Project 1: Cocoa Agroforestry Land Use Classifier

Developed a classifier procedure for discriminating perennial cocoa agroforestry land cover.
Compared a multi-spectral optical image from RapidEye, acquired in the dry season, and multi-seasonal C-band SAR of Sentinel 1.
Hand engineered Grey Level Co-occurrence Matrix (GLCM) texture features (images) from Sentinel-1 Image time series covering six dry and four wet seasons from 2015 to 2017.
Optimised Random Forest Ensemble classifier models using different input feature combinations; multi-spectral reflectance, vegetation indices, co-(VV) and cross-(VH) polarised SAR intensity and GLCM texture measures.
Evaluated accuracy metrics and uncertainty Shannon entropy - information loss in classification.
Built a classifier and uncertainty measures that provide reliable validation of class discrimination at different spatial resolution.

SAR Backscatter Time-series over different land uses

Sentinel-1 SAR GLCM Texture-Based Discrimination of Cocoa Agroforestry land use

Land use classification compared for Sentinel-1 SAR and Rapid-Eye

April 10, 2020

Project 2: Cocoa Agroforestry Canopy Gap Predictor

Estimated canopy cover distribution using in-situ digital hemispherical photographs (DHPs) sampling and estimates of canopy gap fraction.
Built neural network and random forest regression models of cocoa agroforestry canopy gap fraction and Sentinel-1A SAR backscatter intensity and features.
Developed a combination of different backscatter variables for predicting the canopy gap variability in agroforestry cocoa production landscapes.
Utilise a semi-variogram analysis of canopy gap distribution and spatial clustering distances in different cocoa production landscapes.
Provided new insights into the scale of spatial variability of canopy gaps in relation to farm and landscape management through cocoa agroforestry land use.
Built a proof-of-concept to support development of management tools or strategies on tree inventorying and decisions regarding incentives for shade tree retention and planting in cocoa landscapes.

Inventory of canopy cover distribution (ground truth)

Field and spatial data analysis workflow

April 11, 2021

Project 3: Estimator of Tree Spatial Pattern in Agroforests

Assessed the temporal changes in shade management and spatial dynamics in canopy tree structure and live tree biomass.
Conducted field survey along an age gradient of “family farms” in cocoa agroforests created from forest (fCAFS) and savannah (sCAFS) land cover.
Evaluated the temporal changes in farm structure, relative tree abundance, and live aboveground biomass of the major canopy (shade tree) strata.
Used spatial point process and linear mixed effect analyses to assess contributions of associated perennial trees (AsT) on farm rejuvenation patterns.
Provided insights into farmers’ temporal allocation of uses and prioritization of different tree species associations with age of cocoa agroforests.
Developed recommendations for landscape-specific tree management and considerations in proposing on-farm tree conservation incentives.

Spatial sampling and point pattern analysis steps

Spatial pattern of cocoa trees across farm type and age gradients

July 15, 2022

Project 4: Forest (and Tree) Cover Loss from 2001-2021 in Cashew production landscapes

An analysis of trends in Tree and Forest Cover change in cashew production (the project) areas covering three agro-ecological zones.
Ivory Coast (CDI) is ranked second global producer of cashew nuts, after India. In CDI, cashew production contributes an important proportion (1.3%) of national GDP (FAO 2018) and to the livelihood of an important number of small-scale farmers.
However, increasing production of cashew is strongly tied to increases in the production land (Figure 1). Cashew production is carried out in about 19 (of the 33) regions in CDI (Bassett et. al. 2018) and predominantly in the Sudan and Sudano-Guinean agro-ecological zones (FAO 2018), which are largely characterised as fragile ecosystems.

April 14, 2023

Project 5: Spatial estimation of Arid Plant distribution clusters

As a proxy for vegetation photosynthetic activity, vegetation indices (VIs) were estimated from the visible to near-infrared, i.e. VIS-NIR, spectral bands from remote sensing images.
Sentinel-2 Level-2A (S-2A) images we collected using the Google Earth Engine (GEE) cloud computing platform. For now, we have considered scenes within the period from January to December 2021 without cloud filtering. The clouds were later masked and the images co-registered as image collections.
In an exploratory analysis of vegetation cover, 5 different vegetation indices (VIs) were computed and stacked for processing and analysis.
As no ground truth data is yet available for the period of exploratory analysis (2021), we applied an unsupervised image classification procedure was applied.
This classification experiment hypothesised a combination of VIs to provide complementary information and features to enhance thresholding of vegetation and land cover types. Nonetheless, for VI-based classification, it is inappropriate to assume similar vegetation class thresholds for all VI types - the widely applied thresholds may not be optimal for all VIs, vegetation types and context.
The K-means clustering algorithm was applied with 5 clusters selected based on visual observations from the classes obtained by VI thresholding.
On this basis clustering results, we applied five thresholds of potential vegetation classes (< 0.2: No Vegetation, 0.2 – 0.3: Low vegetation, 0.3 – 0.4: Sparse vegetation, 0.4 – 0.5: Moderate vegetation, and > 0.5: Moderate-High vegetation).
Image thresholding procedure and unsupervised classification (image clustering) were conducted using the GEE cloud computing platform and the resultant images exported as Geo Tiff files for further (supervised) analysis.