Application of unmanned aerial systems for crop discrimination in smallholder farms.

Abstract

Agriculture is the cornerstone of global food security, serving as humanity’s principal source of sustenance and the primary supplier of critical crops. A crucial challenge facing society is ensuring food security for a rapidly growing population, projected to exceed nine billion by 2050. With limited opportunities to expand arable land, improving agricultural productivity has become indispensable to meet escalating global food demand. Thus, there is a need for robust, precise, holistic agricultural intelligence systems to monitor and optimise crop production. This is particularly so, in regions characterised by heterogeneous smallholder farming systems dominated by mixed cropping. The ability to identify and monitor individual crop types is fundamental for optimising resource allocation, informing targeted interventions, and ultimately enhancing agricultural productivity. Unfortunately, the use of traditional groundbased methods for crop identification has been deemed labour-intensive, time-consuming, and spatially limited, rendering them inadequate for large-scale or frequently updated crop assessments. Thus, the efficacy of remote sensing technologies has been proven in acquiring synoptic and multi-temporal data that is crucial for agricultural monitoring and management. Among the suite of remote sensing platforms, unmanned aerial systems (UASs) have garnered significant attention due to their capacity for near-real-time data acquisition at high spatial resolutions. Equipped with increasingly sophisticated yet miniaturised and lightweight sensors, UAS offers a flexible and cost-effective alternative to traditional aerial and satellite imagery, particularly for localised agricultural applications. The advancements in geospatial technologies have facilitated critical data collection on various farm tasks, with crop discrimination and classification as key focus areas. However, despite the evident potential of UAS in agriculture, several bottlenecks have been identified, including lack of comprehensive information regarding optimal UAS configuration, sensor characteristics tailored for specific crop discrimination, and robust data processing and analytical methodologies applicable across diverse cropping systems. Given this background, this study sought to (i) systematically review the current state, challenges and opportunities in the application of unmanned aerial systems for crop discrimination, (ii) determine the optimal field parameters, specifically the number of crop species and crop row widths, that facilitate accurate crop discrimination and (iii) develop techniques that distinguish crop types in a mixed cropping setting, owing to the flexibility and cost-effectiveness of UASs particularly for localised agricultural applications. This thesis addresses the overarching challenge of achieving accurate and reliable crop discrimination, explicitly focusing on the prevalent mixed-cropping system of maize (Zea mays) and soybean (Glycine max), which are of significant economic and nutritional importance in regions of sub-Saharan Africa. The thesis focused on multiple investigations employing a range of remote sensing data modalities and analytical techniques to tackle the complexities inherent in distinguishing spectrally and structurally similar crops within heterogeneous agricultural environments. The first objective was to systematically review the current state, challenges and opportunities in the application of unmanned aerial systems for crop discrimination. This was followed by examining the spectral separability of maize and soybean across different growth stages using hyperspectral data. Thirdly, the thesis evaluated the utility of spectral, textural and morphological features derived from UAS-based RGB imagery to distinguish maize and soybean from other objects. This was followed by developing a novel technique for shadow detection in RGB datasets. Lastly, the thesis developed a hybrid approach by integrating segmentation and pixel-based classification to provide a comprehensive understanding of effective remote sensing strategies for enhanced crop discrimination in mixedcropping systems. The ultimate goal is to contribute to advancing precision agriculture practices, particularly in resource-constrained settings, where accurate and timely information on crop distribution is paramount for sustainable agricultural development. This explores different facets of crop discrimination in mixed-cropping systems, which are a characteristic of smallholder farming systems in most developing countries. The second objective sought to investigate the spectral separability of maize and soybean at different phenological stages based on field experiments. To achieve this, hyperspectral data spanning the visible to near-infrared spectrum (400–1100 nm) were employed to evaluate the spectral signatures of these two crops across five critical growth stages. The integration of statistical analysis (ANOVA), distance (Jeffries-Matusita distance) and divergence metrics (Transform Divergence, Kullback-Leibler Divergence), and machine learning algorithms (Partial Least Squares-Discriminant Analysis (PLS-DA)) provided a robust framework for optimising band selection and identifying critical phenological stages for discrimination. The key findings of this study revealed that peak spectral separability occurred during the reproductive stages (85– 110 days after planting), with the red spectral region (600–700 nm) exhibiting maximum divergence, attributed to differences in chlorophyll dynamics. Notably, PLS-DA achieved nearperfect classification accuracy (100% F1-score) at the mid-grain filling stage (DAP 85), highlighting the efficacy of leveraging red-edge (680–750 nm) and near-infrared (700–1100 m) bands during this period. Conversely, minimal separability was observed during early vegetative stages due to spectral overlap. This research underscores the need to consider phenological timing and specific spectral regions for effective crop discrimination using hyperspectral data, offering valuable insights for designing targeted remote sensing surveys. The third objective leveraged the increasing accessibility and affordability of UAS-based RGB imagery to evaluate the utility of spectral, textural, and morphological features for distinguishing maize and soybean in a mixed-cropping environment. High-spatial-resolution RGB images were captured during the tasselling stage of maize (48 days after planting) using a DJI Matrice 300 drone. Due to persistent cloud cover and rainfall during the summer, data acquisition was constrained; consequently, only DAP48 observations were obtained during the reproductive and maturity stages. By extracting a comprehensive set of 26 variables encompassing spectral indices, textural features, and morphological transformations, the study employed a random forest (RF) algorithm for supervised classification. The results emphatically demonstrated the superior performance of morphological features, achieving the highest classification accuracy (0.93) and F1-score (92%), followed by a combination of textural and morphological features. Spectral features alone proved to be the least effective. Morphological features, capturing canopy structure and plant geometry, outperformed spectral and textural traits, highlighting the limitations of spectral-only approaches in mixed-cropping systems. Although these features focusing on the structural and geometrical features of maize and soybean were successful, the results revealed that RGB datasets in smallholder farms were compromised by shadows, which disproportionately increase spectral overlap. To address the problem of shadows that are prevalent due to mixed cropping with varying plant heights, particularly in smallholder farming systems, the fourth objective developed a novel hue-intensity-green-blue (HIGB) difference technique. The performance of this new technique was rigorously compared against established methods (C3 and normalised saturation-value difference index) using RGB datasets from experimentally manipulated maize and soybean mixtures. The HIGB technique, based on the differences between hue and intensity and the green and blue channels, consistently outperformed the benchmark models (C3 and NSVDI) across various shadow conditions, achieving overall accuracies ranging from 77% to 95%. This robust performance, even in scenarios with dark or obscured shadows, underscored the practical utility of the HIGB technique for improving the reliability of crop discrimination efforts using RGB imagery. The HIGB technique performs robustly under varying lighting conditions, nderscoring its value as a critical preprocessing tool for improving crop discrimination. Furthermore, the thesis proposed an alternative light intensity ratio-based (LIRB) approach for shadow removal using RGB imagery. This method is applicable in areas where shadow pixels are sparse; however, it did not fully meet expectations. The approach struggled to reconstruct or eliminate dense shadows, resulting in the introduction of blurry artefacts. These artefacts significantly compromised the overall objective of accurately detecting actual crop acreage within a mixed cropping system. By understanding the limitations of LIRB, the last chapter focused on developing a hybrid classification framework integrating region-based segmentation and pixel-based machine learning. This approach was proposed to tackle the spectral and structural complexity of heterogeneous agro-ecological landscapes by focusing on vegetation pixels only. This method leverages simple linear iterative clustering (SLIC) superpixels to group spectrally similar pixels into meaningful and targeted regions, followed by extracting texture and structural features from these segments. These multi-faceted features were then used to train robust machine learning classifiers: Random Forest and Extreme Gradient Boosting. The experimental results demonstrated remarkably high detection accuracy, with precision, recall, and F1-scores exceeding 0.98 for both classifiers. Feature contribution analysis revealed that mean intensity and standard deviation features derived from SLIC were the most influential, followed by textural and morphological traits. Integrating diverse features substantially reduced error rates from 8% (SLIC-only) to 1% with multi-feature integration, demonstrating the synergistic benefits of combining segmentation, feature fusion, and ensemble learning. This research strongly suggests the benefits of employing such a robust hybrid approach, combining the strengths of segmentation and pixel-based methods and advanced machine learning classifiers to achieve scalable and high-resolution crop mapping in complex agricultural environments. In conclusion, these findings provide actionable strategies for mapping and monitoring crops in smallholder systems, where technical and financial constraints limit multispectral adoption. By prioritising accessible RGB sensors, simple algorithms, and phenological timing, this work supports scalable precision agriculture in developing countries, ultimately aiding food security and sustainable land management. The research highlights the importance of considering the phenological stage and leveraging specific spectral regions, as demonstrated by the hyperspectral analysis. It also underscores the significant role of morphological features derivable from UAS-based RGB imagery for effective crop differentiation. Finally, the proposed hybrid segmentation-classification approach showcases the potential for integrating diverse features and advanced machine learning algorithms for achieving high accuracy in heterogeneous landscapes. The collective insights from these investigations contribute significantly to precision agriculture, offering valuable methodologies and findings that can be further developed and implemented for improved crop monitoring and management, especially in resource-constrained agricultural systems prevalent in regions like sub-Saharan Africa (SSA) and similar environments worldwide. Future research should focus on translating these ground and UAS-based insights to satellite platforms, enabling broader regional scalability while maintaining accuracy in complex cropping systems. Further research could also focus on integrating these diverse approaches, exploring the transferability of these techniques across different crop types and geographical locations, and developing user-friendly tools for practical implementation by agricultural stakeholders.