-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.tex
329 lines (243 loc) · 10.7 KB
/
main.tex
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
\documentclass[a0]{tumposter}
\usepackage[english]{babel}
\usepackage{blindtext}
\usepackage{multicol}
\usepackage{amsmath}
\usepackage{graphicx}
% for printing fontsizes
\usepackage{printlen}
\uselengthunit{mm}
\input{preamble.tex}
\usepackage[utf8]{inputenc}
\title{
Early Classification for Agricultural Monitoring \\ from Satellite Time Series
}
\author{
Marc Rußwurm,\footnotemark[1] Romain Tavenard,\footnotemark[2] Sébastien Lefèvre,\footnotemark[2] Marco Körner\footnotemark[1]
}
\header{
Remote Sensing Technology \\
TUM Department of Civil, Geo and Environmental Engineering \\
Technical University of Munich
}
\begin{document}
\maketitle
\renewcommand{\subsection}[1]{\textbf{#1}}
\begin{minipage}[t]{.65\textwidth}
\section{Objective}
\begin{tikzpicture}[scale=16.8]
\node[label={[name=sat,text height=1.5ex,text depth=.25ex]Satellite Data}, anchor=north](x) at (-1,0){$\M{X} = (\V{x}_0, \V{x}_1, \dots , \V{x}_T)$};
\node[below=of x, label={[yshift=1.3em, xshift=1.8em, font=\tiny, text=white]below:ESA Sentinel 2 Satellite}](s2){\includegraphics[width=11cm]{images/sentinel2}};
\node[below=0em of s2, text width=11cm](eqbox){
\small
\begin{itemize}
\item collected at regular temporal intervals of 2-3 days
\item measurements of 13 spectral bands
\item data available globalls
\end{itemize}
};
\node[label={[name=cm,text height=1.5ex,text depth=.25ex]{Early} {Classification} Model}, anchor=north](f) at (0,0){ $\yhat_t, {\deltat} = f(\xuptot)$};
\node[below=0em of f, text width=13cm, font=\small](info){
\begin{description}
\item[$\xuptot$] observation until $t$ \\
\item[$\yhat_t$] class prediction scores \\
\item[$\deltat$] probability of stopping.
\end{description}
};
% \node[below=of f](example){\input{images/example.tikz}};
\node[circle, below=of info, text width=11cm, fill=tumbluedark, text=white](example){Classifying a \\ satellite time series \\ {\color{tumbluelight}\textbf{accurately}} \color{white} as {\color{ecolor}\textbf{early}} \color{white} as possible};
\node[label={[name=ctm,text height=1.5ex,text depth=.25ex]Crop Type Labels}, anchor=north](y) at (1,0){
$\V{y} = \small (y_\text{corn}, y_\text{barley}, \dots) \in \mathbb{R}^{13}$
};
\node[below=of y, label=below:crop type labels](labels){\includegraphics[width=11cm]{images/parcels}};
\node[below=of labels, text width=11cm](lbls){
\small
\begin{itemize}\setlength\itemsep{.1em}
\item European Common Agricultural Policy (CAP)
\item collected yearly in entire Europe
% \item slowly made publicly available
% \item today, gathered on a national basis
% \item in future harmonized within Europe's INSPIRE directive
\end{itemize}
};
\draw[very thick, -{Stealth[scale=.5]}, tumblue] (x) -- (f);
\draw[very thick, -{Stealth[scale=.5]}, tumblue] (f) -- (y);
\coordinate(bottomleftcolumn) at (example.south -| s2);
\coordinate(bottomrightcolumn) at (example.south -| lbls);
\scoped[on background layer]
{
\node[fit=(ctm)(lbls)(bottomrightcolumn), fill=tumbluelight!20, inner sep=1em, rounded corners]{};
\node[fit=(f)(example)(cm), fill=tumorange!20, inner sep=1em, rounded corners]{};
\node[fit=(sat)(eqbox)(bottomleftcolumn), fill=tumbluelight!20, inner sep=1em, rounded corners]{};
}
\end{tikzpicture}
\section{Method}
Based on previous work (Rußwurm et al., 2019) applied to crop type mapping from remote sensing data.
%A network output to indicate a probability of stopping $\deltat$
\begin{minipage}[t]{.49\textwidth}
\subsection{Mechanism}
\input{images/backprop_stopping_rule.tikz}
\backpropstoppingrulefull
% \begin{tikzpicture}
% \node[fill=tumbluelight!20, rounded corners](a) at (0,0){
% \input{images/backprop_stopping_rule.tikz}
% \backpropstoppingrulefull
% };
% \node[right=of a]{\input{images/qualitative_example.tikz}};
% \end{tikzpicture}
% \vspace{-10em}
% probability of not having classified before (known at training time)
% $$P(t) = \deltat \cdot \prod_{\tau=0}^{t-1} 1 - p_\tau$$
\end{minipage}
\begin{minipage}[t]{.49\textwidth}
\subsection{Loss function}
composite loss function
$$\mathcal{L}(\V{x}, \V{y}) = \sum_{t=0}^T P(t;\deltauptot) \mathcal{L}_t(\xuptot, \V{y})$$
A Loss function including accuracy and earliness
\begin{tikzpicture}[node distance=-.5em, draw=tumbluelight, rounded corners]
% \node(allloss){$\mathcal{L}(\V{x}, \V{y}) = \sum_{t=0}^T P(t;\deltauptot) \mathcal{L}_t(\xuptot, \V{y})$};
%
\node(loss){$\mathcal{L}_t(\xuptot, \V{y})$};
\node[right=of loss](equals){$=$};
\node[right=of equals, fill=accuracycolor!20, rounded corners, label={[font=\small]Classification Loss}](classificationloss){$\alpha \mathcal{L}_c (\xuptot, \V{y})$};
\node[right=of classificationloss](minus){$-$};
\node[right=of minus, fill=earlinesscolor!20, rounded corners, label={[font=\small]Earliness Reward}](earlinessreward){$(1 - \alpha)\mathcal{R}_e(t, \ycorrect_t)$};
\node[below=1em of classificationloss, text width=10cm, xshift=-4em](clossexp){$\mathcal{L}_c = -\log(\ycorrect_t)$ \\
\vspace{.5em}
\tiny
cross entropy loss for accurate classifications \par};
\node[below=1em of earlinessreward, text width=13cm, xshift=-2em](erewardexp){$\mathcal{R}_e(t, \ycorrect_t) = \ycorrect_t \left(1 - \frac{t}{T}\right)$ \\
\vspace{.5em}
\tiny
reduces loss for earlier classifications $1-\frac{t}{T}$ if the correct class $\ycorrect_t$ has been predicted \par};
\draw[thick] (classificationloss) -- (clossexp);
\draw[thick] (earlinessreward) -- (erewardexp);
\end{tikzpicture}
\end{minipage}
{\tiny
Rußwurm, M., Lefèvre, S., Courty, N., Emonet, R., Körner, M., and Tavenard, R. End-to-end learning for early classification of time series. arXiv preprint arXiv:1901.10681, 2019.
\par
}
\vspace{-1em}
\section{Application}
\begin{minipage}[t]{.5\textwidth}
\subsection{Agriculture}
\small
\vspace{1em}
\textbf{Early Crop Detection}
\begin{itemize}
\item early assessment of cultivated crops
\item basis for early crop yield estimation
\end{itemize}
\vspace{.3em}
\textbf{Extraction of Crop Phenology}
\begin{itemize}
\item extraction of vegetation specific events
\item monitoring time of classification
\item regional or temporal variations
\end{itemize}
\vspace{.3em}
\textbf{Generalization}
\begin{itemize}
\item end-to-end trainable
\item no region-specific expert knowledge
\item applicable globally
\end{itemize}
\end{minipage}
\begin{minipage}[t]{.49\textwidth}
\subsection{Dataset and Area of Interest} \par
\begin{minipage}{.6\textwidth}
\small
\vspace{1em}
Hollfeld region Bavaria
\begin{itemize}
\item 49k field parcels
\item 6 main crop types
\item covering 40km by 30 km
\item central Germany
\end{itemize}
\vspace{1em}
\tiny Challenge: Class imbalance \par
\input{images/partition_histograms.tikz}
% \vspace{-4em}\includegraphics[width=.3\textwidth]{images/holl.pdf}
\end{minipage}
\begin{minipage}{.35\textwidth}
% \input{images/partition_histograms.tikz}
\tiny {\color{tumblue}regions with labels} and \\ {\color{tumorange} location of dataset. \par} \par
\includegraphics[width=.9\textwidth]{images/oberfrankenineurope.pdf}
\tiny partition in {\color{traincolor} train}, {\color{validcolor} validation}, and {\color{evalcolor} evaluation} \par
\includegraphics[width=.9\textwidth]{images/holl.pdf}
\end{minipage}
\end{minipage}
\end{minipage}
\hfill
\begin{minipage}[t]{.32\textwidth}
\section{Results}
\subsection{Qualitative Example} \par
{\footnotesize Single example showing reflectance data $\M{x}$ and predictions $\yhat$ along with the stopping time $\tstop \sim \text{Ber}(\deltat)$. \par}
\vspace{.5em}
\input{images/example.tikz}
\vspace{-.5em}
\subsection{Losses during Training} \par
{\footnotesize The combined loss $L_t$, as well as earliness $L_e$ and accuracy $L_e$ losses during training. \par}
\vspace{.5em}
\input{images/loss-accuracyplots.tikz}
\vspace{-1em}
\subsection{Stopping Condition Parameterization} \par
{\footnotesize Stopping times throughout the training grouped by crop category. The parameterization of early classification is learned for different crop types at different times during training. \par}
\input{images/trainingstoppingclasses.tikz}
\subsection{Balancing Earliness and Accuracy} \par
{\footnotesize Evaluting the effect of the trade-off parameter $\alpha$ on the accuracy and earliness ($\tstop$). Runs repeated three times to evaluated the sability of the results. \par}
\begin{table}
\scriptsize
\hspace{0em}\begin{tabular}{lcccccc}
\toprule\small
\textbf{$\alpha$} & accuracy & $\meantstop$ & precision & recall & $f_1$ & $\kappa$ \\
\cmidrule(lr){0-0}\cmidrule(lr){1-1}\cmidrule(lr){2-2}\cmidrule(lr){3-3}\cmidrule(lr){4-4}\cmidrule(lr){5-5}\cmidrule(lr){6-6}\cmidrule(lr){7-7}
.0 & .25 $\pm$ .22 & .10 $\pm$ .17 & .19 $\pm$ .20 & .25 $\pm$ .17 & .16 $\pm$ .20 & .12 $\pm$ .19 \\
.2 & .81 $\pm$ .03 & .40 $\pm$ .02 & .70 $\pm$ .01 & .74 $\pm$ .01 & .71 $\pm$ .01 & .71 $\pm$ .04 \\
.4 & .80 $\pm$ .09 & .47 $\pm$ .03 & .71 $\pm$ .02 & .74 $\pm$ .01 & .71 $\pm$ .02 & .71 $\pm$ .10 \\
.6 & .85 $\pm$ .02 & .88 $\pm$ .07 & .73 $\pm$ .04 & .74 $\pm$ .03 & .73 $\pm$ .03 & .77 $\pm$ .03 \\
.8 & .84 $\pm$ .01 & .93 $\pm$ .05 & .72 $\pm$ .02 & .75 $\pm$ .01 & .73 $\pm$ .02 & .76 $\pm$ .02 \\
1.0 & .83 $\pm$ .03 & 1.00 $\pm$ .00 & .72 $\pm$ .03 & .75 $\pm$ .01 & .72 $\pm$ .03 & .75 $\pm$ .04 \\
\bottomrule
\end{tabular}
% \caption{Varying the weighting factor $\alpha$ for \emph{early reward} loss formulation (\cref{sec:earlynessreward}).}
% \label{tab:alpha}
\end{table}
\subsection{Extracting Vegetation Characteristics} \par
{\footnotesize Stopping time per crop category reveals characteristic variations in type of vegetation confirmed by date of harvest (\druschdatum) from local authorities. \par}
\vspace{.5em}
\input{images/classboxplots.tikz}
\end{minipage}
%\frame{
\begin{footer}
\begin{minipage}{.15\textwidth}
\textbf{ICML Workshop}\\
\includegraphics[width=5cm]{images/AI4SG}
\end{minipage}
\begin{minipage}{.275\textwidth}
\textbf{Technical University of Munich}\footnotemark[1]\\
TUM Department of Civil, Geo and Env. \\
Remote Sensing Technology \\
Arcisstr. 21, 80333 Munich, Germany
\end{minipage}
\begin{minipage}{.275\textwidth}
\textbf{IRISA-Obelix}\footnotemark[2]\\
Université Bretagne Sud \\
IRISA, UMR 6074 CNRS \\
Campus de Tohannic, 56000 Vannes, France
\end{minipage}
\begin{minipage}{.2\textwidth}
\textbf{Data \& Code} \\
% \vspace{1em}
{github.com/rtavenar/early\_rnn} \\
{twitter.com/MarcCoru} \\
www.lmf.bgu.tum.de/vision
\end{minipage}
\begin{minipage}{.05\textwidth}
\hfill\includegraphics[width=5cm]{images/qr-code}\\
\end{minipage}
\end{footer}
%}
\end{document}