UAV remote sensing of spatial variation in banana production

Brian L. MachovinaA,C, Kenneth J. FeeleyA, and Brett J. MachovinaB

ADepartment of Biological Sciences, Florida International University, Miami, FL 33199, USA; and The Fairchild Tropical Botanic Garden, Coral Gables, FL 33156, USA.

BDepartment of Economics and Geosciences, United States Air Force Academy, CO 97331, USA. CCorresponding author. Email:

Abstract. Remote sensing through Unmanned Aerial Vehicles (UAV) can potentially be used to identify the factors influencing agricultural yield and thereby increase production efficiency. The use of UAV remains largely underutilised in tropical agricultural systems. In this study we tested a fixed-wing UAV system equipped with a sensor system for mapping spatial patterns of photosynthetic activity in banana plantations in Costa Rica. Spatial patterns derived from the Normalised Difference Vegetation Index (NDVI) were compared with spatial patterns of physical soil quality and banana fruit production data. We found spatial patterns of NDVI were significantly positively correlated with spatial patterns of severalmetrics of fruit yield and quality: bunchweight, number of hands per bunch, length of largestfinger, and yield. NDVI was significantly negatively correlated with banana loss (discarded due to low quality). Spatial patterns of NDVI were not correlated with spatial patterns of physical soil quality. These results indicate that UAV systems can be used in banana plantations to help map patterns of fruit quality and yield, potentially aiding investigations of spatial patterns of underlying factors affecting production and thereby helping to increase agricultural efficiency.

Additional keywords: crop productivity, Musa, NDVI.

Received 12 April 2016, accepted 3 October 2016, published online 23 November 2016


Bananas (Musa acuminata) are the developing world’s fourth most valuable food crop (Frison et al. 2004) and globally are the 12thmost important plant cropbyvalue andquantity.Worldwide, over 100 megatons (Mt) of bananas are grown annually on an estimated area of ~5million ha (FAOSTAT 2014). Export production, with a volume exceeding 15 Mt and an estimated value of ~US$5 billion annually, is concentrated primarily in Latin America, where over 80% of banana exports originate (FAO 2009; Robinson and Sauco 2010; Evans 2012). Costa Rica is the world’s second largest exporter of bananas (FAOSTAT 2014).

Banana cultivation involves many financial costs and requires extensive use of expensive agrochemicals as nutrient sources and biocides, often causing downstream environmental effects (Worobetz 2000; Marín et al. 2003; Astorga 2005). Better understanding of the spatial patterns of variables that determine banana production could potentially lead to increases in yields (Cassman 1999; Mueller et al. 2012), decreases in production costs per unit yield, and increase profits. An important strategy for improving agricultural productivity and food security is utilising new technologies to gather information on crop ecology that can help better direct management decisions (Gebbers and Adamchuk 2010; Foley et al. 2011). As a core element of precision agriculture, remote monitoring of crop photosynthesis and yields can reveal patterns of stressors affecting crops, enabling managers to adjust treatments to

specifically target threatened or affected areas while avoiding treating areas not requiring attention.

Remote-sensing platforms with sensors for measuring electromagnetic reflectance patterns from vegetation offer opportunities to identify geographic patterns of crop stressors and can be used to help investigate underlying causes of stress and improve the agricultural management decision-making process (Jackson 1986; Plant 2001). Ground-based sensors, as well as sensors mounted on satellites andmanned airplanes, have been used to monitor a variety of parameters in managed and natural systems; parameters measured include water stress (Jones 1999; Takács et al. 2006; Tamás and Lénárt 2006; Jones and Schofield 2008), pest damage (Nutter et al. 2002; Prabhakar et al. 2011; Hillnhütter et al. 2012), and disease (Pozdnyakova et al. 2002; West et al. 2003; Zhang et al. 2003; Apan et al. 2004; Mahlein et al. 2010), as well as underlying physical variables affecting production, such as leaf area index (Hoffmann and Blomberg 2004; Steltzer and Welker 2006), topography (Florinsky 1998; Hirano et al. 2003), soil quality and nutrient availability (Goel et al. 2003; Apan et al. 2004). Stressors are often visible through remote sensors before the effects can be perceived by the human eye, offering advantages to address problems earlier in their cycle of damage (Jones 2004) and at larger spatial scales. The utilisation of spectral reflection patterns of near-infrared (NIR) and red light are used via the commonly applied normalised difference vegetation index (NDVI) (Rouse et al. 1973) to examine spatial

Journal compilation � CSIRO 2016


Crop & Pasture Science, 2016, 67, 1281–1287
patterns of agricultural productivity patterns (Leon et al. 2003; Tamás and Lénárt 2006). NDVI, which indicates the amount of red light absorbed and NIR light reflected, is closely correlated with photosynthetic activity of plants, and spatial patterns of photosynthetic activity can be visualised as varying levels of NDVI. Increased photosynthesis increases crop yields, and spatial patterns of NDVI early in crop development have successfully been used to predict harvest levels many months later (Leon et al. 2003; Dobermann and Ping 2004; Zarco-Tejada et al. 2005).

Small Unmanned Aerial Vehicles (UAV) are rapidly increasing in popularity as a potential tool for monitoring many agricultural practices (Swain et al. 2007; Knoth et al. 2010; Swain et al. 2010; Laliberte et al. 2011; Turner and Watson 2011; Zhang and Kovacs 2012). UAV that include multi-rotor, fixed-wing, and lighter-than-air (i.e. balloon or kite) platforms (Inoue et al. 2000) can, in some situations, offer advantages of acquiring aerial imagery at lower costs than manned airplanes or satellites with user-friendly methodology such as easier flight training, rapid field deployment, and quick turnaround of image processing, especially when target areas are small and numbers of images are low. Small, lightweight sensor systems can capture NIR and red light, enabling monitoring of NDVI of vegetation by small, low-cost UAV (Tamás and Lénárt 2006; Manera et al. 2010). Their use, however, can be limited by aviation laws, safety concerns, short flight times, weather, or small payload capacity (Hardin and Jensen 2011).

The goal of this research was to determine (1) if variability in banana plantation productivity is related to variation in NDVI, (2) if this variation is driven by variation in soil properties, and (3) if an open-market inexpensive UAV is a useful means of acquiring the NDVI data in a banana plantation.


The UAV system was evaluated for remote vegetation sampling potential in commercial banana plantations located near the city of Rio Frio in Heredia, Costa Rica (108190300N, 838530110W; Fig. 1a). The study area was located at ~100m a.s.l. on flat topography east of the mountain range that runs north–south through Costa Rica. Between 2008 and 2012, the area received a mean annual rainfall of 4900mm (Fig. 1b) and had a mean annual temperature of 258C. The region experiences extensive low altitude cloud cover during much of the year, indicated by a paucity of cloud free satellite imagery available from the LANDSAT platforms (http://earthexplorer., accessed 11 October 2016). The region was dominated by agricultural activities including banana, pineapple, heart of palm (Bactris gasipaes), and tropical ornamental plant cultivation. The UAV system was evaluated during the first week of April 2014.

The harvesting methods in these banana plantations provided a unique opportunity to compare remotely sensed data to banana production data. Bananas are harvested from specific areas along numbered cable lines which vary in length from ~100 to 300m that transport bunches to processing facilities, and several standard measurements of banana fruit production and quality are recorded for each cable line. Approximately every

9 months, a banana plant produces a single bunch, which comprises 5–10 hands and which each produce 10–20 bananas (fingers). Typically, the area of harvest encompassed ~50m on each side of a cable line. In this study, we compared remotely sensed data to six banana fruit productionmeasurements: number of boxes produced per ha (one box = 44 kg), mean weight of a bunch, mean loss (proportion of bananas discarded from packing due to unacceptable quality), mean number of hands per bunch, mean size of largest banana per bunch, and the mean thickness of a banana on the second hand. Production variables were provided as totals or averages from 4-week periods. The mean value per cable line for each variable that was compared with remotely sensed data was calculated as the mean of the combined values recorded during the 13 4-week sampling periods of 2013 and the first 6 recorded 4-week sampling periods of 2014, providing a mean value from 76 weeks of production data.

Supplied by MarcusUAV, Inc. (, accessed 11 October 2016), the fixed wing UAV system (Fig. 2) was a 2.5-kg delta-wing design with a 175-cm wingspan, powered by two 2700 mAh, 14.7 v, 4-cell LiPO batteries. Manual flight control during takeoff and landing was




3786 Meters


–16 Meters

100 Kilometers500







R ai

nf al

l ( m

m )

0 1 3 5 7 9 1113 15 1719 21 23 25 27 29 3133 35 37 39 4143 45 4749 51

Fig. 1. (a) An elevationmap indicating study area in relation to San Jose, the capital of Costa Rica, and (b) weekly mean rainfall beginning 1 January (2008–2012) at the study area.

1282 Crop & Pasture Science B. L. Machovina et al.
performed with a Spektrum DX8 RC controller. Mission planning and automated flight control was performed using Mission Planner 1.22.99 on a laptop computer, relayed via a ground-based radio-modem antenna. A small video camera mounted in the nose of the UAV relayed live video footage of the flight path to a ground-based tracking antenna. All automated flight operations and video processing were managed via a single laptop computer connected to the antenna system. Flight plans were made creating survey grids using the Auto Waypoint and Polygon tools in Mission Planner on imagery downloaded from Ovi Satellite Maps, which provided better high-resolution coverage of the region than Google Maps. Takeoff and landings were performed via manual control, but image-capture flight patterns were under automated control by the flight control software.

The fixed-wing UAV was outfitted with a 90-g Tetracam ADC Micro (, accessed 11 October 2016), which was mounted on a motorised roll stabiliser. The Tetracam Micro captures Near-Infrared, Red, and Green wavelengths similar to Landsat Thematic Mapper bands TM2, TM3 and TM4. Wavelengths recorded are Infrared: 760–900 nm (recorded on red channel), Red: 630–690 nm (recorded on green channel), and Green: 520–600 um (recorded on blue channel). The system has a 3.2-megapixel resolution (2048� 1536) sensor and a fixed 8.43-mm lens. Images were stored on 16-GB removable storage cards. Geographic locations of camera trigger points were recorded by the Tetracam from the UAV’s flight controller GPS.

Prior to the flights, images of a white Teflon calibration plate were recorded with the Tetracam under ambient light

(a) (b)


Fig. 2. Fixed Wing UAV showing (a) approximate size, (b) antenna system for location tracking and live video capture, and (c) UAV mounted on launcher.

UAV remote sensing of bananas Crop & Pasture Science 1283
conditions. The UAV was launched from a dual slide-rail launcher constructed from PVC piping and powered by a 15- m, triple-cord bungee line staked into the ground ~30m in front of the UAV. A foot operated trigger released the UAV. Launches were performed from an athletic field located within 0.5–2 km of the origins of the onset of imagery capture. Landings occurred at the same location as launches, and were achieved via manual triggering of a parachute deployment or by manually slide landing the UAV on the grassy field. Three flights were performed, reaching 260-m altitude image capture elevation, traveling at 16m/s, lasting from 20 to 22min, flying linear distances of 11.7 km, 16.4 km, 16.5 km and recording imagery covering 165 ha, 186 ha, 164 ha respectively. Images were recorded with ~60% forelap and 40% sidelap and a pixel resolution of 10 cm.

Post-flight images were transferred to a laptop and visually sorted to remove takeoff/landing images lower than 260-m altitude and blurry images. Images were processed into false- colour infrared images and NDVI classified images using the Teflon standard images and Pixel Wrench, the image processing software supplied by Tetracam. Using Agisoft Photoscan Professional, we attempted to mosaic and orthorectify images from each of the flights, but only the second flight provided sufficient image quality and overlap to enable the creation of a quality mosaicked single image using automated methods of the software. All banana production data comparisons were performed on data extracted from the mosaic from this flight.

The orthorectified mosaic of the flight was imported into ArcGIS. A vector map indicating locations of cable lines, supplied by growers, was also imported. A total of 23 cable lines with active banana production areas were identified. Along each of these cable lines, 20 locations were identified visually for sampling NDVI values from the NDVI mosaic. NDVI was calculated as (NIR – R)/(NIR +R). Twenty sample locations along each cable line were sampled via a stratified random sampling method by dividing each side of a cable into 10 approximately-equal-sized zones and randomly selecting the approximate centre of one of the four quadrants in each zone. At each sampling location, the closest 10-m-diameter (78.5m2) circular area that covered only bananas (no roads, paths, canals, or other vegetation types) was selected and the mean NDVI value for the circular area was calculated. The mean NDVI value for a cable line was calculated as the combined mean of pixels in all 20 sample location areas along each cable line. A vector map indicating locations of samples for determining soil classifications, supplied by growers, was also imported. These classifications were made based on soil core samples previously made by growers at the specific locations. Soils at sample sites were classified on a four-tier scale (I-IV) of most to least favourable classes, respectively, for banana cultivation based on physical soil characteristics including texture, structure, portion of coarse fragments, consistence, and drainage. At each soil core sample location, a 10-m-diameter (78.5m2) circular area was selected from the NDVI mosaic. Only soil sample

100 m



Order now and get 10% discount on all orders above $50 now!!The professional are ready and willing handle your assignment.