Integrating Sentinel-1/2 and machine learning models for mapping fruit tree species in heterogeneous landscapes of Limpopo

Abstract

From ancient times to this century, Africa has relied chiefly on agriculture for survival. Crop type maps are crucial for agricultural management, sustainable farming systems, and realizing food security. Agronomists, agricultural extension officers, policymakers, and the government rely on crop type spatial distribution information to make informed decisions and optimize resource allocation for sustainable agricultural management. Attaining food security for all is an urgent need in Africa. However, the farming landscapes predominately comprise fragmented smallholder heterogeneous farms. The farming systems include intercropping and cultivating different crops that require different management strategies. This results in within-class spectral similarities and intra-spectral variability due to similar canopy structures and different phenologies, which complicates the application of remote sensing in crop type mapping. The free availability of Copernicus products such as Sentinel 1 and 2 have high temporal, spectral, and spatial resolution suitable for mapping smallholder agriculture. Thus, this research aimed to integrate Sentinel-1/2 and machine learning models for mapping fruit tree species in heterogeneous landscapes of Limpopo. First, the research tested the applicability of sampling techniques and five mapping classifiers (i.e., Random Forest (RF), Support vector Machine (SVM), Adaptive Boosting (AdaBoost), Gradient Boosting (GB), and eXtreme Gradient Boosting (XGBoost) in mapping fruit trees and co-existing land use types. The original dataset was under-sampled randomly into two balanced datasets (i.e., Dataset 1 and Dataset 2) consisting of 100 and 150 sample points. Furthermore, the imbalanced ratio from the original dataset was reduced by applying different sampling strategies to extract four imbalanced datasets (i.e., at 40%, 50%, 60%, and 70%), which resulted in the formation of Dataset 3, Dataset 4, and Dataset 5, respectively. These samples, together with the original dataset (i.e., Dataset 7), were used as input to Sentinel‑2 (S2) data using adaptive boosting (AdaBoost), gradient boosting (GB), random forest (RF), support vector machine (SVM), and eXtreme gradient boost (XGBoost) machine learning algorithms. The results showed that reducing the amount of imbalanced ratio by randomly under-sampling the original imbalanced dataset could increase the classification accuracy to 71% using the SVM classifier and 60% of the original dataset. Individually, the majority of the crop types were classified with an F1 score of between 60% and 100%. Secondly, the research independently assessed the effectiveness of Sentinel-1 (S1) and Sentinel-2 (S2) data for fruit tree mapping using random forest (RF) and support vector machine (SVM) classifiers. Four models were tested using each sensor independently and fusing both sensors. From the fused model, features were ranked using the RF mean decrease accuracy (MDA) and forward variable selection (FVS) to identify optimal spectral windows to classify fruit trees. The best fruit tree map with an overall accuracy (OA) of 0.91.6% with a kappa coefficient of 0.91% was produced using the RF MDA and FVS model and SVM classifier. The application of SVM to S1, S2, S2 selected variables and S1S2 fusion independently produced OA = 27.64, Kappa coefficient = 0.13%; OA= 87%, Kappa coefficient = 86.89%; OA = 69.33, Kappa coefficient = 69. %; A = 87.01%, Kappa coefficient = 87%, respectively. The green (B3), SWIR_2 (B10), and vertical horizontal (VH) polarization bands were identified as the optimal spectral features for S2 and S1 data, respectively. The third part of the research identified the optimal growth window period in which fruit trees can be detected with high accuracy. Phenological metrics were extracted from 12 months (i.e., January to December) of Sentinel-2 (S2) data and were used to classify fruit trees using a random forest (RF) classifier in a Google Earth Engine environment. The results showed that fruit trees can be detected and mapped with high accuracy during winter months (i.e., April-July) with an overall accuracy (OA) of 84.89% and a kappa coefficient of 83%. The user accuracy ranged from 62 to 100%, while the producer accuracy ranged from 60 to 100%. The fruit trees were mostly differentiated from co-existing land use types using the short infrared and the red-edge bands. The fourth part of the thesis attempted to increase fruit tree classification accuracy by classifying optimal Sentinel-2 images acquired during the fruit trees' critical growth stages using a Deep Neural Network (DNN) model. This was achieved by applying phenological metrics derived from Sentinel-2 images acquired during optimal crop-growing seasons (i.e., flowering, fruiting, harvesting). The DNN models were optimized by tuning the hyperparameters to achieve the best classification results. The DNN produced an OA of 86.96%, 88.64%, 86.76%, and 87.25% for April, May, June, and July images, respectively. The results indicate the DNN models were robust and stable across the selected fruit growth periods. This research has shown that earth observation (EO) data such as Sentinel 1 and 2 can be used to map fruit trees in fragmented sub-tropical horticultural landscapes characterized by different environmental conditions and different crop cultivars operating under different management practices. The research results will assist agricultural stakeholders (i.e., farm managers, agronomists, agricultural extension officers, and policymakers) in allocating agricultural resources, devising effective agricultural management strategies, and attaining sustainable agriculture and food security.