Optimal spatial resolution of Unmanned Aerial Vehicle (UAV)-acquired imagery for species classification in a heterogeneous grassland ecosystem

Figures & data

Figure 1. Study area and area selected for classification test. Background is a Quickbird image acquired on 14 July 2013 in true color composition.

Figure 2. UAV system, camera, and a sample photo. Sample photo is NIR–green–blue color composition; vegetation show red color.

Figure 3. Photos showing a more homogeneous (left, photo taken on 11 June 2015) and a more heterogeneous (right, photo taken on 9 July 2015) species composition.

Figure 4. Workflow of imagery processing and multi-scale species classification.

Figure 5. Estimation of scale parameters for multiresolution segmentation of 5 cm image using ESP. The abrupt increases of rate of change (dark grey color curve) indicate the optimal scales for segmentation.

Figure 6. Classification tree built for imagery acquired on 11 June 2015 with spatial resolution of 10 cm. Classes in the grey colored boxes represent classified species. Bold letters show selected features as classifiers; specific threshold values for different features were not shown for the sake of comprehension.

Table 1. Error matrix of classification model for imagery with different spatial resolutions.

	Predicted
Observed	Awnless brome (dense)	Awnless brome (sparse)	Fescue	Goldenrod	Milkweed	Total	Producer’s accuracy (%)
A
Awnless brome (dense)	28	5	0	0	1	34	82
Awnless brome (sparse)	3	34	1	0	3	41	83
Fescue	0	1	42	0	2	45	93
Goldenrod	4	0	0	37	10	51	73
Milkweed	0	0	0	8	35	43	81
Total	35	40	43	45	51	214	–
User’s accuracy	80%	85%	98%	82%	69%	–	82
B
Awnless brome (dense)	33	3	0	5	0	41	80
Awnless brome (sparse)	4	22	0	0	2	28	79
Fescue	0	5	33	0	0	38	87
Goldenrod	1	0	0	41	2	44	93
Milkweed	5	2	0	4	34	45	76
Total	43	32	33	50	38	196	–
User’s accuracy	77%	69%	100%	82%	89%	–	83
C
Awnless brome (dense)	30	7	0	0	2	39	77
Awnless brome (sparse)	2	26	2	0	3	33	79
Fescue	0	4	28	0	0	32	88
Goldenrod	3	0	0	30	6	39	77
Milkweed	0	2	0	0	42	44	95
Total	35	39	30	30	53	187	–
User’s accuracy	86%	67%	93%	100%	79%	–	83
D
Awnless brome (dense)	22	1	0	0	0	23	96
Awnless brome (sparse)	17	11	5	2	6	41	27
Fescue	0	2	22	1	0	25	88
Goldenrod	3	0	0	22	7	32	69
Milkweed	0	1	0	0	42	43	98
Total	42	15	27	25	55	164	–
User’s accuracy	52%	73%	81%	88%	76%	–	73

Tables A–D correspond to imagery acquired on 11 June 2015 with spatial resolution of 5, 10, 15, and 20 cm, respectively.

Figure 7. Comparison of object features (i.e., green NDVI and Standard deviation) for imagery acquired on 11 June 2015 with different spatial resolutions.

Figure 8. Comparison of producer’s and user’s accuracy of classification using imagery acquired on 11 June 2015 with different spatial resolutions, respectively.

Figure 9. Original images (left, NIR–green–blue color composition) and corresponding classification maps (right). Images were acquired on 11 June and 9 July 2015 and with spatial resolution of 10 cm.