IEEE Consumer Electronics Magazine - March/April 2023 - 114

Special Section on Security, Trust, and Services for AI-Empowered Sensing in the Next Generation IoT
and output gate, are used in the fundamental
structure of LSTM. Adding, discarding, and
updating memory cells are all implied by the
input gate, forget gate, and output gate. The
overall architecture of LSTM is depicted in
Figure 4. The extensive computation of a single
LSTM cell is governed by (1)-(6)
inpt ¼ s biasinp þ Winp ht1;xt
c~ tðÞ 
forgett ¼ s biasforget þWforget ht1;xt
ct ¼ forgett ct1 þ inpt~ic tðÞ
outt ¼ s biasout þ Wout ht1;xt
i ¼ tanh ctoutt
i ¼ tanh bc

i biasc þ Wc ht1;xt


h tðÞ
FIGURE 4. Sign-in and Home Page of the smart
farming mobile application.
predicting fertilizer amount for each unknown
land. According to the performance analysis, it
is found that Bi-LSTM is a deep learning network
consisting of two LSTM architectures for accurate
prediction of fertilizer amount. The data
flow diagram of the proposed IoT-based fertilizer
recommendation system is given in Figure 2.
Step-by-Step Process
Step 1: Initially, the sensor output data were
checked for missing values of 80% for each
land data. Those records will be removed
from the server database. Then, the data
record containing 100% information that will
be considered as input for the Bi-LSTM network,
namely, temperature, humidity, nitrogen
(N), phosphorus (P), and potassium (K).
Step 2: For each record, the three columns of fertilizers
will be calculated based on the agricultural
expert's advice. Finally, the
preprocessed fertilizer recommended data
will be considered for training the learning
model.
Step 3: The data are split as training and testing
of 75:25 ratio for processing the Bi-LSTM prediction
algorithm. During an evaluation, the
unknown 25% of testing data will be compared
with the real fertilizer recommendation
values to know the accuracy.
Step 4: A self-loop memory cell and a network of
gating units, such as input gate, forget gate,
114
(1)
(2)
(3)
(4)
(5)
(6)
where biasinp;cand winp;c are the weight and bias
of the input gate and intermediate cell,
respectively.
Step 5: Bi-LSTM stores information utilizing past
and future values in two layers of LSTM with
memory cells allocated to forward and backward
pass,13 respectively. As a result, several
hidden states (ðht
;ht
!
Þ) were kept and
then concatenated to create a single hidden
state that contained both past and future
data. The LSTM memory cell uses the current
(c0,h0) and hidden input to compute
the next cell state at time step t. Whether to
ignore or add information from the cell is
decided by each memory. As a result, the
memory cell of the forwarded LSTM unit
from the bidirectional technique calculates
the first output cell (h1) and new cell state
(ct) using starting state values. The forward
and backward pass of Bi-LSTM uses the classic
long short-term unit operation to accomplish
this operation at any time step t. Both
directions' final hidden states (ht) are
concatenated into a single output
ht ¼ s ht
;ht
!
where s denotes the aggregating function
includes addition, multiplication, concatenation,
or average.
Step 6: The prediction value fromBi-LSTMnetwork
is treated as recommendation of fertilizers for
IEEE Consumer Electronics Magazine

IEEE Consumer Electronics Magazine - March/April 2023

Table of Contents for the Digital Edition of IEEE Consumer Electronics Magazine - March/April 2023