Estimation of canopy properties in beech forests from an unmanned aerial vehicle equipped with a standard RGB standard RGB camera. Tutor aziendale.
Estimation of canopy properties in beech forests from an unmanned aerial vehicle equipped with a standard RGB camera Master di II livello in Sistemi Informativi Territoriali e Telerilevamento (SITT) A. A. 2014/2015 Tutor aziendale Donatella GUZZI Tutor universitario Leonardo DISPERATI
Candidato Francesco CHIANUCCI
INTRODUCTION Crowns and canopy are the active interface between plant-atmosphere Carbon and nutrient cycling Radiation interception, (photosynthesis, productivity) High sensitivity to disturbance (monitoring and research programs);
Ground measurements of canopy attributes in forestry are challenging Direct measures: destructive, time-consuming, spatially limited Indirect ground methods not suitable for real-time applications, spatially limited
Satellite-based information Spatially extensive information (landscape-global scale); Often not suited for local-regional objectives (spatial and temporal resolution) High-cost per (desired) scene and not profitable revisit time for small-medium scale
INTRODUCTION Recent upsurge in the UAV systems availability Low-cost sensors Technological advances in UAV platforms, sensors and softwares High resolution at small-medium scale Rapid revisit time
Visible and non-visible spectrum sensors available for UAV NDVI and other NIR indices from dedicated cameras Thermal cameras Visible-spectrum indices (GLA, ExG) from metric and commercial cameras
Challenges Applicability of UAV-indices in forestry still at an early experimental stage Small sensor required for many UAVs: reliability of commercial instrumentation? Visible-spectrum indices still not checked from UAV in forestry
OBJECTIVES Test the effectiveness of UAV for estimating forest canopy attributes Small platform (i.e., widest suitability) Fixed-wing (medium scale)
Test visible-spectrum indices ‘Green coordinates’ indices Commercial camera (i.e., widest range of sensors available for UAV)
Comparison with indirect optical methods Calibration with ground indirect optical methods
MATERIAL AND METHODS Test site: Alpe di Catenaia Central Italy 87% forest cover, mainly deciduous forests
10 plots Beech forests 0.5 – 1.0 ha Different stand density (108-3324 trees ha-1)
Field estimates Fisheye photography (DHP)
180° FOV Leaf area index
Cover photography (DCP)
30° FOV Canopy cover
MATERIAL AND METHODS Aerial images Fixed-wing eBee (SenseFly) platform RGB camera Canon ELPH 110 RGB payload
Acquisition Two autonomous flights in July, 2015 (approx. 25’ each) eMotion autonomous flight control Longitudinal overlap: 80% Side overlap: 30% Altitude: 170 m (~7.5 cm pixel resolution)
MATERIAL AND METHODS Orthomosaiced Cropped image Visible spectrum index:
Automated training set
2 2
Vegetation: G>R &G>B & G>25 Non vegetation: GLA ≤ 0 Dark pixels: G ≤ 25 Dubious: residual pixels
CIELAB colour space conversion GLA map a* map (green/red)
b* map (blue/yellow)
MATERIAL AND METHODS Orthomosaiced Cropped image Visible spectrum index:
Automated training set
2 2
Vegetation: G>R &G>B & G>25 Non vegetation: GLA ≤ 0 Dark pixels: G ≤ 25 Dubious: residual pixels
Mean of the training set groups CIELAB colour space conversion GLA map Canopy cover map
a* map (green/red)
b* map (blue/yellow)
MATERIAL AND METHODS Orthomosaiced Cropped image Visible spectrum index:
Automated training set
2 2
Vegetation: G>R &G>B & G>25 Non vegetation: GLA ≤ 0 Dark pixels: G ≤ 25 Dubious: residual pixels
Mean of the training set groups CIELAB colour space conversion GLA map
Canopy cover map
Ground canopy cover VS Leaf inclination angle distribution (measured or assumed)
MATERIAL AND METHODS Orthomosaiced Cropped image Visible spectrum index:
Automated training set
2 2
Vegetation: G>R &G>B & G>25 Non vegetation: GLA ≤ 0 Dark pixels: G ≤ 25 Dubious: residual pixels
Mean of the training set groups CIELAB colour space conversion GLA map
Canopy cover map VS
+ Leaf inclination angle distribution (measured or assumed)
RESULTS
UAV estimates of canopy cover significantly agreed with DCP UAV classification poorly discriminated small gaps UAV estimates are more closer to crown cover estimates from DCP (within-crown gaps are excluded)
Within-crown gaps considered
Within-crown gaps excluded
RESULTS UAV estimates of leaf area index significantly agreed with DHP Accuracy of UAV leaf area estimates is influenced by leaf angle distribution assumption: The commonly assumed spherical distribution (mean inclination: 57°) overestimated leaf area index The planophile distribution (mean inclination: 27°) measured in short trees underestimated leaf area index Best fitting was obtained assuming plagiophile leaf angle distribution (mean inclination: 45°)
DISCUSSION UAV is an effective medium to obtain accurate forest canopy estimates Accurate canopy retrieval in dense canopies High resolution at nadir
Standard RGB cameras produced reliable results CIELAB colour space largely scene-illumination-insensitive Low-cost sensors Suitable for widest range of UAV platforms
Accuracy of UAV leaf area estimates is influenced by leaf inclination distribution Not measurable in nadir direction Difficulty of direct field measures (tall forest canopies) Levelled camera approach suitable (only at accessible height) Levelled camera can be mounted on multirotor UAVs
