Impact of Parcel Size
Figures 14 and 15 show the impact of parcel size on parcel-
based classification without temporal modeling on both sites.
For Site04, when keeping only large parcels (area >3 ha), over-
all accuracy improves by 15%, 5.9%, and 8.7% for radar, opti-
cal, and combined optical/radar attributes, respectively (Figure
14). Indeed, due to the limited spatial resolution of Sentinel-1
images, radar attributes are less robust on small parcel sizes.
When considering parcels larger than >3 ha, radar images
achieve similar results as optical images, with an overall accu-
racy of 79.3%. Finally, combining optical and radar data lead
to better results in all cases, and especially for large parcels.
As for Site77, similarly to Site04, when considering only
parcels larger than >3 ha, the overall accuracy is greatly
improved by radar attributes (+ 8.1% reaching 97.1%), which
confirms that radar images are less robust for smaller parcels.
Discussion
In the following, we will compare the contribution of optical
and Radar Sentinel time series and the impact of parcel size.
We then detail the impact of modeling the temporal structure
and finally explain the impact of the site characteristics on
the results.
Classification Without Temporal Modeling
First, we will compare the results of both sites using parcel-
based crop type prediction based on image observations only
(without temporal modeling).
Optical versus Radar Sentinel Time Series
Depending on the image distribution and the cloud cover, op-
tical data may not lead to good results for crop type mapping.
In our case, temporal missing data interpolation was used,
which led to some uncertainties and decreased the classifica-
tion accuracy. On the other hand, radar data are less robust for
small parcels (
≤
1.5 ha). However, this issue can be reduced
by refining the preprocessing framework of radar data.
Table 7. Site04—Confusion matrices using combined optical and radar attributes.
Ground truth
Corn Barley
O.
cereals Rapeseed Sunflower
O.
oilseeds Protein
Forage
crops Meadows
Fruit
trees Vineyards
Olive
groves
Arom.
crops Vegetables
Classification
Optical and radar without temporal modeling
Corn
31
-
-
-
1
2
-
-
-
-
1
-
-
3
Barley
-
47
9
-
-
-
-
3
37
-
-
-
13
-
O. cereals
-
5
684
-
5
1
-
5
59
-
-
4
25
1
Rapeseed
-
-
1
56
-
-
-
-
10
-
-
-
-
-
Sunflower
-
-
1
-
78
-
-
-
2
-
-
-
13
4
O. oilseeds
1
-
-
-
7
18
-
-
4
-
1
-
2
7
Protein
-
2
6
-
-
-
10
1
7
-
-
-
2
-
Forage crops -
-
18
-
1
-
-
173
239
6
3
9
64
2
Meadows
-
3
16
-
3
1
-
47
1294
-
9
27
77
3
Fruit trees
1
-
1
-
-
-
-
-
52
56
1
9
5
-
Vineyards
-
-
-
-
-
-
-
-
18
-
74
4
19
-
Olive groves -
-
4
-
-
1
-
1
113
1
6
311 26
1
Arom crops
-
-
1
-
-
-
-
-
64
-
2
6
480
2
Vegetables
-
-
2
-
6
4
-
2
18
-
4
6
9
60
Structured optical and radar
Corn
28
-
7
-
1
-
-
-
1
-
-
-
1
-
Barley
-
20 123
-
-
-
-
-
47
-
-
-
-
-
O. cereals
1
1
676
-
2
-
-
-
102
-
-
-
6
1
Rapeseed
-
5
21
35
-
-
-
-
6
-
-
-
-
-
Sunflower
-
-
6
-
67
-
-
1
13
-
-
-
8
3
O. oilseeds
4
-
4
-
7
10
-
-
14
-
-
-
-
1
Protein
-
2
21
-
-
-
-
-
5
-
-
-
-
-
Forage crops 1
-
15
-
3
-
-
25
461
-
-
-
9
1
Meadows
-
1
16
-
1
-
-
2
1456
-
-
1
3
-
Fruit trees
-
-
-
-
-
-
-
-
2
122
1
-
-
-
Vineyards
-
-
-
-
-
-
-
-
3
-
112
-
-
-
Olive groves -
-
1
-
-
-
-
-
3
2
-
458
-
-
Arom crops
-
1
17
-
-
-
-
-
26
-
-
-
511
-
Vegetables
-
-
20
-
10
-
-
-
16
3
-
-
4
58
Table 5. Site04—Effect of temporal modeling on classification
metrics, using combined radar and optical attributes.
Class
No temporal modeling Temporal modeling
F-score
User.
Acc
Proc.
Acc F-score
User.
Acc
Proc.
Acc
Corn
0.89 0.95 0.83 0.78 0.84 0.73
Barley
0.40 0.85 0.26 0.18 0.67 0.11
Other cereals
0.85 0.83 0.86 0.78 0.73 0.84
Rapeseed
0.92 1
0.86 0.71 1
0.55
Sunflower
0.79 0.78 0.79 0.71 0.75 0.68
Other oilseeds
0.57 0.74 0.46 0.41 1
0.26
Protein
0.49 1
0.32 0
0.07 0
Forage crops
0.47 0.73 0.34
0.1
0.88 0.05
Meadows
0.76 0.67 0.87
0.80
0.67 0.98
Fruit trees
0.61 0.86 0.47
0.97
0.96 0.97
Vineyards
0.69 0.74 0.65
0.99
0.99 0.98
Olive groves
0.74 0.82 0.67
0.99
0.99 0.99
Aromatic crops 0.74 0.65 0.86
0.93
0.94 0.92
Vegetables
0.61 0.72 0.53
0.67
0.9 0.54
Table 6. Site77—Effect of temporal modeling on accuracy
metrics, using combined radar and optical attributes.
Class
No temporal modeling
Temporal modeling
F-score
User.
Acc
Proc.
Acc F-score
User.
Acc
Proc.
Acc
Corn
0.94
0.93 0.95 0.88 0.83 0.93
Barley
0.90
0.94 0.86 0.82 0.78 0.85
Other cereals
0.95 0.96 0.94
0.95
0.94 0.97
Rapeseed
0.96 0.97 0.94
0.97
0.98 0.95
Protein
0.95 0.93 0.97
0.95
0.97 0.94
Fiber plants
0.97
1
0.95 0
0.1 0
Forage crops
0
0.1 0
0.70
0.78 0.65
Meadows
0.87 0.81 0.93
0.95
0.94 0.97
Fruit trees
0.01 0.1 0
0.94
1
0.89
Vegetables
0.89
0.91 0.88 0.45 0.97 0.30
438
July 2020
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
