Understanding the composition and the changes of New Zealand’s woody vegetation communities is important for effective management. However, past national-scale classification maps emphasised mature rather than seral vegetation communities and forests were mapped at relatively coarse spatial resolution. The integration of Sentinel-2 (S-2) and PlanetScope (PS) satellite imagery provides an opportunity for forest mapping with low cost and high accuracy by combining the high spectral resolution of S-2 with the high spatial resolution of PS. In addition, describing phenological changes in spectral response using time- series data may help to distinguish vegetation types, thereby allowing for more specificity within vegetation classification. However, it is not known whether phenological features derived from vegetation indices (VI) are helpful for accurate land cover classification.

The thesis aims to optimise automated species classification for New Zealand’s indigenous forests with advanced remote sensing techniques. The study was undertaken in a podocarp forest in New Zealand’s central north island. Nine land cover classes were classified, including conifer, low-layer vegetation, broadleaf evergreen, highland softwood, wetland vegetation, water, dead tree, lowland softwood, and low- density vegetation and bare soil. The current study started with the evaluation of machine learning methods and integrated S-2-PS imagery, followed by including phenological features in the classification procedure and finally analysing the contribution of phenological features in detail.

Chapter 2 investigated the feasibility of the integrated image for detailed forest mapping. Free satellite data (S-2, PS, fused data) were compared with commercial data (WorldView-2, and WorldView-2 resampled to S-2 and PS spatial resolutions) by conducting pixel-based classifications with three machine learning classifiers (Support Vector Machine radial basis function kernel, Random Forest, Artificial Neural Network). Spectral features (single bands and VIs), textural features, and an 8-m resolution digital terrain model (DTM) were used in classifications; the relative importance of these input features was also assessed. It was found that the overall classification accuracy was dependent on the combination of classifier and imagery, with different combinations resulting in a range of accuracies between 66.9% and 95.6%. The integrated PS and S-2 product has a higher spatial resolution than S-2. It also achieved the best overall accuracy (95.6%), which was even greater than that of WorldView-2 (95.1%). The DTM was the most important feature for all scenarios; Gray-Level Co-Occurrence Matrix-Mean was the most important texture variable for WorldView-2 and integrated images. Original bands, as well as Greenness Index (GI), Normalised Green (Norm-G), and Simple NIR and Red Ratio (SR-NIRR), were also crucial for vegetation classification.

Chapter 3 tested the potential of phenological features in the classification of native vegetation. Phenological features, based on S-2 imagery, were extracted, including the enhanced vegetation index (EVI), enhanced vegetation index 2 (EVI2) and normalised difference vegetation index (NDVI). Google Earth Engine (GEE) harmonic analysis and TIMESAT double logistic fitting function were used to extract phenological features. Pixel-based classifications were performed using a Neural Network on six different scenarios. The accuracy of the classification scenarios was determined, and the importance score for each feature was evaluated. Using only the fused PS and S-2 bands, the land cover in the study area was classified with 90.1% accuracy. Adding phenological features increased the classification accuracy to 93.1%. When combined with vegetation indices (VIs), texture features, and a DTM, the addition of phenological features increased the classification accuracy to 96.6%. Including GEE-generated phenological features resulted in better classification accuracies than TIMESAT features. In terms of feature importance evaluation, EVI2- and NDVI-generated phenological features all had high importance scores; the effectiveness of EVI features could potentially have been limited by the quality of the blue band. The results demonstrate that it is possible to produce a more accurate classification of New Zealand’s native vegetation by using phenological features.

Changes in VIs over time can describe vegetation phenology; however, it is unknown which phenological features, derived from VIs, contribute most to land cover classification. Feature selection could potentially solve this problem. In Chapter 4, phenological feature importance and feature selection were tested by using S-2 time-series data and PS. PS and S-2 data were fused to a base image with the same spatial resolution as PS and 8 bands containing spectral data from S-2; this image was used to produce 30 VIs. Phenological features – amplitude (AMP) and phase (PH) were extracted from these 30 VIs using harmonic analysis in GEE. For the purpose of accurately classifying forests and identifying the most important phenological features, three classification scenarios (fused bands & VIs, fused bands & phenological features, fused bands & VIs & phenological features) were developed using a Neural Network. Variable Selection Using Random Forest (VSURF) was applied to the three scenarios to evaluate the impact of feature selection. Results indicate that VSURF could reduce the time needed to complete the classification while maintaining a comparable level of classification accuracy. Phenological features improved accuracy from 90% to 94%, driven mostly by Red-Edge Triangular Vegetation Index (RTVIcore)-AMP&PH, Normalised Near-Infrared (Norm-NIR)-PH, GI-PH, Water Body Index (WBI-PH), NDVI-PH, Norm-G- PH, Red-Edge Normalised Difference Vegetation Index (NDVIre)-PH, Leaf Chlorophyll Content (LChloC)-AMP, and Simple NIR and Blue Ratio (SR-NIRB)-PH. These features reflect changes in the structure, biochemical, and physiological characteristics of the canopy. A lack of ground-based measurements of seasonal dynamics precluded an evaluation of the accuracy of these phenological aspects and an explanation of their distinctive contribution to the model.

Overall, the findings show that the integrated S-2-PS imagery can improve the classification of New Zealand’s natural vegetation. The cost savings of the integrated approach might prove critical, especially as New Zealand’s indigenous forests cover a vast land area that would require substantial financial investment if it were all captured by commercial satellites. The use of phenological features and feature selection algorithms also offers a potential cost-saving when classifying land cover from satellite imagery, as the platforms for phenological feature extraction, feature selection, and classification are free to use. Using the methods developed in the research, there are opportunities to classify highly-diverse New Zealand’s native vegetation at a finer scale.