting machine learning classifiers for land cover and land use (LCLU) mapping, but model selection, a critical step of the machine learning classification, has usually been ignored in the past research. In this paper, step-by-step guidance (for classifier training, model selection, and map production) with supervised learning model selection is first provided. Then, model selection is exhaustively applied to different machine learning (e.g. Artificial Neural Network (ANN), Decision Tree (DT), Support Vector Machine (SVM), and Random Forest (RF)) classifiers to identify optimal polynomial degree of input features (d) and hyperparameters with Landsat imagery of a study region in China and Ghana. We evaluated the map accuracy and computing time associated with different versions of machine learning classification software (i.e. ArcMap, ENVI, TerrSet, and R). The optimal classifiers and their associated polynomial degree of input features and hyperparameters vary for the two image datasets that were tested. The optimum combination of d and hyperparameters for each type of classifier was used across software packages, but some classifiers (i.e. ENVI and TerrSet ANN) were customized due to the constraints of software packages. The LCLU map derived from ENVI SVM has the highest overall accuracy (72.6%) for the Ghana dataset, while the LCLU map derived from R DT has the highest overall accuracy (48.0%) for the FNNR dataset. All LCLU maps for the Ghana dataset are more accurate compared to those from the China dataset, likely due to more limited and uncertain training data for the China (FNNR) dataset. For the Ghana dataset, LCLU maps derived from tree-based classifiers (ArcMap RF, TerrSet DT, and R RF) routines are accurate, but these maps have artefacts resulting from model overfitting problems.

The goal of this study was to map and quantify the number of newly constructed buildings in Accra, Ghana between 2002 and 2010 based on high spatial resolution satellite image data. Two semi-automated feature detection approaches for detecting and mapping newly constructed buildings based on QuickBird very high spatial resolution satellite imagery were analyzed: (1) post-classification comparison; and (2) bi-temporal layerstack classification. Feature Analyst software based on a spatial contextual classifier and ENVI Feature Extraction that uses a true object-based image analysis approach of image segmentation and segment classification were evaluated. Final map products representing new building objects were compared and assessed for accuracy using two object-based accuracy measures, completeness and correctness. The bi-temporal layerstack method generated more accurate results compared to the post-classification comparison method due to less confusion with background objects. The spectral/spatial contextual approach (Feature Analyst) outperformed the true object-based feature delineation approach (ENVI Feature Extraction) due to its ability to more reliably delineate individual buildings of various sizes. Semi-automated, object-based detection followed by manual editing appears to be a reliable and efficient approach for detecting and enumerating new building objects. A bivariate regression analysis was performed using neighborhood-level estimates of new building density regressed on a census-derived measure of socio-economic status, yielding an inverse relationship with R² = 0.31 (n = 27; p = 0.00). The primary utility of the new building delineation results is to support spatial analyses of land cover and land use and demographic change.