1. Introduction
Agricultural production is critical for growth in many developing economies [
1] and is indispensable for food security in sub-Saharan Africa (SSA). The world’s population is gradually increasing and projected to reach 9 billion by the year 2050 [
2]. Considering that approximately 815 million people in the world are chronically undernourished [
2], it is imperative to address extant food insecurity challenges by increasing agricultural production (to the tune of 50% more) to feed the growing population [
2]. Yet, the need to increase production is often constrained by resource limitations, production inefficiencies, and natural/human threats in many smallholder farming systems [
3]. Globally, about one-third (38%) of the terrestrial surface is classified as agricultural land and pastures [
4], and further expansion of both land-use types is expected to cause negative ecological impacts like biodiversity loss and deforestation [
5]. Therefore, sustainable management of current cropland areas is important to improve productivity and address yield gaps [
4]. However, this requires periodic monitoring of cropland area and composition to understand changes and develop or deploy appropriate agronomic tools and interventions for farm-level decision-support.
Generally, mapping of croplands includes the identification of crops and their areal coverage to generate relevant agricultural statistics (at various geographical scales), support yield forecasting, and assess agroecological/environmental changes [
6]. In most SSA countries, conventional methods such as periodic agricultural surveys are adopted to map or quantify cropland areas, usually at irregular frequencies. Such methods are cost-prohibitive due to extensive and repetitive field surveys, mostly conducted on smaller areas/units of lands, and based on biased sampling techniques which may skew area estimates [
7]. However, current remote sensing tools and technologies offer capabilities for the rapid and cost-effective acquisition of data to continuously map cropland changes at varying scales [
7,
8,
9].
In Africa, about 80% of farmlands are cultivated by smallholder farmers, and in East Africa, these farmers account for 75% of the agricultural outputs [
3]. Most of the smallholder farms are characterized by mosaic landscapes and mixed crop farming on small pieces of land, thereby posing a significant challenge to the monitoring of changes in these agricultural landscapes [
10]. The increasing accessibility of high-resolution remotely sensed data is promising for mapping and monitoring vegetation growth and dynamics in these smallholder farms. Optical sensors that are coupled with unmanned aerial vehicles (UAVs) are emerging as a reliable source of high-resolution and multispectral imageries that can be used to distinguish crops within and between fields [
11,
12,
13]. Moreover, the imagery reflectance data can be processed to generate vegetation indices which can further aid crop identification and differentiation [
14]. Specifically, the Normalized Difference Vegetation Index (NDVI) [
15], Green Normalized Difference Vegetation Index (GNDVI) [
16] and Enhanced Vegetation Index (EVI2) [
17], have proven useful for various agronomic use-cases [
18,
19,
20]. In addition, relevant insights can be derived from UAV-acquired imagery data by applying advanced analytics, including machine learning techniques to achieve high prediction accuracy of various parameters for agronomic decision support.
Many researchers have utilized UAV-acquired data for crop and yield mapping in smallholder farming systems, in combination with various remote-sensing and machine learning tools/methods [
18]. Chew et al. [
21] mapped croplands in Rwanda using a deep convolutional neural network (CNN) on UAV-acquired imagery and achieved high classification accuracy (up to 96%) for staple crops like banana and maize compared to legumes (49%), which are mostly cultivated under intercropped conditions [
21]. In Ghana, UAV-acquired RGB and near-infrared (NIR) spectral bands (and calculated vegetation indices) were used to delineate and map smallholder maize-cultivated farms with an accuracy of 94% [
22], while several imagery products (including thermal imagery, multispectral band, and vegetation indices) were used to monitor crop growth within farmers’ fields in the Czech Republic [
23]. The methods applied in these various studies differ as well. For instance, croplands in Zimbabwe were mapped by implementing automatic classification of UAV imageries, including ensemble classification methods and decision-level fusion (and NDVI thresholding) to identify the croplands and determine spatiotemporal cropland changes [
24]. As explained by Lu and Weng [
25], in addition to the right imagery, the right choice of classification methods is crucial to successfully map land cover. Machine learning methods, supervised or unsupervised, have been widely used for land-cover classification studies using remotely-sensed data [
26,
27,
28]. For instance, support vector machine (SVM) has gained much attention [
29] and has been applied widely in land cover classification using imageries with highly reliable outputs [
29,
30,
31]. Phan and Kappas [
32] compared land cover classification with k-nearest neighbor (KNN), random forest (RF), and SVM, and reported that higher overall classification accuracy and least sensitivity to training size samples was achieved with SVM. Other studies [
28,
33,
34] have shown that RF is also promising for land-cover classification. Rodriguez-Galiano et al. [
33] showed that high (>90%) overall accuracy was achieved (kappa = 0.92) by applying RF in a land cover classification of a complex area to classify 14 different land categories. While each of these classification techniques has unique strengths under specific application contexts, they can perform poorly in other contexts, and it is often important to compare performance across classifiers relative to the specific classification task.
In Rwanda, Banana is one of the important staple crops for food and nutritional security in the country, and it supports household livelihood as a source of income [
35]. However, Banana production is threatened by the Banana Xanthomonas Wilt (BXW), an infectious crop disease that can cause up to 100% yield loss per infected stand [
36,
37], and spatially explicit data about areas of banana production are generally outdated or non-existent. The identification and delineation of cropland area is a critical first step towards targeting, controlling, and preventing crop-specific diseases, such as BXW, nationally. Similar to other sub-Saharan African Countries, national estimates about the location and areas of banana farms in Rwanda are predominantly estimated or extrapolated based on traditional survey and hierarchical reporting, which are often conducted post-season, at irregular intervals, and usually at sparse locations [
9,
38,
39]. Rapid and spatially explicit assessment of banana farms at high-resolution and frequent intervals can support the need for timely extension delivery in banana production systems and support food security in Rwanda by providing reliable data at useable granularity for local action and interventions. The combination of high-resolution imageries, such as from new-generation satellites and UAVs, and evolving analytical methods, such as machine learning classification techniques, are promising to generate (near) real-time outputs and insights for timely decision-support. Yet, due to varying geographical contexts and the cost of collecting ground-truth data from representative locations, methods for classifying croplands should be evaluated for their reliability or relative accuracy [
40], especially to advance credible national monitoring systems. For instance, the classification accuracy of supervised classification methods can be affected by the volume, dimensionality, and quality of the dataset [
41]. Further, depending on the desired level of confidence in the classification process, the usefulness/reliability of landcover classification outputs can be determined by the robustness of the ground-truth training data. Limited research exists to guide the understanding of the impact of differing levels of sample points on classification accuracy, however, such assessment can generate useful information to guide the selection of the most reliable landcover classifier that can be reliable for national mapping purposes in Rwanda. Therefore, this study was conducted to assess the accuracy of multi-classifier models for mapping of banana landcover at a village level, and to evaluate if a combination of ground-truth georeferenced data within digitized point datasets from high-resolution UAV imageries can enhance the classification outcomes based on improved data robustness.
4. Discussion
The rapid classification of land cover with UAV-acquired imagery across multiple villages can enhance timeliness and accuracy of cropland area estimates, especially to advance decision-support for the specific crop (such as banana) or farming systems. Although this research was conducted as a study focused on four banana-producing villages in Rwanda, it demonstrates the potential to adapt UAV as a tool to support the national framework for agricultural assessment. The availability of spatio-temporally rich and crop-specific map products, based on high-resolution UAV-acquired imageries and reliable classification routine, can support the delivery of tailored recommendations and extension support to farmers, depending on the status and dynamics of their cropland conditions.
Access to robust ground-truth data is indispensable for the implementation of training and validation of the classification models. However, ground-level data points are rarely available at sufficient volume to fully calibrate and test models, and this study is not an exception. For our study, it was noteworthy that the accuracies of classifier models were comparable at different data richness levels, notwithstanding the general differences in the performance of the models relative to each other. The initial 750 georeferenced data points that were collected from banana farms provided a good basis for implementing the classification routine. Yet, the high-resolution orthomosaic UAV images (i.e., RGB band combination) offered a major advantage to support manual digitization of banana farms and generate more georeferenced data points, based on visual identification of the banana stands and canopy. This flexibility to generate additional data points (up to 3x the initial feature points) enabled the testing of hypotheses regarding the performance of ML classifiers at a different level of data richness, to evaluate the potential trade-off between these methods, as a guide for future and larger scale classification within similar/the same geography. The accuracy of RF and CART increased with increasing training data samples, although by a smaller margin. The random forest model performed well, on average above 96% accuracy, for all data levels tested. This implies that while in general better model performance can be expected with more ground-truth data, the performance of classification models tends to flatten out when a sufficient number of the data points is reached. Modelers need to strike a balance between model performance and the cost of data collection; model performances can be optimized at moderate data richness, with careful parameter tuning, for effective mapping of banana croplands in our case. This understanding can help to evolve sampling strategies to minimize the difficulty associated with acquiring ground truth data in heterogeneous landscapes for future national banana mapping projects.
Other studies [
21,
22] have utilized UAV-acquired imageries and object-oriented modeling techniques to map croplands, with varied accuracies and outcomes. In our efforts to identify a suitable classification model, it was clear that the classification accuracy metrics contrasted between the different classifier models. Our results suggest that the random forest model can be effective for identifying banana cropland and agree with previous research where the model has been reported to be promising for cropland mapping [
11,
21,
22]. The least accurate model, the SVM, may have performed very poorly compared to RF and CART due to the model’s training complexities. SVM classified some banana farms as other vegetation because they have similar spectral characteristics. Since the model constructs hyperplanes in a multi-dimensional space to classify the target data [
29,
48], it may have performed poorly when the target classes in the data overlap, including shades of green color that are characteristics of banana and other vegetations. Generally, CART and RF were expected to perform at similar accuracy levels for distinguishing the landcover classes, considering that RF is an ensemble of decision trees. However, the accuracy metrics of both models which shows that RF performed better than CART may be indicative of the inherent limitation of CART, which is prone to overfitting by penalizing the fitted model to minimize the training noise at the expense of the overall data [
17]. This problem is addressed in the RF model, which minimizes bias-related overfitting of the data, specifically by recognizing the nuances of vegetation characteristics and accounting for the high dimensionality of the data [
33,
34].
The successful delineation and estimation of the current banana-cultivated land area within each village provide additional evidence regarding the potential application of UAV for rapid assessment of cropland area for local or national planning and agronomic decision-support. It is noteworthy that the estimated area of the banana-cultivated land area is comparable to the current village-level estimates reported in national agricultural statistics. This rapid mapping and assessment of the banana-cultivated area in Rwanda is important to target extension resources within critical banana production areas, and provide useful data that can further support the mitigation of banana disease (including BXW) risk by providing timely targeted resources for control at a local, regional, and national scale. For instance, the risk of BXW contagion and incidence is linked to the presence and density of the host plant, banana. Therefore, clusters of farmlands that are cultivated with banana can be identified more easily with the landcover maps and relevant extension (information and personnel) resources can be deployed to support the farmers both for proactive and reactive actions to minimize the potential or existing threat of the disease.
To fully unlock the opportunity for rapid mapping of banana croplands, UAV-derived imagery can be fused with satellite data for national-scale mapping of banana croplands. Therefore, the future research direction can go beyond the spectral-based classification by including texture-based and object-oriented algorithms to extract more information about the unique features of banana farmlands. Also, it will be relevant to assess how classified land-cover data can be combined with other ancillary data (e.g., weather, slope, soil) to implement specific agronomic use-cases at the local and national level, such as prediction of (BXW) disease incidence and risk in Rwanda’s banana production system.